Content uploaded by Peter Arijs
Author content
All content in this area was uploaded by Peter Arijs on Jan 28, 2016
Content may be subject to copyright.
Universiteit Gent
Faculteit Toegepaste Wetenschappen
Vakgroep Informatietechnologie
Proefschrift tot het bekomen van de graad van
Doctor in de Toegepaste Wetenschappen:
Elektrotechniek
Academiejaar 1999-2000
Planning van Ring Gebaseerde
Telecommunicatienetwerken
Planning of Ring-Based
Telecommunication Networks
Peter Arijs
Promotor: Prof. Dr. Ir. Piet Demeester
Universiteit Gent
Faculteit Toegepaste Wetenschappen
Vakgroep Informatietechnologie
Sint-Pietersnieuwstraat 41
B-9000 Gent, België
Tel: +32-9-2643316
Fax: +32-9-2643593
WWW: http://www.intec.rug.ac.be
INHOUDSTAFEL
TABLE OF CONTENTS
INHOUDSTAFEL - TABLE OF CONTENTS __________________________________________i
LIJST VAN AFKORTINGEN - LIST OF ABBREVIATIONS ___________________________ix
SAMENVATTING ______________________________________________________________ xv
1. Inleiding_________________________________________________________________ xv
2. Netwerk planning _________________________________________________________ xv
2.1 Evolutie van de telecommunicatiesector ___________________________________________ xv
2.2 Overzicht van de telecommunicatiemarkt __________________________________________xvi
2.3 Belang van netwerk planning __________________________________________________ xvii
2.4 Overzicht van netwerk planningsproblemen _______________________________________ xvii
2.5 Overzicht van netwerk planningstechnieken _______________________________________xviii
3. Technologie en architectuur voor het transportnetwerk _________________________ xix
3.1 Transportnetwerk structuur _____________________________________________________ xix
3.2 Transportnetwerk functionaliteit _________________________________________________ xx
3.3 Transportnetwerk technologieën _________________________________________________xxi
3.3.1 Synchrone digitale hiërarchie (SDH) ___________________________________________ xxi
3.3.2 Golflengte divisie multiplexering (WDM) ______________________________________ xxii
3.4 Pakket geschakelde netwerken _________________________________________________xxiii
3.5 Integratie van pakket geschakelde netwerken in transportnetwerken ____________________ xxiv
3.6 Herstelmechanismen voor transportnetwerken _____________________________________ xxiv
3.6.1 Trail protectie ____________________________________________________________ xxv
3.6.2 Subnetwerk connectie protectie (SNCP)________________________________________ xxv
3.6.3 Ring gebaseerde protectiemechanismen ________________________________________ xxv
3.6.4 Restauratie mechanismen __________________________________________________ xxvii
4. Vergelijking van SDH netwerk architecturen _________________________________xxix
4.1 Inleiding___________________________________________________________________ xxix
4.2 Invoergegevens _____________________________________________________________ xxx
4.3 Netwerk ontwerp ___________________________________________________________xxxiii
4.3.1 Routering ______________________________________________________________xxxiii
4.3.2 Evaluatie _______________________________________________________________ xxxiv
4.3.3 Dimensionering__________________________________________________________ xxxiv
4.3.4 Metrieken ______________________________________________________________ xxxiv
ii
4.4 Resultaten ________________________________________________________________ xxxiv
4.4.1 Resultaten voor SNCP ____________________________________________________ xxxv
4.4.2 Resultaten voor MS-SPRing ________________________________________________ xxxvi
4.4.3 SNCP versus MS-SPRing_________________________________________________ xxxvii
5. Planning en configuratie van een WDM ring_________________________________xxxix
5.1 Inleiding__________________________________________________________________ xxxix
5.2 Planningsproblemen ___________________________________________________________ xl
5.2.1 Lange termijn planning_______________________________________________________ xl
5.2.2 Ring dimensionering_________________________________________________________ xl
5.2.3 SDH gebaseerde WDM ringen ________________________________________________ xli
5.3 Optimale routering op een SPRing _______________________________________________xlii
5.4 Golflengte toekenning op een SPRing____________________________________________ xliii
5.5 Optimalisatie van de hybride DPRing/SPRing architectuur ___________________________ xliv
5.6 Optimalisatie van SDH gebaseerde WDM ringen___________________________________ xlvi
6. Planning van geïnterconnecteerde WDM ringen_______________________________xlvii
6.1 Inleiding___________________________________________________________________xlvii
6.2 Beschouwde planningsproblemen voor geïnterconnecteerde WDM ringen _______________xlvii
6.3 Invoergegevens _____________________________________________________________xlvii
6.4 Het principe van het equivalente netwerk ________________________________________ xlviii
6.5 Ring routering ______________________________________________________________ xlix
6.6 Ring dimensionering ____________________________________________________________ l
6.7 Ring identificatie ______________________________________________________________lii
6.7.1 Ring generatie______________________________________________________________ lii
6.7.2 Ring selectie ______________________________________________________________ liii
6.7.3 Ring optimalisatie__________________________________________________________ liii
6.7.4 Resultaten ________________________________________________________________ liv
6.8 Vergelijking van ring gebaseerde en vermaasde netwerk architecturen ____________________ lv
7. Topologische planning van het toegangsnetwerk _______________________________lvii
7.1 Inleiding____________________________________________________________________lvii
7.2 Toegangsnetwerk planning ____________________________________________________ lviii
7.3 Planning van de boom topologie __________________________________________________lx
7.3.1 Probleemstelling ____________________________________________________________ lx
7.3.2 Oplossingsmethodes _________________________________________________________ lx
7.3.3 Vergelijking van de methodes_________________________________________________ lxi
7.4 Planning van de ring topologie __________________________________________________lxii
7.4.1 Probleemstelling ___________________________________________________________lxii
7.4.2 Oplossingsmethode_________________________________________________________lxii
7.4.3 Resultaten ________________________________________________________________lxii
7.5 Planning van de hybride ring-boom topologie______________________________________ lxiv
7.5.1 Probleemstelling __________________________________________________________ lxiv
7.5.2 Oplossingsmethode_________________________________________________________ lxv
7.5.3 Resultaten _______________________________________________________________ lxvi
8. Finaal overzicht__________________________________________________________lxvii
iii
CHAPTER 1 - INTRODUCTION ___________________________________________________ 1
1.1 Ring, ring_________________________________________________________________ 1
1.2 Contents of this work _______________________________________________________ 1
1.3 Published results___________________________________________________________ 3
CHAPTER 2 - NETWORK PLANNING FOR THE NEW MILLENNIUM ________________ 7
2.1 Introduction ______________________________________________________________ 7
2.2 Evolution of the telecommunications sector_____________________________________ 7
2.2.1 Services _____________________________________________________________________ 7
2.2.2 Technology __________________________________________________________________ 8
2.2.3 Liberalization_________________________________________________________________ 9
2.3 Overview of the telecommunications market___________________________________ 10
2.4 Importance of network planning_____________________________________________ 12
2.5 Different planning problems ________________________________________________ 13
2.5.1 Planning time frame___________________________________________________________ 13
2.5.2 Network planning area_________________________________________________________ 14
2.5.3 Network planner _____________________________________________________________ 15
2.6 Network planning techniques _______________________________________________ 16
2.6.1 Strategic network planning: making investment decisions _____________________________ 17
2.6.2 Tactical network planning: solving dimensioning problems ____________________________ 17
2.7 Conclusion _______________________________________________________________ 19
2.8 References _______________________________________________________________ 20
CHAPTER 3 - TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES ___ 21
3.1 Introduction _____________________________________________________________ 21
3.2 Transport network structure________________________________________________ 21
3.3 Transport network functionalities ___________________________________________ 23
3.3.1 Layering and partitioning ______________________________________________________ 23
3.3.1.1 Partitioning _______________________________________________________________ 24
3.3.1.2 Layering _________________________________________________________________ 24
3.3.1.3 Circuit, path and transmission media layers ______________________________________ 26
3.3.2 Multiplexing ________________________________________________________________ 26
3.3.3 Cross-connecting _____________________________________________________________ 27
3.3.4 Consolidation, segregation and grooming __________________________________________ 28
3.3.5 Management ________________________________________________________________ 29
3.3.6 Connection monitoring ________________________________________________________ 30
3.3.6.1 Inherent monitoring ________________________________________________________ 30
3.3.6.2 Non-intrusive monitoring ____________________________________________________ 31
3.3.6.3 Intrusive monitoring ________________________________________________________ 31
3.3.6.4 Sublayer monitoring ________________________________________________________ 31
iv
3.3.7 Maintaining network integrity ___________________________________________________ 31
3.4 Transport network technologies _____________________________________________ 32
3.4.1 Introduction _________________________________________________________________ 32
3.4.2 Synchronous Digital Hierarchy (SDH) ____________________________________________ 32
3.4.2.1 The early days: PDH________________________________________________________ 32
3.4.2.2 SDH ____________________________________________________________________ 33
3.4.2.3 SDH layers _______________________________________________________________ 35
3.4.2.4 SONET __________________________________________________________________ 36
3.4.2.5 SDH equipment____________________________________________________________ 37
3.4.3 Wavelength Division Multiplexing (WDM) ________________________________________ 38
3.4.3.1 Optical fiber ______________________________________________________________ 38
3.4.3.2 Multi-wavelength transmission________________________________________________ 39
3.4.3.3 The all-optical network ______________________________________________________ 40
3.4.3.4 Management of the optical network ____________________________________________ 43
3.4.3.5 WDM equipment __________________________________________________________ 45
3.5 Packet-switched network technologies ________________________________________ 48
3.5.1 Asynchronous Transfer Mode (ATM)_____________________________________________ 48
3.5.2 Internet Protocol (IP) networks __________________________________________________ 49
3.6 Integration of packet-switching in transport network technologies ________________ 50
3.6.1 IP over ATM over SDH over WDM ______________________________________________ 51
3.6.2 IP over ATM over WDM ______________________________________________________ 51
3.6.3 IP over SDH over WDM _______________________________________________________ 52
3.6.4 IP over SDL over WDM _______________________________________________________ 52
3.6.5 IP over Gigabit Ethernet over WDM______________________________________________ 53
3.6.6 Optical packet switching _______________________________________________________ 53
3.6.7 Proprietary mechanisms________________________________________________________ 54
3.7 Reliable network architectures ______________________________________________ 55
3.7.1 Trail protection ______________________________________________________________ 56
3.7.1.1 Linear protection___________________________________________________________ 56
3.7.1.2 Path protection ____________________________________________________________ 57
3.7.2 Subnetwork Connection Protection (SNCP) ________________________________________ 58
3.7.3 Ring based architectures _______________________________________________________ 58
3.7.3.1 Dedicated protection ring ____________________________________________________ 59
3.7.3.2 Shared protection ring_______________________________________________________ 61
3.7.3.3 Ring interconnection________________________________________________________ 63
3.7.3.4 Drop & continue for interconnected SNCP rings __________________________________ 64
3.7.3.5 Drop & continue for interconnected MS-SPRing__________________________________ 66
3.7.3.6 Ring switched matched nodes for shared protection ring interconnection_______________ 68
3.7.4 Restoration mechanisms _______________________________________________________ 69
3.7.5 Comparison between different protection and restoration mechanisms ___________________ 71
3.7.6 Multi-layer recovery __________________________________________________________ 74
3.8 Conclusion _______________________________________________________________ 76
3.9 References _______________________________________________________________ 77
v
CHAPTER 4 - COMPARISON OF SDH NETWORK ARCHITECTURES _______________ 83
4.1 Introduction _____________________________________________________________ 83
4.2 Input to the network design problem _________________________________________ 85
4.2.1 Network topology ____________________________________________________________ 85
4.2.2 Network traffic ______________________________________________________________ 86
4.2.3 Recovery strategies ___________________________________________________________ 86
4.2.3.1 Path protection using end-to-end SNCP _________________________________________ 86
4.2.3.2 Multiplex Section Shared Protection Rings (MS-SPRing) ___________________________ 86
4.2.3.3 Combination of SNCP path protection and MS-SPRing ____________________________ 87
4.2.4 Link model__________________________________________________________________ 87
4.2.5 Node scenarios_______________________________________________________________ 87
4.2.5.1 NS-1: Based on LO ADMs___________________________________________________ 89
4.2.5.2 NS-2: Based on LO ADMs and HO ADMs ______________________________________ 91
4.2.5.3 NS-3: Based on HO ADMs, LO MUXs, and DXC 4/4 _____________________________ 92
4.2.5.4 NS-4: Based on HO ADMs and DXC 4/3/1 ______________________________________ 94
4.2.5.5 NS-5: Based in HO ADMs and LO MUXs_______________________________________ 95
4.2.5.6 Summary of node scenarios __________________________________________________ 96
4.2.6 Cost model__________________________________________________________________ 96
4.3 Network design method ____________________________________________________ 97
4.3.1 Routing phase _______________________________________________________________ 98
4.3.1.1 End-to-end SNCP path protection______________________________________________ 98
4.3.1.2 MS-SPRing _______________________________________________________________ 98
4.3.1.3 Combination of SNCP path protection and MS-SPRing ____________________________ 99
4.3.2 Traffic evaluation phase _______________________________________________________ 99
4.3.2.1 Link evaluation____________________________________________________________ 99
4.3.2.2 Node evaluation __________________________________________________________ 100
4.3.3 Equipment dimensioning phase_________________________________________________ 101
4.3.3.1 Link dimensioning ________________________________________________________ 101
4.3.3.2 Node dimensioning________________________________________________________ 102
4.3.4 Key performance indicators____________________________________________________ 107
4.4 Results _________________________________________________________________ 108
4.4.1 Results for SNCP____________________________________________________________ 108
4.4.1.1 Comparison between the different node scenarios ________________________________ 108
4.4.1.2 Results for the extended network _____________________________________________ 112
4.4.1.3 Results for different traffic scenarios __________________________________________ 113
4.4.2 Results for MS-SPRing _______________________________________________________ 116
4.4.2.1 Comparison between the different node scenarios ________________________________ 116
4.4.2.2 Results for the extended network _____________________________________________ 118
4.4.2.3 Results for different traffic scenarios __________________________________________ 119
4.4.3 Comparison between the different architectures ____________________________________ 122
4.5 Conclusion ______________________________________________________________ 127
4.6 References ______________________________________________________________ 129
vi
CHAPTER 5 - DIMENSIONING AND CONFIGURATION OF A WDM RING __________ 131
5.1 Introduction ____________________________________________________________ 131
5.2 WDM ring planning ______________________________________________________ 132
5.2.1 Long term planning __________________________________________________________ 132
5.2.1.1 Determination of the architecture _____________________________________________ 132
5.2.1.2 Determination of the ring topology____________________________________________ 133
5.2.2 WDM ring dimensioning______________________________________________________ 133
5.2.2.1 Ring loading _____________________________________________________________ 133
5.2.2.2 Wavelength assignment ____________________________________________________ 135
5.2.2.3 OADM cost influence______________________________________________________ 136
5.2.3 SDH-on-WDM ring dimensioning ______________________________________________ 137
5.3 Ring loading ____________________________________________________________ 139
5.3.1 Problem formulation _________________________________________________________ 139
5.3.2 Bounds on the wavelength requirements__________________________________________ 140
5.3.2.1 General demand pattern ____________________________________________________ 140
5.3.2.2 Hub demand pattern (single star) _____________________________________________ 141
5.3.2.3 Long demand pattern ______________________________________________________ 142
5.3.2.4 Uniform homogeneous demand pattern ________________________________________ 142
5.3.2.5 Adjacent demand pattern ___________________________________________________ 143
5.3.3 Comparison of demand patterns ________________________________________________ 144
5.3.4 Mathematical programming algorithm ___________________________________________ 145
5.3.5 Comparison of DPRing and SPRing _____________________________________________ 145
5.4 Wavelength assignment ___________________________________________________ 146
5.4.1 Problem formulation _________________________________________________________ 146
5.4.2 Results ____________________________________________________________________ 148
5.5 Hybrid DPRing/SPRing architecture ________________________________________ 149
5.5.1 Problem formulation _________________________________________________________ 149
5.5.2 Results ____________________________________________________________________ 150
5.5.3 Extension for rings with drop & continue _________________________________________ 153
5.6 SDH-over-WDM ring design _______________________________________________ 157
5.6.1 Problem formulation _________________________________________________________ 157
5.6.2 Bounds and additional constraints_______________________________________________ 158
5.6.3 Results ____________________________________________________________________ 160
5.7 Conclusion ______________________________________________________________ 162
5.8 References ______________________________________________________________ 164
CHAPTER 6 - PLANNING OF INTERCONNECTED WDM RINGS ___________________ 167
6.1 Introduction ____________________________________________________________ 167
6.2 Planning issues for interconnected WDM rings _______________________________ 169
6.3 Problem formulation _____________________________________________________ 170
6.3.1 Input parameters ____________________________________________________________ 170
vii
6.3.1.1 Network topology_________________________________________________________ 170
6.3.1.2 Traffic __________________________________________________________________ 171
6.3.1.3 Rings___________________________________________________________________ 171
6.3.1.4 Ring interconnection_______________________________________________________ 171
6.3.1.5 Transparency and wavelength conversion ______________________________________ 171
6.3.1.6 Costs ___________________________________________________________________ 172
6.3.1.7 Availability ______________________________________________________________ 172
6.3.2 Network design objectives_____________________________________________________ 173
6.3.2.1 Ring routing _____________________________________________________________ 173
6.3.2.2 Ring dimensioning ________________________________________________________ 173
6.3.2.3 Ring identification ________________________________________________________ 173
6.4 Equivalent network ______________________________________________________ 173
6.5 Ring routing ____________________________________________________________ 175
6.5.1 Integer linear programming____________________________________________________ 175
6.5.2 Shortest path first routing _____________________________________________________ 176
6.5.3 Results ____________________________________________________________________ 177
6.6 Ring dimensioning _______________________________________________________ 179
6.6.1 Integer linear programming____________________________________________________ 180
6.6.2 Heuristic solution method _____________________________________________________ 181
6.6.3 Results ____________________________________________________________________ 182
6.7 Ring identification _______________________________________________________ 188
6.7.1 Ring generation in an existing topology __________________________________________ 189
6.7.1.1 Generating rings containing a given node_______________________________________ 189
6.7.1.2 Generating all possible rings in a given network topology__________________________ 191
6.7.2 Ring pre-selection ___________________________________________________________ 192
6.7.3 Ring optimization ___________________________________________________________ 193
6.7.3.1 Tabu search heuristic ______________________________________________________ 193
6.7.3.2 Exhaustive search _________________________________________________________ 195
6.7.4 Results ____________________________________________________________________ 195
6.7.4.1 Performance of the tabu search heuristic _______________________________________ 195
6.7.4.2 Impact of the main parameter settings _________________________________________ 196
6.8 Comparison of ring and mesh based architectures _____________________________ 197
6.8.1 Cost comparison ____________________________________________________________ 198
6.8.1.1 Link cost ________________________________________________________________ 199
6.8.1.2 Node cost _______________________________________________________________ 200
6.8.2 Availability comparison_______________________________________________________ 200
6.9 Conclusion ______________________________________________________________ 203
6.10 References ______________________________________________________________ 205
CHAPTER 7 - TOPOLOGICAL PLANNING OF THE ACCESS NETWORK ___________ 207
7.1 Introduction ____________________________________________________________ 207
7.2 Access network evolution__________________________________________________ 208
viii
7.2.1 Current technologies _________________________________________________________ 208
7.2.2 Evolution towards broadband access networks _____________________________________ 209
7.3 Access network planning __________________________________________________ 210
7.3.1 Considered planning problem __________________________________________________ 210
7.3.2 Access network planning framework ____________________________________________ 211
7.3.3 Access network topologies ____________________________________________________ 212
7.4 Tree network planning____________________________________________________ 214
7.4.1 Problem formulation _________________________________________________________ 214
7.4.2 Solution methods____________________________________________________________ 214
7.4.2.1 Zoom-in approach_________________________________________________________ 215
7.4.2.2 Iterative path finding approach_______________________________________________ 215
7.4.2.3 Artificial minimum spanning tree approach _____________________________________ 216
7.4.3 Results ____________________________________________________________________ 216
7.4.3.1 Sample networks__________________________________________________________ 216
7.4.3.2 Numerical results _________________________________________________________ 217
7.4.3.3 Discussion of results _______________________________________________________ 219
7.5 Ring network planning____________________________________________________ 220
7.5.1 Problem formulation _________________________________________________________ 220
7.5.2 Solution method_____________________________________________________________ 220
7.5.3 Results ____________________________________________________________________ 223
7.6 Hybrid ring-tree network planning _________________________________________ 227
7.6.1 Problem formulation _________________________________________________________ 227
7.6.1.1 Quantifying network reliability_______________________________________________ 228
7.6.1.2 Network cost_____________________________________________________________ 229
7.6.2 Solution method_____________________________________________________________ 230
7.6.2.1 General optimization scheme ________________________________________________ 230
7.6.2.2 Cluster optimization _______________________________________________________ 230
7.6.2.3 Ring optimization _________________________________________________________ 232
7.6.2.4 Interaction between cluster and ring optimization ________________________________ 234
7.6.3 Results ____________________________________________________________________ 234
7.6.3.1 Reliability versus cost in tree networks ________________________________________ 234
7.6.3.2 Reliability versus cost in ring network _________________________________________ 235
7.6.3.3 Interaction between cluster and ring optimization ________________________________ 237
7.6.3.4 Conclusion ______________________________________________________________ 238
7.7 Conclusion ______________________________________________________________ 238
7.8 References ______________________________________________________________ 240
CHAPTER 8 - CONCLUSIONS __________________________________________________ 243
8.1 Last call… ______________________________________________________________ 243
8.2 Where do we go from here?________________________________________________ 246
8.3 References ______________________________________________________________ 247
LIJST VAN AFKORTINGEN
LIST OF ABBREVIATIONS
A
AAL ATM Adaptation Layer
ADM Add-Drop Multiplexer
ADSL Asymmetric Digital Subscriber Line
AIS Alarm Indication Signal
AP Access Point
APS Automatic Protection Switching
ATM Asynchronous Transfer Mode
AUG Administrative Unit Group
AU-n Administrative Unit of order n
B
BER Bit Error Rate
BGP Border Gateway Protocol
BML Business Management Layer
C
CAPEX Capital Expenditures
CATV Common Antenna Television
CLEC Competitive Local Exchange Carrier
C-n Container of order n
CoS Class of Service
CP Connection Point
CRC Cyclic Redundancy Check
D
D&C Drop & Continue
DBFA Dual Band Fiber Amplifier
DCF Dispersion Compensating Fiber
DPRing Dedicated Protection Ring
DPT Dynamic Packet Transport
DS Digital Signal
DSF Dispersion Shifted Fiber
DTM Dynamic Transfer Mode
DWDM Dense Wavelength Division Multiplexing
x
DXC Digital Cross-Connect
E
EDFA Erbium Doped Fiber Amplifier
EGP Exterior Gateway Protocol
ELT Expected Loss of Traffic
EMI Electro-Magnetic Interference
EML Element Management Layer
F
FDM Frequency Division Multiplexing
FEC Forward Error Correction
FITL Fiber-In-The-Loop
FTTB Fiber-To-The-Building
FTTC Fiber-To-The-Cabinet
FTTH Fiber-To-The-Home
FWM Four Wave Mixing
G
GIS Geographic Information System
GSM Global System for Mobile communications
H
HDLC High-level Data Link Control
HO Higher Order
HOP Higher Order Path
I
IETF Internet Engineering Task Force
IGP Interior Gateway Protocol
ILEC Incumbent Local Exchange Carrier
ILP Integer Linear Programming
IP Internet Protocol
IRR Internal Rate of Return
IS-IS Intermediate System to Intermediate System
ITU International Telecommunication Union
IXC Interexchange Carrier
K
KPI Key Performance Indicator
KSP K Shortest Paths
L
LAN Local Area Network
LEX Local Exchange
x
i
LO Lower Order
LOP Lower Order Path
LP Linear Programming
M
MAC Medium Access Control
MEMS Micro Electro-Mechanical Switches
MPLS Multi-Protocol Label Switching
MS Multiplex Section
MS-DPRing Multiplex Section Dedicated Protection Ring
MSOH Multiplex Section Overhead
MSP Multiplex Section Protection
MS-SPRing Multiplex Section Shared Protection Ring
MTBF Mean Time Between Failures
MTTR Mean Time To Repair
MUX Multiplexer
N
NE Network Element
NEL Network Element Layer
NIC Network Interface Card
NML Network Management Layer
NPV Net Present Value
NS Node Scenario
NUT Non-preemptible Unprotected Traffic
NZDF Non Zero Dispersion shifted Fiber
O
OA Optical Amplifier
OADM Optical Add-Drop Multiplexer
OAM Operations, Administration and Maintenance
OCh Optical Channel
OCh-DPRing Optical Channel Dedicated Protection Ring
OCh-SPRing Optical Channel Shared Protection Ring
OC-n Optical Channel of order n
OEO Optical Electrical Optical (conversion)
OMS Optical Multiplex Section
OMS-SPRing Optical Multiplex Section Shared Protection Ring
ON Optical Node
OPEX Operational Expenditures
OSC Optical Supervisory Channel
OSI Open Systems Interconnection
OSPF Open Shortest Path First
OSS Operational Support System
OTDM Optical Time Division Multiplexing
x
ii
OTN Optical Transport Network
OTS Optical Transmission Section
OXC Optical Cross-Connect
P
PABX Private Automatic Branch Exchange
PDH Plesiochronous Digital Hierarchy
PMD Polarization Mode Dispersion
POH Path Overhead
PON Passive Optical Network
POP Point Of Presence
POS Packet over SONET
PPP Point-to-Point Protocol
PVC Permanent Virtual Circuit
Q
QoS Quality of Service
R
RBOC Regional Bell Operation Companies
RIP Ring Interworking on Protection
RS Regeneration Section
RSMN Ring Switched Matched Nodes
RSOH Regeneration Section Overhead
RSVP Resource Reservation Protocol
S
SC Street Cabinet
SCM Subcarrier Modulation
SDL Simple Data Link
SDM Space Division Multiplexing
SLA Service Level Agreement
SME Small and Medium Enterprises
SMF Single Mode Fiber
SML Service Management Layer
SNCP Subnetwork Connection Protection
SNR Signal to Noise Ratio
SOH Section Overhead
SONET Synchronous Optical Network
SPRing Shared Protection Ring
SSF Service Signal Fail
STM-n Synchronous Transfer Mode of order n
STS-n Synchronous Transport Signal of order n
SVC Switched Virtual Circuit
x
iii
T
TCP Termination Connection Point
TCP Transport Control Protocol
TDM Time Division Multiplexing
TM Terminal Multiplexer
TMN Telecommunications Management Network
TUG Tributary Unit Group
TU-n Tributary Unit of order n
U
UDP User Datagram Protocol
UMTS Universal Mobile Telecommunications System
V
VC Virtual Connection
VCI Virtual Connection Identifier
VC-n Virtual Container of order n
VDSL Very high-speed Digital Subscriber Line
VP Virtual Path
VPI Virtual Path Identifier
VPN Virtual Private Network
VT Virtual Tributary
VWP Virtual Wavelength Path
W
WDM Wavelength Division Multiplexing
WP Wavelength Path
X
XC Cross-Connect
x
iv
SAMENVATTING
Planning van ring gebaseerde
telecommunicatienetwerken
1. Inleiding
De snelle groei van telecommunicatiediensten en technologie, gekatalyseerd door de
liberalisering van de sector, legt steeds toenemende eisen op huidige telecommunicatie-
netwerken. Het ontwerp en voortdurend uitbreiden van het netwerk is een complex proces,
waarbij rekening dient gehouden te worden met tal van factoren zoals kost, schaalbaarheid,
flexibiliteit, betrouwbaarheid, enzovoort. Geavanceerde planningstechnieken zijn vereist om
deze complexiteit op te vangen. Dit werk behelst een aantal planningsproblemen en
oplossingstechnieken voor ring gebaseerde telecommunicatienetwerken. Ringen vormen een
eenvoudig maar krachtig concept om betrouwbare netwerken uit te bouwen. Dit wordt in deze
thesis onderstreept door de toepassing van ringen in diverse gebieden van het netwerk aan te
tonen.
2. Netwerk planning
In hoofdstuk 2 wordt een kort overzicht gegeven van de recente evoluties binnen de
telecommunicatiesector, en hoe deze het netwerk planningsproces sturen. We schetsen het
doel en de verschillende toepassingen van netwerk planning, en geven aan hoe het netwerk
planningsprobleem praktisch kan opgelost worden aan de hand van een aantal economische
en wiskundige technieken.
2.1 Evolutie van de telecommunicatiesector
De evolutie van de telecommunicatiesector is de laatste jaren in een ware
stroomversnelling terechtgekomen. Dit is te merken aan de steeds sterker toenemende groei
van de hoeveelheid informatie die wereldwijd wordt uitgewisseld via telecommunicatie-
netwerken. Deze sterke groei wordt gedragen door een aantal factoren. Enerzijds werden de
laatste jaren een aantal uiterst succesvolle diensten geïntroduceerd, die de vrijheid en het
gemak van communiceren stimuleerden. Bekende voorbeelden die ondertussen gemeengoed
zijn geworden, zijn de mobiele communicatie (GSM) en het gebruik van het internet voor het
uitwisselen van allerhande informatie in de vorm van digitale data. Teneinde deze nieuwe
diensten aan te bieden, dienen de huidige telecommunicatienetwerken aangepast te worden
met behulp van nieuwe technologieën. Een tweede factor die de groei van de
telecommunicatie drijft, is dan ook de snelheid van technologische innovatie die sterk is
SAMENVATTING
x
vi
toegenomen. De nieuwe produkten die daaruit voortspruiten laten operatoren niet alleen toe
nieuwe diensten aan te bieden, maar drukken tevens de kostprijs van het netwerk en
vergemakkelijken de uitbouw ervan. Bekende voorbeelden van recente technologische
innovaties binnen de telecommunicatiesector zijn: golflengte multiplexering op optische
vezels (WDM), nieuwe modulatietechnieken op koper draden (zoals ADSL), terabit routers
die kolossale hoeveelheden informatie over het internet sturen, enzovoort. Ten laatste heeft
ook de liberalisering van de telecommunicatiesector een belangrijke invloed gehad. Hierdoor
konden nieuwe spelers tot de markt toetreden, en de bestaande operatoren beconcurreren. De
hieruit voortvloeiende prijzenslag heeft een gunstig effect op de tarieven voor de
eindgebruikers, maar heeft een vermindering van de winstmarge tot gevolg voor de
operatoren. De netto winst van de operator kon nog wel in grote mate gehandhaafd worden
dankzij een stijgende omzet (wegens de toenemende hoeveelheid verkeer), interne
kostenbesparingen, en door gebruik te maken van schaalvoordelen (getuige hiervan zijn de
vele overnames en allianties in de sector). Daarnaast kan in een klimaat van tanende
winstmarges en toenemende investeringen, een efficiënte netwerk planning in belangrijke
mate bijdragen tot het verbeteren van de rendabiliteit van de operator.
2.2 Overzicht van de telecommunicatiemarkt
Binnen de telecommunicatiemarkt kunnen een zestal spelers geïdentificeerd worden.
De eindgebruikers zijn ongetwijfeld de belangrijkste spelers. Zij consumeren de diensten,
zorgen ervoor dat netwerken nodig zijn, en netwerkapparatuur verkocht wordt. Daarom is het
belangrijk dat de dienstenleveranciers de noden van de klant goed kunnen inschatten,
teneinde hierop terdege te kunnen inspelen. Verschillende types klanten vragen immers
verschillende types diensten: zo zijn de vereisten van residentiële klanten bijvoorbeeld totaal
verschillend van deze van zakenklanten. Terwijl sommige dienstenleveranciers slechts een
minimale hoeveelheid aan netwerkinfrastructuur bezitten, zijn andere dienstenleveranciers
tegelijkertijd netwerk operatoren. Netwerk operatoren bezitten een netwerk en verkopen
connectiviteit aan hun klanten. Deze klanten kunnen eindgebruikers zijn indien de operator
ook diensten levert, maar het kunnen ook andere dienstenleveranciers of zelfs operatoren zijn.
Teneinde een netwerk uit te bouwen, is er nood aan netwerkapparatuur. De leveranciers van
netwerkapparatuur verkopen de bouwblokken van het netwerk. Terwijl de grote apparatuur-
leveranciers (bijvoorbeeld Alcatel, Lucent, Nortel) produkten ontwikkelen voor elk segment
van het telecommunicatienetwerk, focussen opstartende leveranciers op één enkel segment of
nicheprodukt. De toenemende keuze aan types apparatuur en de leveranciers ervan, geven
operatoren meer mogelijkheden, maar anderzijds bemoeilijkt deze ruime keuze ook de
planning van het netwerk. Teneinde de wildgroei van netwerk technologieën in goede banen
te leiden, zijn welomlijnde standaarden van groot belang. Standaardisatie instituten creëren
aldus één gemeenschappelijke markt in plaats van vele verbrokkelde markten, wat in het
voordeel is van zowel operatoren als leveranciers van apparatuur. Ten laatste spelen ook de
regelgevende organen een belangrijke rol binnen de huidige telecommunicatiemarkt. In het
initiële stadium van liberalisering van de markt, zien regelgevende organen toe op een
geleidelijke overgang naar een competitieve markt teneinde de consumenten en operatoren te
beschermen.
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
vii
2.3 Belang van netwerk planning
Netwerk planning bepaalt de stappen die dienen gezet te worden teneinde het
gewenste netwerk op een gestage, efficiënte en economische manier uit te bouwen. Door het
nauw opvolgen van de netwerk evolutie, het accuraat voorspellen van de vraag, en het gebruik
van geschikte planningstechnieken kan een gestage uitbouw van het netwerk in de hand
worden gewerkt. Hierbij is het van belang dat kritische beslissingen niet te vroeg worden
genomen (resulterend in overbodige uitgaven) maar ook niet te laat worden genomen
(resulterend in verlies van marktpositie). Bij de efficiënte uitbouw van het netwerk dient de
technologie en architectuur van het netwerk zodanig gekozen te worden dat alle diensten van
de operator op een eenvoudige maar gegarandeerde manier kunnen aangeboden worden.
Bovendien dient rekening gehouden te worden met toekomstige groei van verkeer en
diensten. Ten laatste dient een economische uitbouw van het netwerk ervoor te zorgen dat het
netwerk zodanig wordt uitgebouwd dat de winst van de operator wordt gemaximaliseerd.
Door de enorme investeringen en operationele kosten die gepaard gaan bij de uitbouw en het
onderhouden van een netwerk, kan een kost-efficiënte netwerk planning de winst van een
operator gevoelig beïnvloeden. Hierbij is vooral het hefboomeffect van de kostenbesparingen
op de winst van groot belang in een klimaat van toenemende investeringen en krimpende
winstmarges.
Binnen deze thesis wordt netwerk planning op twee manieren gehanteerd. Enerzijds
wordt gekeken naar de situatie waarbij het hoofddoel bestaat uit de kost minimalisatie van een
goed gedefinieerde netwerk structuur. In dit geval zijn alle beslissingen omtrent technologie
en architectuur keuze reeds genomen, en dient enkel nog de uitbouw van het netwerk op een
zo economisch mogelijke wijze te gebeuren. Anderzijds wordt ook de situatie bekeken
waarbij de technologie en architectuur nog niet gekozen zijn. Door een netwerk planning uit
te voeren voor verschillende technologie en architectuur keuzes, kunnen de voor- en nadelen
van elk scenario tegen elkaar afgewogen worden. Als dusdanig kan netwerk planning dus ook
bijdragen aan de keuze van de netwerk technologie en architectuur.
2.4 Overzicht van netwerk planningsproblemen
De aard van het netwerk planningsprobleem wordt gedreven door de
planningstermijn, het geografisch of functioneel gebied van het te plannen netwerk en de
identiteit en het standpunt van de netwerk planner.
Afhankelijk van de planningstermijn spreken we van verschillende soorten planning.
Lange termijn ‘strategische’ planning kijkt 5 tot 10 jaar vooruit, en focusseert op strategische
beslissingen zonder hierbij veel details over het netwerk in rekening te brengen (bijvoorbeeld
omschrijving van het zakenplan, definitie van de aan te bieden diensten, voorspellingen van
de vraag, keuze van technologie, …). Middellange termijn ‘tactische’ planning blikt 1 à 2 jaar
vooruit en is meer begaan met het dimensioneren van het netwerk. Korte termijn
‘operationele’ planning houdt zich bezig met de dagdagelijkse operationele taken, zoals het
routeren van verkeer. Binnen deze thesis worden vooral tactische en in mindere mate
strategische planningstaken bekeken.
Het netwerk kan in 3 functionele gebieden onderverdeeld worden, elk met hun eigen
planningstaken. Het toegangsnetwerk, verbindt de individuele gebruiker met de lokale
centrale binnen de zone. Gezien de grote investeringen die in dit deel van het netwerk
SAMENVATTING
x
viii
plaatsvinden, kan een degelijke netwerkplanning resulteren in belangrijke kostbesparingen.
Strategische planning binnen het toegangsnetwerk beslaat de keuze van de technologie en
architectuur. Tactische planning houdt zich bezig met de optimalisatie van de topologie
(bijvoorbeeld locatie van centrales en transmissielijnen). Het geschakelde netwerk zet op
dynamische wijze verbindingen op tussen verschillende schakelcentrales. Hierbij wordt
vooral onderscheid gemaakt tussen het telefonie en pakket geschakelde netwerk. Tactische
planning van het geschakelde netwerk omhelst dimensionering van de schakelcentrales en van
de tussenliggende transmissielijnen zodat alle verkeer op elk moment van de dag kan
gerouteerd worden. Ook het ontwikkelen van routeringsalgoritmes om dynamisch connecties
op te zetten kan deel uitmaken van het (operationele) planningsproces binnen het geschakelde
netwerk. Het transportnetwerk verzorgt statische verbindingen van hoge capaciteit tussen
diverse toegangspunten in het netwerk, die bijvoorbeeld kunnen gebruikt worden door het
geschakelde netwerk. Hiervoor dient de juiste technologie, topologie en architectuur gekozen
te worden (strategische planning) alsook de benodigde capaciteit binnen het netwerk
(tactische planning). In deze thesis worden verschillende deelaspecten van de planning van
het transportnetwerk en het toegangsnetwerk beschouwd.
Het doel van netwerk planning kan gevoelig verschillen naargelang de identiteit van
de netwerkplanner. Gevestigde operatoren (zoals de vroegere monopolisten) moeten de
bestaande apparatuur, die over de jaren heen in het netwerk werd geïnstalleerd, in rekening
brengen. Dit legt extra randvoorwaarden op, die moeten in rekening gebracht worden bij
capaciteitsuitbreidingen. Pas indien het onderhouden of uitbreiden van bestaande apparatuur
te kostelijk wordt, of deze niet langer in staat is nieuwe diensten aan te bieden, kan
overgestapt worden naar een nieuwe technologie. Nieuwe operatoren, moeten een netwerk
van de grond af opbouwen, en moeten dus nog belangrijke beslissingen nemen in verband met
de te adopteren netwerktopologie (in welke gebieden toegang te verschaffen, hoe de
verschillende toegangspunten met elkaar te verbinden, …). Bij het nemen van technologische
beslissingen, moeten nieuwe operatoren met minder randvoorwaarden rekening houden, en
kunnen meteen voor de – op dat ogenblik – meest geschikte technologie kiezen. Alhoewel
leveranciers van netwerkapparatuur zelf niet over een netwerk beschikken, kunnen ook zij
gebaat zijn met netwerk planning. Enerzijds kunnen zij deze netwerk planning aanbieden aan
hun klanten (de netwerk operatoren), door voor hen studies uit voeren of door operationele
planningstaken te integreren in de beheerssystemen van de apparatuur. Anderzijds kunnen zij
intern planningstechnieken gebruiken om de toepassing van hun produkten te vergelijken met
concurrerende produkten, en op basis daarvan de prijszetting te bepalen. In deze thesis wordt
netwerkplanning hoofdzakelijk bekeken vanuit het oogpunt van de operator, alhoewel
sommige aspecten ook betrekking hebben op leveranciers van netwerkapparatuur.
2.5 Overzicht van netwerk planningstechnieken
Voor strategische netwerk planning, is het niet nodig een gedetailleerd beeld van het
netwerk te hebben. De meeste strategisch beslissingen zijn gerelateerd aan de economische
impact van het uit te voeren project. Teneinde economische studies uit te voeren die de
waarde van een project meten, is het niet voldoende om te kijken naar cijfers als
investeringen, omzet en winst, omdat deze onvoldoende rekening houden met het tijdsaspect
(zoals afschrijvingen en de tijdswaarde van geld). Daartoe dienen meer geavanceerde
economische metrieken gebruikt te worden, zoals netto courante waarde, terugverdientijd en
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
ix
investeringsrendement. Meerdere metrieken, die elk een ander aspect belichten, laten toe een
meer gefundeerde keuze te maken tussen verschillende projecten.
Tactische en operationele planning houdt zich meer bezig met de optimale uitbouw
van het netwerk, eens de meeste beslissingen omtrent technologie en architectuur genomen
zijn. Hiertoe dient o.a. onderzocht te worden welke toegangspunten te kiezen, hoe deze
toegangspunten te verbinden met transmissielijnen, hoe verkeer te routeren in het netwerk,
welke dimensies de apparatuur moet hebben, enzovoort. Een van de voornaamste objectieven
hierbij is minimalisatie van de investeringskost. Als dusdanig kan het planningsprobleem
gemodelleerd te worden als een wiskundig optimalisatieprobleem van combinatorische aard.
Hierop kunnen bestaande optimalisatietechnieken toegepast worden, ofwel kan een specifieke
oplossingsmethode toegepast worden, die op maat van het probleem wordt gemaakt.
Bestaande optimalisatietechnieken die typisch gebruikt worden voor netwerk planning zijn
lineaire programmering en heuristische zoekmethodes (zoals gesimuleerde afkoeling,
genetische algoritmes en taboe zoeken). Deze laatste garanderen echter geen optimale
oplossing, maar laten wel toe om een goede oplossing te bekomen in een korte rekentijd.
3. Technologie en architectuur voor het transportnetwerk
In hoofdstuk 3 wordt een overzicht gegeven van de technologie en architectuur van
hedendaagse en toekomstige transportnetwerken. Er wordt gestart met een overzicht van de
structuur en belangrijkste functies van het transportnetwerk. Hierna, worden de technologieën
besproken die dit kunnen realiseren. Hierbij wordt de nadruk gelegd op SDH en WDM
transportnetwerken, maar ook pakket geschakelde transportnetwerken (zoals ATM en IP), die
voortdurend aan belang winnen, worden niet uit het oog verloren. In het bijzonder de
convergentie van beide technologieën wordt bekeken vanuit de invalshoek van IP over WDM
transport. Tenslotte belicht dit hoofdstuk uitvoerig de verschillende herstelmechanismen die
een transportnetwerk resistent tegen fouten kunnen maken. De werking van zowel ring
gebaseerde mechanismen, als oplossingen voor vermaasde netwerken, wordt besproken en de
pro en contra van elke strategie worden aangehaald.
3.1 Transportnetwerk structuur
Het doel van elk netwerk is het uitwisselen van informatie tussen de verschillende
gebruikers. Teneinde een schaalbare en kost-efficiënte netwerkstructuur te bekomen, wordt
het netwerk typisch gestructureerd in aan aantal hiërarchische niveaus. Op het laagste niveau,
het toegangsnetwerk, wordt elke gebruiker verbonden met de lokale centrale in zijn zone. Het
transportnetwerk zorgt voor de verbinding tussen de verschillende lokale centrales en
tussenliggende schakelcentrales, zodat gebruikers in verschillende zones kunnen
communiceren. Het transportnetwerk zelf wordt ook nog eens onderverdeeld in hiërarchische
niveaus. Het metropool netwerk zorgt voor de verbinding van lokale centrales binnen
eenzelfde stad. Het regionaal netwerk, verbindt metropole netwerken en centrales van
verschillende steden in dezelfde regio. Finaal worden de regionale netwerken verbonden via
het kernnetwerk. Elk van deze hiërarchische niveaus vervult andere taken, en verschilt dus in
functionaliteit en planningsproblemen die dienen opgelost te worden.
SAMENVATTING
x
x
3.2 Transportnetwerk functionaliteit
In deze paragraaf behandelen we een aantal functionele taken en concepten waarin
een transportnetwerk dient te voorzien.
We beginnen met de concepten van lagering en partitionering, die toelaten om een
netwerk bestaande uit meerdere technologieën op een efficiënte en generische manier te
beschrijven. Partitionering laat toe een netwerk te verdelen in een aantal duidelijk
afgebakende subnetwerken (zoals hiërarchische niveaus of administratieve domeinen), die elk
afzonderlijk beheerd worden en verbonden worden aan de hand van links. Lagering laat toe
een netwerk op te splitsen in een aantal netwerklagen, waarbij aangrenzende lagen in een
klant/dienaar (Engels: client/server) relatie staan met elkaar. Elke onderliggende laag bedient
de bovenliggende laag (zijnde de klant) van transportfaciliteiten. Zo wordt een link in een
hogere laag ondersteund door een netwerkconnectie in de onderliggende laag.
Een belangrijke functie waarin een transportnetwerk dient te voorzien, is het
multiplexeren van signalen. Multiplexering laat toe om verschillende (tributaire) signalen
samen te nemen als één geaggregeerd signaal, dat dan met verminderde kost en complexiteit
kan vervoerd worden doorheen het netwerk. Dit maakt deel uit van het adaptatie proces, bij
het overgaan van een hogere netwerklaag naar een lagere netwerklaag. Multiplexering, kan
technisch gesproken, op verschillende domeinen plaatsvinden. De meest gekende voorbeelden
zijn multiplexering in de tijd, in de ruimte en in het frequentie domein.
Een andere belangrijke functie is cross-connecting, die toelaat om een volledig
vermaasd logisch netwerk op te bouwen binnen een veel minder vermaasde onderliggende
netwerkstructuur. Een cross-connect termineert verschillende geaggregeerde links en laat toe
tributaire signalen te verbinden tussen deze links. Als dusdanig kunnen semi-permanente
tributaire connecties opgezet worden tussen elke twee punten binnen het netwerk, waarbij elk
zo een connectie een geaggregeerd signaal op zichzelf is (en dus als link fungeert voor de
bevattende tributaire signalen).
Onder het consolideren van verkeer, verstaan we het samennemen van verschillende
verkeersstromen op één transportfaciliteit, teneinde de vullingsgraad van deze faciliteit te
bevorderen. De segregatie van verkeer is de inverse actie, waarbij verkeer opgesplitst wordt
in verschillende verkeersstromen (bijvoorbeeld naar verschillende locaties). Het proces
waarbij consolideren en segregeren gecombineerd wordt, teneinde de totale netwerkefficiëntie
te bevorderen, wordt grooming genoemd.
Naast de transportnetwerk functies die zich inhouden met het bewerken van verkeer,
zijn er ook functies die het beheer of management van het netwerk en het verkeer tot zich
nemen. Hieronder vallen o.a. fout management (het detecteren en opvangen van fouten),
configuratie management (het configureren van netwerkelementen teneinde diensten aan te
bieden), rekening management (tarificatie van de gebruikers voor het gebruik van het
netwerk), prestatie management (het superviseren van de kwaliteit van het netwerk en de
diensten) en beveiliging (bepaalde gebruikers toegang tot het netwerk verhinderen of
beperken). Vooral prestatie management (Engels: performance management) neemt een
belangrijke rol in. Specifieke technieken voor prestatie management kunnen in de Engelse
tekst teruggevonden worden.
Een laatste belangrijke taak van het transportnetwerk is het waken over de
foutbestendigheid van het netwerk. Dit kan enerzijds gebeuren door pro-actieve maatregelen
te treffen, zowel op software implementatie vlak, en door het gebruik van betrouwbare
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
xi
hardware componenten. Anderzijds kunnen bij het optreden van een fout ook reactieve
handelingen uitgevoerd worden, zoals het herrouteren van een deel van het verkeer.
3.3 Transportnetwerk technologieën
In deze paragraaf stellen we technologieën voor die de hoger beschreven
transportnetwerk functies kunnen vervullen. In deze transportnetwerken, wordt communicatie
tussen twee punten verzorgd door een semi-permanente connectie. Zo een connectie is een
communicatiekanaal met een vaste capaciteit en route, welke niet gewijzigd wordt tijdens de
communicatie. Transportnetwerken zijn een typische voorbeeld van connectie georiënteerde
netwerken, waarbij communicatie in 3 fazen verloopt: het opzetten van de connectie, het
uitwisselen van informatie, en het afbreken van de connectie. Een connectie kan enkel
opgezet worden, als er voldoende voorzieningen (zoals capaciteit) aanwezig zijn in het
netwerk. Eens de connectie opgezet is, kan een kwalitatieve communicatie gegarandeerd
worden.
3.3.1 Synchrone digitale hiërarchie (SDH)
De synchrone digitale hiërarchie (SDH) is een wereldwijde standaard om
tijdsgemultiplexeerde signalen op een efficiënte wijze te versturen. Het basissignaal is een
STM-1 signaal, dat een debiet heeft van 155.52 Mb/s. SDH ondersteunt verschillende
multiplexeringsstappen, gebruik makend van zogeheten virtuele containers (VC), om diverse
signalen met een lager debiet onder te brengen in een STM-1. STM-1 signalen kunnen op hun
beurt gemultiplexeerd worden in een STM-N signaal (bijvoorbeeld STM-16 van 2.5 Gb/s).
SDH kan makkelijkst beschreven worden aan de hand van zijn lagenstructuur (zie
Figuur 1). De laagste laag, de fysische media laag, beschrijft het fysische medium waarover
het signaal verstuurd wordt (bijvoorbeeld coaxiale kabel of optische vezel) en moduleert het
STM-N signaal over dit fysisch medium. De bovenliggende laag, de regenerator sectie (RS)
laag, zorgt voor een regelmatige regeneratie van het STM-N signaal, dat gedegradeerd wordt
door imperfecties in de fysische media laag (zoals verzwakking, vervorming, …). De
volgende laag is de multiplex sectie (MS) laag. Deze MS laag is verantwoordelijk voor het
multiplexeren van signalen in een STM-N signaal. Een multiplex sectie is typische
gedefinieerd tussen twee multiplexers, en doorloopt een aantal regenerator secties. De MS
laag ondersteunt de zogeheten pad lagen, die verantwoordelijk zijn voor het assembleren van
verkeer in virtuele containers, en die cross-connecting flexibiliteit bieden teneinde het verkeer
te routeren. We onderscheiden de hogere orde (HO) padlaag, die VC-4 (150 Mb/s) containers
vervoert en daarboven de lagere orde (LO) padlaag die VC-3 (48 Mb/s), VC-2 (8.5 Mb/s),
VC-11 (1.5 Mb/s) en VC-12 (2 Mb/s) containers vervoert. Naast de multiplexeringsstappen,
voorziet deze lagenstructuur eveneens in de benodigde overhead bytes, om het netwerk te
beheren en superviseren.
SAMENVATTING
x
xii
Circuit laag netwerken
VC-11 VC-12 VC-2 VC-3
VC-3 VC-4
+ AUG + MSOH
+ RSOH (= STM-N)
Optische vezel, coax kabel, ...
circuit
laag
lagere orde
pad laag
hogere orde
pad laag
multiplex
sectie laag
regenerator
sectie laag
fysische
media laag
pad laag
sectie laag
Figuur 1: Lagenstructuur van SDH
SDH werd in het begin van de jaren ’80 geïntroduceerd, en is dus een gevestigde
technologie waarvoor de apparatuur reeds vele jaren bestaat. SDH apparatuur kan
onderverdeeld worden in 3 categorieën: multiplexers, add-drop multiplexers (ADM) en
digitale cross-connects (DXC). Multiplexers aggregeren meerdere tributaire signalen in één
STM-N signaal. ADMs zijn een uitbreiding van multiplexers, met twee geaggregeerde STM-
N poorten en één poort voor het toevoegen of afhalen (= add/drop) van lokaal getermineerd
verkeer. Verkeer dat niet lokaal getermineerd wordt, passeert op transparante wijze door de
ADM. DXCs hebben, in tegenstelling tot ADMs, meerdere geaggregeerde STM-N poorten en
ook poorten voor lokaal getermineerd verkeer, tussen dewelke verkeer kan worden gecross-
connecteerd. Zowel ADMs en DXCs kunnen op LO en/of HO niveau werken.
3.3.2 Golflengte divisie multiplexering (WDM)
Golflengte divisie multiplexering (WDM), laat toe om verschillende signalen te
moduleren op één optische vezel, elk op een andere golflengte. Hierdoor kunnen sterke
kostbesparingen bekomen worden ten opzichte van het alternatief, waarbij meerdere parallelle
vezelparen gebruikt worden. Door meerdere signalen op één vezel te multiplexeren, kunnen
deze gezamenlijk worden versterkt, waardoor veel minder versterkers nodig zijn. Bovendien
is het ook economisch voordelig om signalen optisch (in plaats van elektrisch) te gaan
verwerken in de knopen van het netwerk.
Net zoals in SDH, kunnen we ook in WDM verschillende netwerklagen
identificeren: de optische kanaal laag (OCh), de optische multiplex sectie laag (OMS) en de
optische transmissie sectie laag (OTS). De OCh laag is het equivalent van de padlaag in
SDH, en voorziet in transport van golflengten doorheen het netwerk. De OMS laag is het
optische equivalent van de SDH MS laag, en ondersteunt de OCh laag. De OTS laag is het
equivalent van de SDH RS laag en zorgt voor het correct transport van een golflengte-
gemultiplexeerd signaal over verschillende types optische vezel.
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
xiii
Wegens de verzwakking en vervorming die een signaal ondervindt, is het momenteel
onmogelijk een optisch signaal over grote afstand te transporteren zonder tussenliggende
elektrische conversie en regeneratie. Daarom werd het concept van opake optische netwerken
ingevoerd. In een opaak optisch netwerk wordt het signaal in elke tussenliggende knoop
geconverteerd naar het elektrisch domein en wordt het als dusdanig geregenereerd. Bovendien
vergemakkelijkt dit ook de supervisie van het signaal, omdat het makkelijker is om een
elektrisch signaal te verifiëren op fouten dan een optisch signaal. Transparante optische
netwerken, waarbij geen tussenliggende elektrische conversies plaatsvinden, zijn veel
moeilijker te verwezenlijken, en zijn momenteel enkel haalbaar voor kleinschalige netwerken,
of als ‘eilandjes’ binnen een groter netwerk. Deze transparante netwerken bieden wel een
aanzienlijk kostvoordeel, wegens de afwezigheid van opto-elektrische convertoren (ook
transponders genoemd).
Naast opto-elektrische conversie, kan een signaal onderweg ook optisch
geconverteerd worden naar een andere golflengte. Het gebruik van golflengte convertoren
zorgt ervoor dat een signaal op elke vezel gebruik kan maken van elke vrije golflengte, wat
een efficiënt gebruik van het netwerk bevordert. Bij de afwezigheid van golflengte
convertoren dient een signaal dezelfde golflengte aan te houden tussen de twee eindpunten,
wat niet alleen de netwerk planning bemoeilijkt maar ook tot blokkering kan leiden, indien
dezelfde golflengte niet vrij is op alle links van het pad.
Het beheer of management van het optisch netwerk staat momenteel nog in zijn
kinderschoenen. Diverse voorstellen werden reeds gemaakt om soortgelijke management
technieken te gebruiken als in andere technologieën (zoals in SDH, maar er wordt ook
gekeken naar internet-georiënteerde protocollen) om een flexibel en onderhoudbare optische
netwerklaag te bekomen. Een overzicht van deze voorstellen kan in de Engelse tekst
teruggevonden worden.
WDM apparatuur kan op een gelijkaardige wijze gecatalogiseerd worden als SDH
apparatuur. WDM lijnsystemen bestaan uit golflengte multiplexers en optische versterkers.
Momenteel kunnen reeds een honderdtal optische kanalen, die elk een 2.5 Gb/s signaal
vervoeren (en in de toekomst wellicht 10 Gb/s), ondersteund worden. Optische add-drop
multiplexers (OADMs), zijn functioneel equivalent aan SDH ADMs, maar werken op
golflengte niveau. Momenteel zijn de meeste commercieel verkrijgbare OADMs weinig
flexibel, en laten enkel toe een beperkte set golflengten te termineren in de OADM. Optische
cross-connects (OXC) kunnen in diverse vormen verwezenlijkt worden. Opake OXCs kunnen
een elektrische schakelmatrix bevatten of een optische schakelmatrix afgebakend door
transponders. Transparante OXCs, met een optische schakelmatrix, kunnen golflengte
conversie ondersteunen of niet. Indien golflengte conversie ondersteund wordt, is de
schakelmatrix veel groter, omdat elke inkomende golflengte moet kunnen gecross-
connecteerd worden naar elke uitgaande golflengte. Naast een grotere schakelmatrix, zijn ook
golflengte convertoren vereist, wat zo een OXC aanzienlijk duurder maakt dan een OXC
zonder golflengte convertoren. Daarom werden ook een aantal tussenvormen uitgedacht, met
beperkte golflengte conversie.
3.4 Pakket geschakelde netwerken
In pakket geschakelde netwerken worden in plaats van semi-permanente connecties,
kleine pakketjes gebruikt om informatie uit te wisselen. Naast de informatie, bevatten de
SAMENVATTING
x
xiv
pakketjes ook voldoende overhead informatie, die de pakketjes toelaten om zelf hun weg door
het netwerk te vinden. In tegenstelling tot transportnetwerken, wordt geen vaste capaciteit in
het netwerk gereserveerd voor pakketjes, maar kan de beschikbare capaciteit gedeeld worden
door alle pakketjes, wat een efficiënter netwerkgebruik toelaat. De twee belangrijkste pakket
geschakelde technologieën in het transportnetwerk worden besproken: de asynchrone transfer
mode (ATM) en het internet protocol (IP). Omdat dit proefschrift enkel planning van SDH en
WDM netwerken behandelt, worden pakket geschakelde netwerken niet verder besproken in
deze samenvatting. Voor meer informatie verwijzen we dan ook graag naar de Engelse tekst.
3.5 Integratie van pakket geschakelde netwerken in transportnetwerken
Een van de redenen waarom pakket geschakelde netwerken vermeld worden in dit
proefschrift, is omwille van de toenemende integratie van pakket geschakelde netwerken in
toekomstige transportnetwerken. Er heerst een consensus dat op de hoogste netwerklagen
(dichtst bij de applicaties), IP het dominante protocol zal blijven voor de komende jaren.
Tegelijkertijd is men het er ook over eens dat WDM de onderste laag van het netwerk voor
zijn rekening zal nemen. Welke tussenliggende lagen er tussen IP en WDM dienen gebruikt te
worden, is een open vraag en talrijke oplossing werden reeds naar voor geschoven. De meest
conservatieve structuur is een meerlagen structuur zoals IP over ATM, over SDH, over
WDM. Daarbij kan gebruik gemaakt worden van bestaande apparatuur en wordt vertrouwd op
de kwaliteit die ATM biedt en de betrouwbaarheid die de herstelmechanismen binnen SDH
bieden. De meest progressieve aanpak is om ATM en SDH te vervangen door een
‘lichtgewicht’ tussenliggende laag, die een enveloppe definieert waarin de IP pakketjes
kunnen verpakt worden. Alle andere netwerkfunctionaliteit wordt dan aan de IP en WDM
laag overgelaten. Dit vereist dat de nodige kwaliteitsgaranties kunnen geboden worden in een
IP netwerk en dat snelle en betrouwbare herstelmechanismen ontwikkeld worden in IP en/of
WDM. Er bestaan eveneens een aantal tussenliggende opties, waarvoor we verwijzen naar de
Engelse tekst.
3.6 Herstelmechanismen voor transportnetwerken
In deze paragraaf wordt een overzicht gegeven van de verschillende
herstelmechanismen die in SDH en WDM netwerken kunnen gebruikt worden. We
onderscheiden protectie- en restauratiemechanismen. Beide mechanismen herstellen het
verkeer dat onderbroken is door een fout (bijvoorbeeld kabelbreuk), door het verkeer te
herrouteren via een alternatieve weg in het netwerk. Protectie herrouteert verkeer via een
vooraf bepaald pad, waarop vrije capaciteit voorzien is, die specifiek voorbehouden is voor
een bepaald foutscenario. Dit maakt de herroutering 100% voorspelbaar. Restauratie
daarentegen, maakt gebruikt van gedeelde vrije capaciteit in het netwerk. Deze capaciteit is
niet toegewezen aan een bepaald foutscenario. Wanneer een fout optreedt, zal een
restauratiealgoritme een alternatief pad berekenen binnen deze vrije capaciteit, voor de
aangetaste connecties. Restauratie is dus minder voorspelbaar, maar laat in principe toe om de
vrije capaciteit in het netwerk beter te benutten. De verschillende protectie- en
restauratiemechanismen worden hierna kort besproken.
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
xv
3.6.1 Trail protectie
Een trail wordt gedefinieerd als de gesuperviseerde overdracht van informatie tussen
twee punten. Bij trail protectie wordt voor elke werkende trail een protectie trail voorzien,
waarop kan teruggevallen worden indien uit de supervisie blijkt dat de performantie van de
werkende trail onvoldoende is. Trail protectie is dus een eind-tot-eind protectiemechanisme
dat zowel in lineaire, ring als vermaasde topologieën kan aangewend worden. Het kan in een
1+1 configuratie toegepast worden, waarbij de protectie trail enkel kan aangewend worden
voor het opvangen van fouten, of in een 1:1 configuratie, waarbij de protectie trail kan
gebruikt worden om informatie van lage prioriteit door te sturen, indien zich geen fouten
voordoen op de werkende trail. In SDH kan het toegepast worden op de LOP, HOP of MS
laag en in WDM op de OCh en OMS laag. Voorbeelden van trail protectie zijn lineaire
multiplex sectie protectie, waarbij een multiplex sectie geprotecteerd wordt door een
parallelle multiplex sectie. Een ander voorbeeld is padprotectie in een vermaasd netwerk,
bijvoorbeeld op de HOP laag. Hierbij wordt een werkende trail tussen twee punten in het
vermaasde netwerk, geprotecteerd door een disjunct gerouteerde protectie trail.
3.6.2 Subnetwerk connectie protectie (SNCP)
Subnetwerk connectie protectie (SNCP) kan een connectie protecteren, ofwel tussen
de eindpunten, ofwel gedeeltelijk (binnen een subnetwerk). SNCP kan ook in verschillende
topologieën toegepast worden, op 1+1 of 1:1 wijze, gelijkaardig aan trail protectie, maar de
supervisie van de connectie is verschillend (zie Engelse tekst). SNCP kan een connectie
protecteren die over verschillende administratieve domeinen wordt getransporteerd. Door alle
subnetwerken onafhankelijk te protecteren, kunnen op deze manier meervoudige fouten van
de werkende connectie, die voorvallen binnen verschillende subnetwerken, op eenvoudige
manier opgevangen worden.
3.6.3 Ring gebaseerde protectiemechanismen
Ringen, opgebouwd uit ADMs, spelen een bijzondere rol voor het toepassen van
herstelmechanismen omdat zij een goedkope en eenvoudige topologie realiseren waarin
steeds 2 disjuncte paden kunnen gevonden worden. Ring gebaseerde protectiemechanismen
kunnen gecatalogiseerd worden aan de hand van een aantal criteria. Het belangrijkste
criterium is het feit of een capaciteitslot op de ring door meerdere niet-overlappende
connecties kan gebruikt worden of niet. Een capaciteitslot is één eenheid van capaciteit
rondom de ring (dus een tijdslot in geval van SDH, een golflengte in geval van WDM). In
geval van een toegewezen protectie ring (Engels: dedicated protection ring, of DPRing) is elk
capaciteitslot rondom de ring toegewezen aan één connectie. In dit geval is de benodigde ring
capaciteit gelijk aan de totale hoeveelheid verkeer die de ring dient te transporteren,
onafhankelijk van het verkeerspatroon. In geval van een gedeelde protectie ring (Engels:
shared protection ring, of SPRing) kan dit capaciteitslot gedeeld worden door meerdere
connecties, wat een beter capaciteitsgebruik van de ring toelaat, afhankelijk van het
verkeerspatroon. Voor de SPRing is geen expliciet protectie pad voorzien voor elke connectie,
maar is de helft van de ringcapaciteit voorbehouden voor protectieverkeer. Deze
protectiecapaciteit kan dan gedeeld worden over verschillende werkend connecties,
naargelang de fout die optreedt.
SAMENVATTING
x
xvi
Een DPRing kan zowel geprotecteerd worden met behulp van 1+1 (of 1:1)
padprotectie of gebruik makend van MS protectie. In het eerste geval (zie Figuur 2) wordt
elke werkende connectie op de ring geprotecteerd door een protectie connectie langs de
andere kant van de ring. Als dusdanig wordt elk capaciteitslot op de ring inderdaad opgebruik
door het werkende en protectie gedeelte van elke connectie. In het tweede geval (zie Figuur 3)
gebeurt de protectie per MS (MS-DPRing), en zullen de ADMs die de fout omsluiten, de
volledig aangetaste MS terugschakelen (Engels: loopback) langs de andere kant van de ring.
Figuur 2a: DPRing geprotecteerd met padprotectie Figuur 2b: DPRing geprotecteerd met padprotectie,
na fout
Figuur 3a: DPRing geprotecteerd met MSP Figuur 3b: DPRing geprotecteerd met MSP, na fout
Ook een SPRing kan zowel geprotecteerd worden met behulp van padprotectie (1:N
in dit geval) of MS protectie (i.e. MS-SPRing), zie respectievelijk Figuur 4 en Figuur 5.
Omdat dezelfde protectiecapaciteit gedeeld wordt door verschillende werkende connecties,
behoeft de SPRing wel een protocol dat het correct gebruik van deze protectiecapaciteit
controleert.
Figuur 4a: SPRing geprotecteerd met padprotectie Figuur 4b: SPRing geprotecteerd met padprotectie,
na fout
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
xvii
Figuur 5a: SPRing geprotecteerd met MSP Figuur 5b: SPRing geprotecteerd met MSP, na fout
Teneinde een grootschalig netwerk uit te bouwen aan de hand van ringen, dringt de
nood zich op om verschillende ringen met elkaar te verbinden. Ringen kunnen
geïnterconnecteerd worden, gebruik makende van 3 mogelijke apparatuurconfiguraties.
Enerzijds kan verkeer dat via de ADM van de ene ring afgehaald wordt, rechtstreeks
verbonden worden met de ADM van de aangrenzende ring. Beide ADMs staan dan rug aan
rug (typisch in hetzelfde gebouw), en de interconnectie gebeurt handmatig (bijvoorbeeld met
een distributie frame). Gezien deze optie weinig flexibiliteit biedt, kan men voor meer
dynamisch verkeer een tweede configuratie verkiezen, waarbij de handmatige interconnectie
vervangen wordt door een cross-connect. Als dusdanig wordt alle verkeer dat van de ring
afgehaald wordt via de ADM, naar de cross-connect gebracht, welke softwarematig kan
geconfigureerd worden om dit verkeer ofwel lokaal te termineren, ofwel door te sluizen naar
een aangrenzende ring, waarvan de ADM ook met deze cross-connect is verbonden. Een
derde, meest flexibele optie, is gelijkaardig aan de tweede optie, maar integreert de ADMs
van de verschillende geïnterconnecteerde ringen in één grote cross-connect. Hierdoor is
slechts één netwerkelement nodig (de cross-connect), en geen ADMs, wat de hoeveelheid
apparatuur en bedrading gevoelig verminderd. De cross-connect dient dan wel alle ring
protectie mechanismen te ondersteunen.
Door een architectuur op te bouwen van meerdere geïnterconnecteerde ringen, kan
een zeer betrouwbaar netwerk bekomen worden. Het is immers mogelijk om meerdere fouten,
die gezamenlijk plaatsgrijpen in verschillende ringen, op een efficiënte manier op te vangen.
De enige zwakke schakel in deze architectuur, is de interconnectie tussen de ringen. Om
hieraan te verhelpen, kunnen ringen verbonden in twee interconnectiepunten. Het drop &
continue protectiemechanisme laat dan toe om, in het geval zich een fout voordoet in een
interconnectiepunt, het verkeer automatisch over te schakelen via het andere
interconnectiepunt. Het principe van drop & continue wordt, zowel voor DPRing als SPRing,
meer uitvoerig besproken in de Engelse tekst.
3.6.4 Restauratie mechanismen
Restauratie berust op een meer autonome bepaling van alternatieve routes in het
geval van een fout, en is daardoor meer flexibel en efficiënt dan protectie wat betreft
capaciteitsgebruik. Restauratie mechanismen kunnen gecatalogiseerd worden aan de hand van
drie belangrijke criteria.
Als eerste criterium beschouwen we wat er gerestaureerd wordt: een pad of een link.
In het geval van padrestauratie, worden alle onderbroken paden als gevolg van een fout
gerestaureerd tussen de eindpunten van elk pad (zie Figuur 6). Linkrestauratie daarentegen,
SAMENVATTING
x
xviii
restaureert alle verkeer op een gefaalde link in één keer, door het verkeer op een alternatieve
route tussen beide eindpunten van de gefaalde link te routeren (zie Figuur 7). Het nadeel van
linkrestauratie is dat het meer capaciteit gebruikt in de nabijheid van de fout en dus minder
efficiënt met capaciteit omspringt. Bovendien is het moeilijk het linkrestauratie concept uit te
breiden naar knoopfouten. Als voordeel heeft linkrestauratie dan wel weer dat het alle verkeer
ineens kan restaureren en daardoor snellere hersteltijden kan realiseren dan padrestauratie.
MUX MUX
M
U
X
M
U
X
M
U
X
MUX MUX
M
U
X
MUX
M
U
X
M
U
X
MUX
M
U
X
M
U
X
Werkende paden
Padrestauratie
Figuur 6: Padrestauratie
MUX MUX
M
U
X
M
U
X
M
U
X
MUX MUX
M
U
X
MUX
M
U
X
M
U
X
MUX
M
U
X
M
U
X
Linkrestauratie
Werkende paden
Figuur 7: Linkrestauratie
Een tweede criterium onderscheidt vooraf berekende en ware tijd restauratie. In het
geval van vooraf berekende restauratie zijn de restauratieroutes voor een aantal vooraf
bepaalde foutscenario's op voorhand berekend, en kan een restauratiepad dus vrij snel opgezet
worden. Daartoe dient wel de toestand van het netwerk zorgvuldig bijgehouden te worden, en
dienen de restauratieroutes herberekend te worden elke keer deze staat wijzigt. Gezien de
berekening van de restauratieroutes in dit geval de hersteltijd niet beïnvloedt, is rekensnelheid
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
xix
minder cruciaal, en kan rekentijd dus ingeruild worden voor optimaliteit van de
restauratieroutes. Bij ware tijd restauratie wordt op het moment van de fout de
restauratieroute berekend. Gezien in dit geval de rekentijd wel deel uitmaakt van de
hersteltijd, zal men dus eerder kiezen voor een snel, maar minder optimaal
restauratiealgoritme. Hoewel ware tijd restauratie typisch trager is dan vooraf berekende
restauratie, kan het wel omspringen met meer foutscenario's en leidt dus tot een
betrouwbaarder netwerk.
Het feit of de restauratie op een gecentraliseerde of gedistribueerde manier
gecontroleerd wordt, is het derde criterium. Bij gecentraliseerde restauratie wordt het
restauratieproces gestuurd vanuit een centrale controle instantie, waar alle actuele
netwerkinformatie beschikbaar is. Alhoewel gecentraliseerde restauratie typisch vrij snel is,
dient de informatie frequent geactualiseerd en verwerkt te worden, wat niet zo schaalbaar is
voor grotere netwerken. Bovendien maakt de afhankelijkheid van één controle instantie, die
alles beheert, deze aanpak minder betrouwbaar. Gedistribueerde restauratie, waarbij de
benodigde informatie verspreid is over het netwerk, lijkt daardoor meer geschikt.
4. Vergelijking van SDH netwerk architecturen
4.1 Inleiding
In hoofdstuk 4, wordt het ontwerp van een lange afstandsnetwerk bekeken, gebruik
makende van SDH en WDM technologie. Dit soort netwerken wordt momenteel gebruikt
door operatoren die zich toespitsen op transport van internationaal verkeer. Het gebruik van
WDM beperkt zich vandaag de dag tot punt-tot-punt transmissie, waarbij de capaciteit op een
optische vezel op een economische wijze drastisch kan verhoogd worden. De
netwerkfunctionaliteit zoals routering, beheer en foutherstel wordt volledig overgelaten aan de
mature SDH laag. Het SDH netwerk, kan op verschillende manieren uitgebouwd worden,
gebruik makend van verschillende herstelmechanismen en verschillende soorten
netwerkapparatuur in tal van configuraties. Teneinde een juiste beslissing te nemen wat
betreft de te adopteren netwerkarchitectuur, is het belangrijk de verschillende alternatieven te
kunnen vergelijken. Hiertoe dient het netwerk eerst ontworpen te worden voor elk
beschouwde herstelmechanisme en elke configuratie van netwerkapparatuur. In dit hoofdstuk
zullen we zo een vergelijkende studie uitvoeren. Gelijkaardige studies werden reeds vermeld
in de literatuur. Onze aanpak verschilt echter op een aantal vlakken van andere studies. Ten
eerste wordt een zeer detaillistisch model gebruikt om de netwerkapparatuur voor te stellen,
en de impact van bepaalde beperkingen op het netwerkontwerp te bestuderen. Ten tweede
worden verschillende mogelijkheden beschouwd om netwerkapparatuur te configureren in een
knoop, terwijl de meeste studies uitgaan van één vaste configuratie van netwerkapparatuur in
een knoop. Ten derde worden verschillende verkeersmatrices gebruikt, met verschillende
granulariteiten van verkeer, om zo de impact van deze granulariteiten te bestuderen op het
ontwerp. In tegenstelling tot sommige andere studies zullen we niet alle mogelijke
herstelmechanismen bekijken, maar enkel de meest interessante: deze zijn padprotectie
(gebruik makend van SNCP) en geïnterconnecteerde MS-SPRingen.
In de hierna volgende paragrafen wordt onze aanpak belicht, die toelaat verschillende
herstelmechanismen en apparatuur configuraties te vergelijken. Eerst worden de gegevens
SAMENVATTING
x
xx
besproken die als invoer voor het planningsvraagstuk dienen. Nadien wordt de
oplossingsmethode geschetst, en finaal worden resultaten naar voor gebracht die toelaten de
verschillende netwerk alternatieven te vergelijken en relevante conclusies te trekken.
4.2 Invoergegevens
In deze paragraaf wordt het planningsprobleem volledig gespecificeerd. We
beginnen met de netwerk topologie, bespreken dan de verkeersmatrices, gevolgd door de
verschillende herstelmechanismen. Ook de verschillende configuraties van netwerkapparatuur
en de kostprijs van deze apparatuur dient gespecificeerd te worden voor aanvang van de
planning.
De netwerk topologie wordt beschreven als een vast stel knopen en tussenliggende
links. De topologie zelf is dus reeds volledig bepaald, enkel de benodigde capaciteit op de
links en in de knopen dient nog bepaald te worden.
De verkeersmatrices omhelzen drie types bidirectioneel verkeer van verschillende
granulariteit, namelijk E1 verkeer (2 Mb/s), E3 verkeer (34 Mb/s) en E4 verkeer (140 Mb/s).
We beschouwen 4 verschillende verkeersmatrices die deze verschillende granulariteiten op
verschillende wijzen combineren. Een eerste matrix bevat verkeer van de 3 beschouwde
granulariteiten, een tweede matrix bevat enkel E1 verkeer, een derde matrix bevat enkel E4
verkeer en een vierde matrix bevat enkel E3 en E4 verkeer. De totale hoeveelheid verkeer is
echter steeds ongeveer even groot voor deze 4 matrices, teneinde relevante conclusies te
kunnen trekken wat betreft hun impact op de netwerk planning.
We beschouwen drie mogelijke herstelmechanismen. Ten eerste wordt padprotectie
beschouwd. Hiertoe wordt het SNCP principe gebruikt tussen de eindpunten van de connectie.
Belangrijk is hierbij dat werkend en protectie pad zowel knoop- als linkdisjunct zijn om elke
enkelvoudige knoop- of linkfout te kunnen overleven. Ten tweede wordt MS-SPRing
beschouwd. Om dit principe toe te passen op een groot netwerk, dienen verschillende ringen
geïdentificeerd te worden in het netwerk, die met elkaar verbonden zijn, en alle knopen van
het netwerk dekken, teneinde alle verkeer te kunnen transporteren. Ten derde bekijken we een
combinatie van SNCP padprotectie en MS-SPRing. Dit betekent dat een connectie ofwel door
SNCP padprotectie ofwel door MS-SPRing wordt geprotecteerd. Hiertoe kunnen dus een
beperkt aantal ringen in het netwerk geïdentificeerd te worden, die niet noodzakelijk alle
knopen dienen te bedekken. Connecties die niet via MS-SPRing kunnen worden
geprotecteerd, kunnen dan via SNCP padprotectie worden geprotecteerd.
Voor de netwerkapparatuur dient onderscheid gemaakt te worden tussen link- en
knoopapparatuur. De linkapparatuur bestaat uit een WDM systeem van 40 golflengten,
gebruik makende van transponders. Op elke golflengte wordt een STM-16 signaal
getransporteerd. We beschouwen geen afzonderlijke SDH lijnsystemen, maar geïntegreerde
lijnsystemen in de SDH ADMs.
Voor de knoopapparatuur beschouwen we volgende netwerkelementen, die in
verschillende configuraties kunnen aangewend worden:
• LO ADMs (zie Figuur 8), die zowel LO als HO verkeer kunnen afhalen en toevoegen
(add/drop) tussen twee inkomende en uitgaande STM-16 links. Deze ADMs
ondersteunen SNCP (op de LO en HO laag) en MS-SPRing (enkel op de HO laag).
Deze ADM is beperkt in de hoeveelheid verkeer die kan afgehaald en toegevoegd
worden: 8 equivalente VC-4s wat betreft HO verkeer (zowel voor SNCP en MS-
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
xxi
SPRing) en LO verkeer dat geprotecteerd is met MS-SPRing. Voor LO verkeer dat
geprotecteerd is met behulp van SNCP, kunnen slechts en 4 equivalente VC-4s
behandeld worden, omdat telkens 2 poorten naar de HO matrix vereist.
.
.
..
.
.
HO matrix
LO matrix
STM-16STM-16
16 STM-1 poorten
LO geprotecteerd
verkeer
HO geprotecteerd
verkeer
Niet
geprotecteerd
verkeer
8 STM-1
poorten
8 trib.
poorten
Figuur 8: Lagere orde (LO) ADM
• HO ADMs (zie Figuur 9), die enkel HO verkeer kunnen behandelen tussen de
inkomende en uitgaande STM-16 links. Deze ADM kan 16 VC-4s afhalen en toevoegen
tussen deze STM-16 links. De HO ADM ondersteunt zowel SNCP en MS-SPRing op
HO niveau.
.
.
..
.
.
HO matrix STM-16STM-16
16 STM-1 poorten
HO
geprotecteerd
verkeer
Niet
geprotecteerd
verkeer
16 trib.
poorten
Figuur 9: Hogere orde (HO) ADM
• LO multiplexers, die LO verkeer multiplexeren in een STM-1 signaal
• DXC 4/3/1, een cross-connect die tot op LO niveau werkt, en toelaat om LO verkeer te
segregeren en consolideren tussen inkomende en uitgaande STM-1 poorten. Bovendien
wordt SNCP ondersteund, zowel op LO als op HO niveau.
• DXC 4/4, een cross-connect die op HO niveau werkt, en dus toelaat om HO verkeer uit
te wisselen tussen verschillende inkomende en uitgaande STM-1 poorten. Bovendien
wordt SNCP ondersteund op HO niveau.
• Regeneratoren, die een STM-16 signaal versterken en digitaal herschapen om onderweg
opgelopen vervormingen te herstellen.
SAMENVATTING
x
xxii
Op basis van bovenstaande netwerkelementen, beschouwen we nu 5 verschillende
zogenaamde knoopscenario's, die elk een andere combinatie van deze netwerkelementen
gebruiken in de knopen van het netwerk. Voor elk scenario wordt een knoop onderverdeeld in
2 gedeeltes, een flexibiliteitgedeelte, en een toegangsgedeelte. Het flexibiliteitgedeelte zorgt
ervoor dat verkeer kan gecross-connecteerd worden tussen alle inkomende en uitgaande STM-
16 links van de knoop zonder dat daarbij blokkering optreedt. De hoeveelheid apparatuur
nodig in het flexibiliteitgedeelte van een knoop hangt dus enkel af van het aantal STM-16
links waarmee deze knoop verbonden is. Het toegangsgedeelte zorgt er anderzijds voor dat
alle verkeer dat start of eindigt in deze knoop op het netwerk kan geplaatst worden. De
hoeveelheid apparatuur nodig in dit toegangsgedeelte hangt dus af van de verkeersmatrix, met
name van de hoeveelheid verkeer die deze knoop als begin- of eindpunt heeft.
Hieronder bespreken we kort de 5 beschouwde knoopscenario's. Voor meer details
en de relevante figuren verwijzen we graag naar de Engelse tekst.
Het eerste beschouwde knoopscenario, NS-1, gebruikt enkel LO ADMs om toegang
te krijgen tot het verkeer op een STM-16 link die getermineerd wordt in de knoop. De LO
ADM wordt gebruikt in het toegangsgedeelte, zowel om LO als HO verkeer af te halen of toe
te voegen. LO verkeer wordt in de LO ADM gegroomd op basis van de bestemmingsknoop.
Onderweg wordt het verkeer als HO verkeer behandeld. In het flexibiliteitgedeelte worden
eveneens LO ADMs gebruikt om verkeer van een STM-16 link af te halen, en dit manueel te
verbinden met een LO ADM die toegang geeft tot een andere STM-16 link.
Het tweede beschouwde knoopscenario, NS-2, gebruikt zowel LO ADMs als HO
ADMs om toegang te krijgen tot het verkeer op een STM-16 link. In het toegangsgedeelte
worden LO ADMs gebruikt voor lokaal LO verkeer, dat opnieuw gegroomd wordt op basis
van de bestemmingsknoop. De HO ADM wordt in het toegangsgedeelte gebruikt voor lokaal
HO verkeer. In het flexibiliteitgedeelte worden enkel HO ADMs gebruikt, omdat deze minder
restrictief zijn. De verbinding tussen 2 ADMs in dit flexibiliteitgedeelte gebeurt opnieuw
manueel.
Het derde beschouwde knoopscenario, NS-3, gebruikt HO ADMs in combinatie met
LO multiplexers en een DXC 4/4. In het toegangsgedeelte kan de HO ADM rechtstreeks
gebruikt worden om lokaal HO verkeer toe te voegen aan een STM-16 link. Lokaal LO
verkeer daarentegen, dient eerst door een LO multiplexer te gaan vooraleer het via de DXC
4/4 en HO ADM op de STM-16 link kan geplaatst worden. Deze LO multiplexer groomt het
LO verkeer op basis van bestemmingsknoop. In het flexibiliteitgedeelte worden eveneens HO
ADMs gebruikt om verkeer tussen STM-16 links uit te wisselen. De verbinding tussen
verschillende HO ADMs gebeurt nu niet langer manueel, maar de ADMs zijn met de DXC
4/4 verbonden zodat deze verbinding niet langer statisch is, maar flexibel.
Het vierde beschouwde knoopscenario, NS-4, gebruikt HO ADMs in combinatie met
een DXC 4/3/1. In het toegangsgedeelte kan de HO ADM rechtstreeks gebruikt worden om
lokaal HO verkeer toe te voegen aan een STM-16 link. Lokaal LO verkeer kan via de DXC
4/3/1 geconsolideerd worden en vervolgens via de HO ADM op de gepaste STM-16 link
geplaatst worden. Aangezien deze DXC 4/3/1 in elke tussenliggende knoop aanwezig is, kan
verkeer dus op LO niveau gecross-connecteerd worden en hoeft het niet langer gegroomd
worden op basis van de bestemming. In het flexibiliteitgedeelte worden eveneens HO ADMs
gebruikt om verkeer van een STM-16 link af te halen en naar de DXC 4/3/1 te brengen alwaar
het LO verkeer op gepaste wijze kan gesegregeerd en geconsolideerd worden teneinde de
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
xxiii
vulling van het netwerk te optimaliseren. De cross-connecting van HO verkeer in het
flexibiliteitgedeelte gebeurt niet via de DXC 4/3/1, maar manueel tussen de HO ADMs.
Het vijfde en laatste beschouwde knoopscenario gebruikt HO ADMs in combinatie
met LO multiplexers en geen DXCs. In het toegangsgedeelte kan de HO ADM rechtstreeks
gebruikt worden om lokaal HO verkeer toe te voegen aan een STM-16 link. Lokaal LO
verkeer daarentegen, dient eerst door een LO multiplexer te gaan vooraleer het via de HO
ADM op de STM-16 link kan geplaatst worden. Deze LO multiplexer groomt het verkeer
opnieuw op basis van bestemmingsknoop. In het flexibiliteitgedeelte worden de HO ADMs
opnieuw gebruikt om manueel verkeer tussen verschillende STM-16 links uit te wisselen.
Finaal is ook een kostmodel nodig om de verschillende alternatieven te evalueren.
Voor de gebruikte kostgegevens van de netwerkapparatuur verwijzen we naar de Engelse
tekst.
4.3 Netwerk ontwerp
Teneinde de verschillende herstelmechanismen en knoopscenario's met elkaar te
kunnen vergelijken, dient het netwerk ontworpen te worden voor elke mogelijke combinatie.
Aangezien we 3 mogelijke herstelmechanismen en 5 knoopscenario's beschouwen, dient het
netwerk ontworpen te worden voor 15 mogelijke gevallen. Om geen 15 verschillende
ontwerpstechnieken te moeten ontwikkelen, splitsen we het netwerkontwerp op in 4 fazen. In
de eerste fase wordt het verkeer gerouteerd. Deze routering is enkel afhankelijk van het
herstelmechanisme, dus dienen we 3 verschillende routeringsalgoritmes te ontwerpen. In de
tweede fase worden de bekomen routes gebruikt om te evalueren hoeveel verkeer over elke
link loopt en hoe het verkeer binnen een knoop eruit ziet. In een derde fase wordt deze
informatie uit de tweede fase aangewend om de capaciteit op elke link te berekenen, evenals
de hoeveelheid netwerkapparatuur in elke knoop. In deze fase dienen we 5 gevallen te
onderscheiden, overeenkomstig met de 5 knoopscenario's. In een laatste faze wordt het
netwerk geëvalueerd op basis van een aantal metrieken. We zullen nu elke fase in iets meer
detail belichten.
4.3.1 Routering
Wat routering betreft, dient een onderscheid gemaakt te worden tussen de 3
beschouwde herstelmechanismen.
Voor SNCP padprotectie dienen per connectie 2 knoop- en linkdisjuncte paden
gevonden te worden in het netwerk. Hiervoor gebruiken we een bestaand algoritme uit de
literatuur dat de kortste kring tussen 2 knopen in een netwerk vindt.
Voor het herstelmechanisme gebruik makend van MS-SPRing volstaat het niet om
het verkeer te routeren, bovendien dient een stel geïnterconnecteerde ringen gevonden te
worden in het netwerk. In onze aanpak wordt het verkeer eerst gerouteerd volgens het kortste
pad in het onderliggende vermaasde netwerk. In een tweede fase worden deze routes gebruikt
om de best passende ringen op te sporen die dit verkeer kunnen dragen.
Wanneer MS-SPRing en SNCP gecombineerd worden, dienen de ringen niet langer
het volledige netwerk te bedekken. In een eerste fase zullen we een beperkt stel MS-SPRingen
identificeren die zo goed mogelijk bij het opgegeven verkeerspatroon passen, en een aantal
combinaties van deze ringen zullen uitgeprobeerd worden. De routering gebeurt eerst over de
ringen, waarbij alle inter- en intra-ring verkeer opgezet wordt. Daarna wordt alle verkeer dat
SAMENVATTING
x
xxiv
niet over de ringen kan worden gerouteerd, geprotecteerd met behulp van SNCP (waarbij de
routes op dezelfde manier bepaald worden als in het geval waar enkel SNCP wordt gebruikt).
4.3.2 Evaluatie
Op basis van de routes bekomen uit de hierboven besproken routering kan een
evaluatie gemaakt worden van de hoeveelheid verkeer die over een link loopt. Hierbij dient
onderscheid gemaakt te worden naargelang het herstelmechanisme en de granulariteit van het
verkeer. In deze fase berekenen we dus hoeveel HO en LO verkeer er op elke link loopt,
zowel voor SCNP en MS-SPRing.
Een zelfde oefening kan gemaakt worden voor de knopen. De hoeveelheid informatie
is hier echter veel groter. Opnieuw dient onderscheid gemaakt te worden naargelang het
herstelmechanisme en de granulariteit van het verkeer. Bovendien dient ook onderscheid
gemaakt te worden tussen verkeer dat door de knoop passeert (en tussen welke twee links dit
gebeurt) en het verkeer dat lokaal wordt afgehaald (en vanwaar het komt).
De exacte informatie die we uit deze evaluatie fase verkrijgen is in detail beschreven
in de Engelse tekst.
4.3.3 Dimensionering
Gebaseerd op de informatie uit de evaluatie fase kan voor elk knoopscenario de
vereiste link- en knoopapparatuur berekend worden. Voor de links dient enkel een
onderscheid gemaakt te worden tussen de gevallen waarbij verkeer tussen de eindpunten
wordt gegroomd (NS-1, 2, 3 en 5) en het geval waarbij LO verkeer optimaal kan gesegregeerd
en geconsolideerd worden in tussenliggende knopen (NS-4). Voor de knopen, dient een aparte
berekening uitgevoerd te worden voor elk knoopscenario en herstelmechanisme. Voor de
formules die toelaten de vereiste link- en knoopapparatuur te bepalen verwijzen we naar de
Engelse tekst, waarin een detaillistisch mathematisch model besproken wordt.
4.3.4 Metrieken
Teneinde de verschillende alternatieven met elkaar te vergelijken, zijn een aantal
representatieve metrieken vereist. Een eerste belangrijke metriek is de totale installatiekost
van het netwerk. Deze kost kan opgesplitst worden in een link- en knoopkost. De linkkost
omhelst de kost van de WDM apparatuur, transponders, versterkers en regeneratoren. De
knoopkost bevat de kost van SDH multiplexers, ADMs en DXCs. Een tweede belangrijke
metriek is de vullingsgraad van het netwerk. Dit is de verhouding tussen de gebruikte
capaciteit en de geïnstalleerde capaciteit. We onderscheiden ook een primaire vullingsgraad,
die de verhouding van werkend verkeer (dus geen protectieverkeer) ten opzichte van de
geïnstalleerde capaciteit weergeeft. Voor MS-SPRing is het enkel nuttig van een primaire
vullingsgraad te spreken, omdat er geen expliciet protectieverkeer is.
4.4 Resultaten
Hieronder worden enkele resultaten samengevat die behaald werden op een netwerk
met 25 knopen en 29 links, en een groter netwerk van 66 knopen en 81 links. We bespreken
eerst afzonderlijk de resultaten voor SNCP en MS-SPRing, en maken nadien een vergelijking
tussen beide mechanismen en het hybride mechanisme.
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
xxv
4.4.1 Resultaten voor SNCP
We bespreken eerst resultaten voor het 25-knopen netwerk en breiden ze daarna uit
voor het grotere netwerk.
De link- en knoopkost (relatief ten opzichte van het duurste knoopscenario) worden
uitgezet in Figuur 10. Wat de linkkost betreft, zijn de knoopscenario's NS-1, 2, 3 en 5
evenwaardig. NS-4 heeft een 12% lagere linkkost dankzij het feit dat verkeer op LO niveau
kan samengenomen worden in alle tussenliggende knopen, en niet enkel tussen de
eindknopen. Dit heeft een sterk effect op de vullingsgraad van het netwerk, die stijgt van 54%
naar 90%. De primaire vullingsgraad stijgt op zijn beurt van 20% naar 30%. Uit deze
resultaten blijkt eveneens dat een groot gedeelte van de netwerkcapaciteit gebruikt wordt voor
protectie.
De knoopkost is het hoogst voor NS-1, omdat hier enkel gebruik gemaakt wordt van
LO ADMs, die zeer restrictief zijn in de hoeveelheid verkeer die kan afgehaald worden. Door
de LO ADMs te combineren met HO ADMs, zoals in NS-2, waarbij de LO ADM enkel
gebruikt worden voor LO verkeer in het toegangsgedeelte van de knoop, daalt de knoopkost
met 32% ten opzichte van NS-1. NS-5, dat enkel HO ADMs gebruikt is nog iets goedkoper.
De scenario's NS-3 en NS-4, die gebruik maken van DXCs zijn dan weer iets duurder. NS-3 is
22% duurder dan NS-2, en NS-4 nog eens 10% duurder dan NS-3 door de zeer dure DXC
4/3/1. Deze 2 scenario's hebben wel een hogere flexibiliteit dan de scenario's zonder DXC.
Uiteindelijk is de totale kost het laagst voor NS-5 en NS-2 en het hoogst voor NS-1.
De extra knoopkost van NS-4 weegt niet op tegen de besparingen in linkkost voor het 25-
knopen netwerk. Voor het grotere netwerk daarentegen is het wel economisch om verkeer op
LO niveau samen te nemen in tussenliggende knopen, zoals in NS-4. In dit geval is NS-4
immers het goedkoopste scenario, zowel qua knoop- en linkkost.
0%
20%
40%
60%
80%
100%
NS-1 NS-2 NS-3 NS-4 NS-5
Relatieve kost
Knoopkost
Linkkost
Figuur 10: Link- en knoopkost voor SNCP knoopscenario's
Wat de granulariteit van de verkeersmatrix betreft blijkt het dat de linkkost van NS-4
vrijwel onafhankelijk is van de granulariteit van het verkeer, maar enkel afhankelijk van de
SAMENVATTING
x
xxvi
hoeveelheid verkeer. Dit is dankzij het feit dat LO verkeer in tussenliggende knopen optimaal
kan samengenomen worden, waardoor de vulling van het netwerk steeds zeer goed is. Voor
de andere knoopscenario's blijkt dat bij een zelfde totale hoeveelheid verkeer, de linkkost
hoger is naarmate meer en vooral kleinere granulariteiten aanwezig zijn in de verkeersmatrix.
Zo heeft een netwerk ontworpen voor verkeersmatrix met E1,E3 en E4 verkeer een hogere
linkkost dan een netwerk met een verkeersmatrix met een equivalente hoeveelheid E4 verkeer
(en geen E1 en E3 verkeer). Naarmate meer E1 verkeer aanwezig is stijgt ook knoopkost voor
de verschillende knoopscenario's. NS-2 is het voordeligst indien enkel E1 verkeer aanwezig
is, omdat enkel LO ADMs gebruikt worden in het toegangsgedeelte en enkel HO ADMs in
het flexibiliteitgedeelte. In het geval enkel E1 verkeer aanwezig is, valt NS-4 zeer duur uit
omdat alle verkeer door de DXC 4/3/1 moet. Indien de hoeveelheid E1 verkeer minder groot
is, is NS-5 te verkiezen.
4.4.2 Resultaten voor MS-SPRing
Ook hier bespreken we eerst resultaten voor het 25-knopen netwerk en breiden ze
daarna uit voor het grotere netwerk.
De link- en knoopkost (relatief ten opzichte van het duurste knoopscenario) worden
uitgezet in Figuur 11. Voor MS-SPRing zijn NS-1, 2, 3 en 5 opnieuw equivalent qua linkkost.
NS-4 is in dit geval 22% voordeliger omdat een betere vullingsgraad bekomen wordt. Deze
vullingsgraad is belangrijk voor MS-SPRing, omdat elke ring moet ontworpen worden op
basis van de zwaarst belaste link. Door deze zwaarst belast link efficiënter te gebruiken, daalt
dus de kost op alle links van de ring.
Door de verbeterde vulling is het aantal HO ADMs benodigd voor NS-4 ongeveer de
helft dan voor NS-2, 3 en 5. Daardoor is NS-4 het voordeligste knoopscenario op basis van de
knoopkost en totale kost. Het nadeel van NS-4 is echter dat het een lagere betrouwbaarheid
biedt. De protectie gebeurt immers op het HO niveau (in de MS-SPRing), terwijl verkeer ook
op LO niveau gerouteerd wordt (in de DXC 4/3/1). Als gevolg kan een fout van de DXC 4/3/1
niet omzeild worden. De knoopkost van NS-5 is ongeveer gelijk aan deze van NS-4, omdat de
meerkost aan HO ADMs in NS-4 de kost van de DXC 4/3/1 ongeveer benaderdt. De linkkost
van NS-5 is echter hoger, en dus ook de totale kost, waardoor NS-5 het tweede voordeligste
knoopscenario is. Het derde voordeligste knoopscenario is NS-3, dat op zijn beurt dan weer
een hogere flexibiliteit biedt dankzij de DXC 4/4. NS-1 en NS-2 zijn minder geschikt voor
MS-SPRing omdat de LO ADMs te restrictief zijn ten opzichte van de HO ADMs voor MS-
SPRing.
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
xxvii
0%
20%
40%
60%
80%
100%
NS-1 NS-2 NS-3 NS-4 NS-5
Relatieve kost
Knoopkost
Linkkost
Figuur 11: Knoop- en linkkost voorMS-SPRing knoopscenario's
Vergelijkbare conclusies worden bekomen op het grotere netwerk. Hier is het
economische voordeel van NS-4 zelfs nog meer uitgesproken.
Wat de granulariteit van de verkeersmatrix betreft, blijkt opnieuw dat de linkkost van
NS-4 vrijwel onafhankelijk is van de granulariteit, maar enkel afhankelijk van de hoeveelheid
verkeer. Voor de andere knoopscenario's blijkt dat bij een zelfde totale hoeveelheid verkeer,
de linkkost tot 20% hoger is dan bij NS-4 naarmate meer en vooral kleinere granulariteiten
aanwezig zijn in de verkeersmatrix. Naarmate meer E1 verkeer aanwezig is, stijgt opnieuw de
knoopkost voor de verschillende knoopscenario's. NS-1 is in alle gevallen het duurst en NS-2
het tweede duurste, wat opnieuw onderstreept dat de LO ADM niet geschikt is voor gebruik
met MS-SPRing. Indien een mix van verschillende granulariteiten aanwezig is heeft NS-4 de
laagste knoopkost, zoniet heeft NS-5 de laagste knoopkost. Dankzij de lagere linkkost heeft
NS-4 opnieuw de laagste totale kost voor alle verkeersmatrices. Opnieuw dient onderstreept te
worden dat dit kostvoordeel dient afgewogen te worden tegen het feit dat fouten van de DXC
4/3/1 niet kunnen opgevangen worden in dit scenario.
4.4.3 SNCP versus MS-SPRing
In deze paragraaf worden de resultaten voor SNCP en MS-SPRing met elkaar
vergeleken, en evalueren we ook de resultaten van het derde herstelmechanisme, dat SNCP en
MS-SPRing combineert. In deze paragraaf concentreren we ons uitsluitend op de
knoopscenario's NS-4 en NS-5, omdat deze uit voorgaande resultaten de meest interessante
bleken te zijn. We bekijken eerst het 25-knopen netwerk (zie Figuur 12), en breiden daarna de
resultaten uit voor het grotere netwerk.
Voor het 25-knopen netwerk blijkt SNCP de laagste linkkost te hebben, zowel voor
NS-4 als NS-5. Ondanks het feit dat MS-SPRing protectiecapaciteit kan delen onder
verschillende werkende connecties, wat niet mogelijk is bij 1+1 SNCP, heeft SNCP een
lagere linkkost. Dit komt voornamelijk omdat het verkeerspatroon op de ringen niet in het
voordeel is van MS-SPRing (zie hoofdstuk 5), omdat een beperkt aantal knopen (zoals de
SAMENVATTING
x
xxviii
grote hoofdsteden en ring interconnectieknopen) het gros van het verkeer aantrekken.
Bovendien dient voor MS-SPRing elke ring gedimensioneerd te worden op basis van de
meest belaste link, terwijl voor SNCP elke link afzonderlijk kan geoptimaliseerd worden.
Wat de knoopkost betreft, is MS-SPRing dan weer de goedkoopste optie voor NS-4
en NS-5. Hiervoor zijn 2 redenen. Enerzijds is de HO ADM niet blokkerend voor MS-
SPRing, wat niet het geval is voor SNCP, zodat soms 2 HO ADMs nodig zijn op een STM-16
link voor SNCP (wat niet het geval is voor MS-SPRing). Ten tweede dient voor elke SNCP
connectie een werkend en protectie pad opgezet te worden, terwijl voor MS-SPRing enkel een
werkend pad nodig is en de protectie in de ringen gebeurt. De 2 paden die in het geval van
SNCP nodig zijn, zorgen dus voor meer verkeer dat in de knopen moet gecross-connecteerd
worden en dus een extra kost.
0%
20%
40%
60%
80%
100%
NS-4 NS-5 NS-4 NS-5 NS-5 + D&C
Relatieve kost
Knoopkost
Linkkost
SNCP MS-SPRing
Figuur 12: Kost SNCP en MS-SPRing voor het 25-knopen netwerk
Voor NS-4 is de totale kost uiteindelijk 20% lager voor MS-SPRing dan voor SNCP
(opnieuw met de belangrijke bemerking omtrent de verminderde betrouwbaarheid). Voor NS-
5 is MS-SPRing 3% goedkoper dan SNCP indien geen drop & continue toegepast wordt en
6% duurder indien wel drop & continue toegepast wordt. De hybride architectuur waarbij
SNCP en MS-SPRing gecombineerd worden, is voor NS-4 23% goedkoper dan SNCP en 3%
goedkoper dan MS-SPRing (zie Figuur 13). Voor NS-5 is de kostreductie van het hybride
ontwerp ten opzichte van SNCP 5% en ten opzichte van MS-SPRing (met drop & continue)
11%.
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
xxix
0%
20%
40%
60%
80%
100%
NS-4 NS-5
Relatieve totale kost
SNCP
MS-SPRing
Hybride
Figuur 13: Vergelijking SNCP, MS-SPRing en hybride architectuur voor het 25-knopen netwerk
Voor het grotere netwerk is NS-4 het goedkoopste scenario, zowel voor MS-SPRing
als SNCP. Voor NS-4 is MS-SPRing in totaal slechts 1.3% goedkoper. Voor NS-5, is MS-
SPRing dan weer 38% duurder, omdat het moeilijk is goed gevulde ringen te vinden indien
verkeer niet kan samengevoegd worden op LO niveau in tussenliggende knopen. De hybride
architectuur is opnieuw goedkoper dan SNCP en MS-SPRing, zowel voor NS-4 en NS-5.
Voor NS-4 is de kostreductie van het hybride ontwerp ten opzichte van SNCP 11 % en ten
opzichte van MS-SPRing 10%. Voor NS-5 is het hybride ontwerp 2.2% goedkoper dan SNCP
en 29% goedkoper dan MS-SPRing.
5. Planning en configuratie van een WDM ring
5.1 Inleiding
Daar waar in het vorige hoofdstuk WDM enkel aangewend werd om de capaciteit op
een vezel op te drijven, bekijken we in dit hoofdstuk ook routering en herstelmechanismen in
de optische laag. Routering en herstel in de optische laag, komt veelal goedkoper uit dan in
een hogere elektrische laag, omdat een grotere hoeveelheid verkeer ineens kan behandeld
worden. We bestuderen in dit opzicht meer in het bijzonder een WDM ring. Deze bestaat uit
optische add/drop multiplexers (OADMs), die in een gesloten lus verbonden worden met
behulp van optische vezelparen. De OADMs laten toe om lokaal verkeer, in de vorm van
golflengten, toe te voegen (multiplexeren) of af halen (demultiplexeren) aan/van de
aangrenzende vezelparen op de ring. Bovendien kunnen de overige golflengten transparant
door de OADM passeren. De OADMs die momenteel beschikbaar zijn, zijn statisch en laten
enkel toe om een stel vooraf gedefinieerde golflengten af te halen, en ondersteunen bovendien
geen herstelmechanismen. In de nabije toekomst zullen echter ook reconfigureerbare OADMs
beschikbaar worden, die een stel golflengten naar keuze kunnen afhalen, en ook
herstelmechanismen ondersteunen. Zowel voor WDM ringen gebaseerd op statische en
SAMENVATTING
xl
reconfigureerbare OADMs dienen een aantal plannings- en configuratieproblemen opgelost te
worden. Dit is het onderwerp van hoofdstuk 5. In de volgende paragraaf worden de diverse
plannings- en configuratieproblemen aangekaart. Daarna zal voor elk probleem een
oplossingsmethode voorgesteld worden, en zullen enkele relevante resultaten gepresenteerd
worden.
5.2 Planningsproblemen
5.2.1 Lange termijn planning
Op de lange termijn dringen zich een aantal strategische beslissingen op wat betreft
de architectuur van de WDM ring. Een eerste belangrijke beslissing, is de keuze van het type
ring: een DPRing of een SPRing. Deze beslissing is belangrijk, omdat zij de interne structuur
van de OADM vastlegt, welke achteraf moeilijk kan gewijzigd worden. Een aantal factoren
spelen mee in het nemen van deze beslissing. Een DPRing is makkelijker te implementeren en
zal dus sneller commercieel verkrijgbaar zijn. Ook de OADMs zullen dus waarschijnlijk
goedkoper zijn dan deze van een SPRing. Anderzijds gebruikt de SPRing typisch minder
capaciteit. Het capaciteitsgebruik is hierbij wel afhankelijk van het verkeerspatroon. Met
name verkeer tussen aangrenzende knopen, leidt tot zeer efficiënt capaciteitsgebruik voor de
SPRing. De capaciteitsvereisten van de DPRing zijn daarentegen onafhankelijk van het
verkeerspatroon, maar enkel van de hoeveelheid verkeer. De DPRing is dus meer robuust
indien enkel de hoeveelheid verkeer accuraat kan voorspeld worden of indien het
verkeerspatroon snel wijzigt.
Een andere lange termijn beslissing, is de keuze van de ring topologie. Hierbij dienen
een aantal ringen gedefinieerd te worden in een netwerk die toelaten om verkeer op een
efficiënte en betrouwbare manier te routeren over deze ringen. De planning van
ringnetwerken is het onderwerp van hoofdstuk 6. In hoofdstuk 5 kijken we enkel naar één
enkele ring. De planningsproblemen die zich stellen op één enkele ring komen uiteraard terug
voor ringnetwerken bestaande uit meerdere ringen.
5.2.2 Ring dimensionering
Op de middellange termijn, dient op basis van het gekozen ring type bepaald te
worden hoe de ring moet uitgerust worden om het opkomende verkeer zo efficiënt mogelijk te
routeren. Efficiënte routering houdt in hoofdzaak in dat we het aantal benodigde golflengten
op de ring wensen te minimaliseren. Hiertoe dienen twee problemen opgelost te worden:
routering en golflengte toekenning.
Het routeringsprobleem stelt zich enkel bij de SPRing. Op een DPRing worden
immers het werkend en protectie pad langs een verschillende kant van de ring gerouteerd, en
gezien er slechts 2 mogelijkheden zijn, wordt dus een volledige golflengte rondom de ring
ingenomen. Op een SPRing wordt enkel werkend verkeer gerouteerd, ofwel in wijzerzin of in
tegenwijzerzin. Er wordt geen expliciet protectie pad voorzien, maar de helft van de
ringcapaciteit wordt voor protectie doeleinden voorbehouden. De ringcapaciteit wordt dus
bepaald door de link van de ring die het meeste werkende verkeer draagt. De totale benodigde
capaciteit (werkend en protectie) op elke link van de ring bedraagt dan 2 maal het verkeer op
deze meest beladen link. Minimalisatie van het aantal golflengten op de ring, houdt dus in dat
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
li
de hoeveelheid verkeer op de meest beladen ring dient geminimaliseerd te worden. Dit
probleem wordt in paragraaf 5.3 in meer detail behandeld.
Ook wat golflengte toekenning betreft, stellen zich geen problemen voor de DPRing.
Aangezien werkend en protectie pad een volledige golflengte innemen, is de toekenning van
golflengten eenvoudig en kunnen geen conflicten optreden. Voor de SPRing, moet aan elk
werkend pad een golflengte toegewezen worden, zodanig dat werkende paden die dezelfde
link delen een verschillende golflengte hebben. Tezelfdertijd wensen we het totaal aantal
benodigde golflengten te minimaliseren door niet overlappende paden zoveel mogelijk te
laten gebruik maken van dezelfde golflengte. Dit probleem kan in principe op eenvoudige
wijze opgelost worden door in elke OADM golflengte convertoren te voorzien, die de
golflengte van een pad link per link kunnen aanpassen naargelang de vrije golflengten op die
link. Het gebruik van deze convertoren houdt uiteraard wel een meerkost in, die niet steeds
gewettigd is. Indien geen golflengte convertoren aanwezig zijn, dient aan elk pad een unieke
golflengte toegekend te worden. Deze golflengte toekenning is minder triviaal, in het
bijzonder voor willekeurige verkeerspatronen. Dit probleem wordt verder in paragraaf 5.4 in
meer detail behandeld.
De minimalisatie van het aantal golflengten is belangrijk teneinde de kost van
transmissie faciliteiten (vezels, versterkers, …) te minimaliseren. Daarnaast speelt ook de
OADM kost een rol. Zoals reeds gezegd is de OADM voor de DPRing typisch goedkoper dan
voor een SPRing. Afhankelijk van de kost verhoudingen, kan een DPRing dus ondanks een
slechter gebruik van capaciteit toch nog voordeliger uitvallen dan een SPRing. Zo zijn
bijvoorbeeld in regionale netwerken de afstanden klein, en is de impact van de OADM kost
op de totale kost groter dan de transmissie kost. In dit geval is de DPRing dus te verkiezen. Er
kan ook overwogen worden om de SPRing en DPRing te combineren in één architectuur. Dit
kan bijvoorbeeld gebeuren door meerdere ringen bovenop elkaar te stapelen, wat nodig is
indien één ring niet volstaat om alle verkeer te transporteren tussen een stel knopen. In dit
geval kunnen enkele ringen in deze stapel van het DPRing type zijn, en andere van het
SPRing type. Zo kan elk ring type aangewend worden voor het verkeerspatroon waarvoor het
best geschikt is. Hoe de kost van zo een hybride architectuur kan geoptimaliseerd worden,
wordt in paragraaf 5.5 in meer detail behandeld.
5.2.3 SDH gebaseerde WDM ringen
In SDH gebaseerde WDM ringen worden vaste OADMs gebruikt, zonder
herstelmechanismen. Het betreft hier dus een WDM ring met enkel werkende capaciteit,
vergelijkbaar met een WDM ring opgebouwd uit punt-tot-punt lijnsystemen. Het enige
verschil hierbij is dat doorgaand verkeer transparant door de OADM kan passeren. Deze
WDM ring kan dan gebruikt worden om meerdere SDH ringen te dragen (zie Figuur 14). Dit
gebeurt door één golflengte rondom de ring te gebruiken als een virtuele vezel, en dus in elke
OADM ook een SDH ADM te voorzien per golflengte (= SDH ring). Zowel protectie als
routering gebeurt dus in de SDH laag, en ook de vraagmatrix wordt in de SDH laag opgesteld
(als aantal VC-4s tussen 2 knopen). Indien meerdere SDH ringen gedragen worden door de
zelfde WDM ring, dient men niet noodzakelijk in elke knoop evenveel SDH ADMs te
voorzien als er SDH ringen zijn. In sommige knopen vereisen bepaalde ringen geen SDH
ADM omdat alle verkeer op de SDH ring de knoop passeert, en er dus geen verkeer dient
afgehaald of toegevoegd te worden. In dit geval kan de SDH ADM vervangen worden door
een golflengte die transparant door de OADM passeert. Door het SDH verkeer op de
SAMENVATTING
x
lii
geschikte ring te plaatsen, zodanig dat de hoeveelheid doorgaand verkeer in de knopen van de
SDH ringen gemaximaliseerd wordt, kunnen een aanzienlijk aantal SDH ADMs bespaard
worden. Dit planningsprobleem zal in paragraaf 5.6 in meer detail behandeld worden.
SDHSDH SDH
SDHSDH SDH
OADM
Figuur 14: SDH gebaseerde WDM ring
5.3 Optimale routering op een SPRing
Het probleem om verkeer te routeren op een SPRing, zodanig dat zo weinig mogelijk
ringcapaciteit nodig is, kan geformuleerd worden met behulp van een lineaire objectieffunctie
en een stel lineaire beperkingen. Voor elke vraag tussen 2 knopen, uitgedrukt als een
hoeveelheid golflengten, dient bepaald te worden hoeveel golflengten er in wijzerzin, en
hoeveel in tegenwijzerzin gerouteerd worden. Zo een lineair probleem kan dan opgelost
worden met behulp van een optimalisatiealgoritme voor lineaire programma's.
In sommige gevallen is het niet nodig om de optimale routering van elke vraag te
kennen, maar is een accurate schatting van de benodigde ringcapaciteit voldoende. Dit is
bijvoorbeeld het geval wanneer het routeringsprobleem een deelprobleem is, dat veelvuldig
dient opgelost te worden binnen een veel groter planningsprobleem (zoals het ontwerp van
een netwerk bestaande uit meerdere ringen). Ook als het verkeer niet met voldoende
zekerheid kan voorspeld worden is het weinig zinvol om de optimale routering te bepalen. In
dit geval is het veel nuttiger om een goede schatting van de benodigde ringcapaciteit te
bekomen, op basis van een aantal parameters die het verkeer ruwweg beschrijven. Teneinde
een accurate schatting te maken, werd verder gebouwd op een bovengrens voor de benodigde
ringcapaciteit die in de literatuur bekend is. Deze bovengrens, gecombineerd met een
eenvoudig af te leiden ondergrens geeft ons zeer nauwe grenzen voor de benodigde
ringcapaciteit, met een maximale fout van één golflengte. Deze grenzen kunnen dan eveneens
gebruikt worden om de formulering van het lineaire programma te versterken, wat toelaat om
sneller een optimaal resultaat te bekomen.
De boven- en ondergrens kunnen nu ook toegepast worden op een aantal
verkeerspatronen die vrij typisch zijn voor ringen, en kunnen gebruikt worden om het
capaciteitsgebruik van de SPRing te vergelijken met de DPRing. In een stervormig
verkeerspatroon vertrekt vanuit elke knoop juist één vraag, en dit steeds naar één en dezelfde
knoop (ook hub genoemd). Dit patroon is het minst geschikt voor een SPRing, omdat het
capaciteitsgebruik van de SPRing in dit geval identiek is aan dat van de DPRing. In een
verkeerspatroon met allemaal lange connecties, bestaat enkel een vraag tussen twee
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
liii
diametraal tegenovergestelde knopen. In dit geval biedt de SPRing enkel een (weliswaar
klein) voordeel tegenover de DPRing indien het aantal knopen op de ring oneven is. In een
uniform verkeerspatroon bestaat tussen elk knopenpaar een vraag. In dit geval kan een
SPRing tot de helft minder capaciteit gebruiken dan een DPRing. Tenslotte is het
verkeerspatroon waarbij enkel een vraag bestaat tussen aangrenzende knopen, het meest
geschikt voor de SPRing, omdat in dit geval de capaciteit optimaal hergebruikt wordt op de
ring.
Het capaciteitsgebruik van een SPRing en een DPRing kan ook vergeleken worden
op basis van meer algemene verkeerspatronen. Hiertoe kan gebruik gemaakt worden van de
lineaire mathematische formulering van het probleem, welke toelaat het probleem op te lossen
met gekende technieken voor lineaire programmeringsproblemen. Uitgemiddeld over een
groot aantal problemen, blijkt dat voor willekeurige verkeerspatronen de SPRing tussen 20%
en 40% minder capaciteit nodig heeft dan de DPRing. Hoe meer knopen de ring bevat, des te
meer uitgesproken deze capaciteitsbesparingen zijn (zie Figuur 15).
0%
20%
40%
60%
80%
100%
4 6 8 10 12 14 16
Aantal knopen op de ring
Relatieve ringcapaciteit SPRing/DPRing
Figuur 15: Vergelijking capaciteitsgebruik SPRing versus DPRing
5.4 Golflengte toekenning op een SPRing
Hier beschouwen we het probleem waarbij aan gerouteerd verkeer op de ring een
golflengte moet toegewezen worden, zodanig dat het aantal benodigde golflengten op de ring
minimaal is. Ook dit probleem kan geformuleerd worden met behulp van een lineaire
objectieffunctie en een stel lineaire beperkingen. Zo een lineair probleem kan dan opnieuw
opgelost worden met behulp van een gekend optimalisatiealgoritme voor lineaire
programma's. De lineaire formulering van het probleem kan bovendien efficiënter gemaakt
worden door enkele kunstgrepen uit te voeren. Zo kunnen connecties die gesplitst worden
over de ring (langs beide kanten van de ring gerouteerd), deels uit de formulering verwijderd
worden omdat de golflengte toekenning eenvoudig op voorhand kan gebeuren. Ook kunnen
SAMENVATTING
x
liv
aan een aantal connecties reeds op voorhand golflengten toegekend worden, om het aantal
mogelijke permutaties van oplossingen te beperken.
Door het golflengte toekenningsprobleem op deze manier op te lossen kan een
vergelijking gemaakt worden tussen het aantal benodigde golflengten in ringen met en zonder
golflengte convertoren. Inderdaad, door te starten van een gegeven routering op de ring,
weten we hoeveel golflengten er nodig zijn in het geval van een ring met golflengte
convertoren. Aangezien in dit geval geen golflengte conflicten kunnen optreden, is dit
eenvoudigweg 2 maal de totale vraag op de meest beladen link van de ring. Bij een ring
zonder golflengte convertoren daarentegen, kan na golflengte toekenning blijken dat een
aantal additionele golflengten nodig zijn. Om dit aantal golflengten te kwantificeren, werden
resultaten gegenereerd voor een groot aantal problemen. Deze problemen gingen uit van een
SPRing met optimale routering (zie paragraaf 5.3), waarvoor dan de golflengte toekenning
gebeurde met behulp van lineaire programmering. In bijna alle gevallen, bleek dat het aantal
benodigde golflengten voor ringen zonder golflengte convertoren identiek is aan het aantal
golflengten in ringen met golflengte convertoren. Met andere woorden, het aantal golflengten
resulterend uit de routering, kan in de meeste gevallen toegekend worden aan de connecties
op de ring, zonder dat er golflengte conflicten optreden. In een heel klein aantal gevallen
(grootte orde 1%), zijn 1 of 2 extra golflengten nodig om de golflengte conflicten op te lossen.
Dit toont aan dat het gebruik van golflengte convertoren weinig zinvol is op ringen met
statisch verkeer. Daarnaast bewijst dit eveneens dat beide problemen van routering en
golflengte toekenning op de ring, die in principe gekoppeld zijn, mogen ontkoppeld worden
zonder dat daarbij veel verlies aan optimaliteit geboekt wordt.
5.5 Optimalisatie van de hybride DPRing/SPRing architectuur
Wanneer meerdere identieke ringen tussen dezelfde knopen vereist zijn om de
capaciteit te verhogen, spreken we van een stapel van ringen. Wanneer verkeer dient
gerouteerd te worden over zo een stapel ringen, dient bepaald te worden van welk type elke
ring is (DPRing of SPRing) en welk verkeer over welke ring gerouteerd dient te worden. Het
uiteindelijke doel hierbij is om de totale kost van de ringstapel te minimaliseren. Dit probleem
is bijzonder interessant, wanneer een DPRing goedkoper is dan een SPRing met evenveel
golflengten (wat realistisch is in de praktijk), omdat in dit geval het betere capaciteitsgebruik
van de SPRing dient afgewogen te worden tegen de meerkost van de SPRing. Teneinde de
kost van deze hybride ring architectuur te optimaliseren, kan dit probleem opnieuw op lineaire
wijze geformuleerd worden en opgelost worden met gekende technieken.
De oplossing van de hybride DPRing/SPRing architectuur kan dan vergeleken
worden met een architectuur waarbij slechts één type ring in de stapel gebruikt wordt (ofwel
DPRing, ofwel SPRing). Deze vergelijking kan gemaakt worden voor verschillende
kostscenario's en verkeerspatronen.
Een eerste vergelijking beschouwt willekeurige verkeerspatronen op een 10-knopen
ring, en een kostscenario waarbij de kost van de DPRing varieert tussen 50% en 100% van de
kost van de SPRing (zie Figuur 16). Indien de kost van de DPRing gelijk is aan de kost van de
SPRing, bestaan alle ringen in de hybride stapel uit SPRingen. De kost van de oplossing met
enkel DPRingen is in dit geval gemiddeld meer dan 50% hoger. Zodra de kost van de DPRing
lichtjes daalt t.o.v. de SPRing (bijvoorbeeld 10%), bestaat de hybride oplossing reeds voor
meer dan de helft uit DPRingen, en is de kost van de hybride oplossing lager dan de oplossing
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
lv
met enkel SPRingen. Deze trend wordt verder gezet naarmate de relatieve kost van de
DPRing daalt t.o.v. de SPRing. Wanneer de kost van de DPRing 60 à 70% van de kost van de
SPRing bedraagt, wordt de oplossing met enkel DPRingen goedkoper dan de oplossing met
enkel SPRingen. De hybride oplossing is in dit geval nog meer dan 20% goedkoper.
0.0%
10.0%
20.0%
30.0%
40.0%
50.0%
60.0%
70.0%
100% 90% 80% 70% 60% 50%
Kost DPRing versus SPRing
Gemiddelde extra kost t.o.v. hybride architectuur
Enkel SPRingen
Enkel DPRingen
Figuur 16: Kost van de hybride architectuur t.o.v. SPRing en DPRing architecturen
Een tweede vergelijking tussen de hybride architectuur en architecturen met één
enkel ring type werd gemaakt voor speciale verkeerspatronen op een 12-knopen ring.
Opnieuw werd de kost van de DPRing geschaald tussen 50% en 100% van de kost van de
SPRing. Voor het stervormige verkeerspatroon en het verkeerspatroon met lange connecties is
de oplossing met DPRingen steeds te verkiezen. De hybride oplossing bestaat dan ook enkel
uit DPRingen. Voor het verkeerspatroon met enkel verkeer tussen aangrenzende knopen is de
oplossing met SPRingen steeds te verkiezen. De hybride oplossing bestaat dan ook enkel uit
SPRingen. Voor het uniforme verkeerspatroon daarentegen, is de hybride architectuur wel
beter dan beide architecturen met één enkel ringtype, indien de kost van de DPRing lager is
dan die van de equivalente SPRing. Het uniforme verkeerspatroon wordt op die manier
uiteengetrokken in 2 verkeerspatronen die beter geschikt zijn voor elk ringtype.
De hybride oplossing biedt ook een interessant kostvoordeel in het geval drop &
continue gebruikt wordt op de ringen, zelfs in het geval de DPRing en SPRing een identieke
kost hebben. In het bijzonder wanneer er ongeveer evenveel inter- als intra-ring verkeer is,
kan de hybride architectuur kostbesparingen tot 20% realiseren tegenover beide architecturen
met één enkel ringtype. De architectuur met enkel DPRingen is vooral geschikt voor ringen
met veel inter-ring verkeer, omdat drop & continue geen impact heeft op de
capaciteitsvereisten van de DPRing. De architectuur met enkel SPRingen is daarom vooral
geschikt voor ringen met veel intra-ring verkeer, omdat hierbij geen drop & continue nodig is.
SAMENVATTING
x
lvi
5.6 Optimalisatie van SDH gebaseerde WDM ringen
Wanneer een WDM ring gebruikt wordt om meerdere gestapelde SDH ringen te
ondersteunen, beschouwen we het optimalisatieprobleem waarbij het aantal SDH ADMs
geminimaliseerd dient te worden. Door het SDH verkeer zodanig te routeren dat doorgaand
verkeer in een bepaalde knoop zoveel mogelijk op dezelfde SDH ring geplaatst wordt, kan zo
een minimale bezetting van SDH ADMs bekomen worden. Hierbij gaan we uit van een
gegeven routering op een SPRing. Voor elke connectie is dus gegeven of ze in wijzerzin of
tegenwijzerzin gerouteerd wordt, en moet nog beslist worden op welke SDH ring deze
connectie dient gerouteerd te worden. Opnieuw kan dit probleem op lineaire wijze
geformuleerd worden en opgelost worden met een optimalisatiealgortime voor lineaire
programma's. We kunnen ook nog een aantal extra randvoorwaarden opleggen, die de
formulering strenger maken en sneller oplosbaar maken.
Er werden resultaten behaald op diverse willekeurige verkeerspatronen (zie Tabel 1).
Voor kleine ringen (5 à 6-tal knopen) kunnen een 25 à 30% van de SDH ADMs bespaard
worden vergeleken met de maximale bezetting waarbij in elke knoop van elke SDH ring een
ADM geplaatst wordt. Voor middelgrote ringen (7 tot 9 knopen) bedragen deze besparingen
40 tot zelfs 50%. Voor nog grotere ringen wordt het probleem snel zeer complex en konden
geen optimale resultaten bekomen worden binnen een aanvaardbare rekentijd. Toch wordt
verwacht dat de trend zich verder zet.
Knopen op
de ring
Gestapelde
SDH ringen
Minimaal/maximaal
benodigde ADMs
Besparingen
in ADMs
5 2 7/10 30.0 %
6 2 8/12 30.0 %
6 2 9/12 25.0 %
7 2 10/14 28.6 %
7 3 14/21 33.3 %
8 3 15/24 37.5 %
8 3 14/24 41.7 %
8 4 18/32 43.7 %
9 4 20/36 44.4 %
9 5 23/45 48.9 %
9 6 30/54 44.4 %
Tabel 1: Resultaten ADM eliminatie voor verschillende ringen
Wanneer we kijken naar meer speciale verkeerspatronen op de ringen, blijkt dat de
ADM minimalisatie bijzonder nuttig is voor het stervormige verkeerspatroon en het
verkeerspatroon met lange connecties. Voor het uniforme verkeerspatroon kan ook nog een
aanzienlijk deel van de ADMs bespaard worden, terwijl voor het verkeerspatroon met
aangrenzende connecties geen besparingen mogelijk zijn. Het is dus net voor deze
verkeerspatronen die minder geschikt zijn voor de SPRing dat de besparingen het grootst zijn,
zodat de SPRing hiervoor toch nog interessant blijkt omdat een aanzienlijk aantal ADMs kan
bespaard worden. Het belangrijkste nadeel van de hier besproken ADM minimalisatie is de
verminderde flexibiliteit van het netwerk, omdat de locaties van de ADMs specifiek zijn voor
het beschouwde verkeerspatroon. Bij snel veranderende verkeerspatronen kan dit leiden tot
blokkering, en daarom is ADM minimalisatie enkel geschikt voor statische verkeerspatronen.
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
lvii
6. Planning van geïnterconnecteerde WDM ringen
6.1 Inleiding
In hoofdstuk 6 beschouwen we de planning van een netwerk bestaande uit meerdere
geïnterconnecteerde WDM ringen. De belangrijke beslissingen die dienen genomen te worden
bij de planning van zo een netwerk, zijn de plaatsing en interconnectie van de verschillende
ringen, en de routering van verkeer over deze ringen. In hoofdstuk 6 wordt een architectuur
beschouwd waar WDM DPRingen vrij (niet gebonden aan een hiërarchie) met elkaar
verbonden worden. Ring interconnectie wordt gerealiseerd met behulp van rug-aan-rug
geplaatste OADMs, zonder gebruik te maken van OXCs. In de interconnectiepunten tussen de
verschillende ringen worden transponders voorzien om het optisch signaal te regenereren. In
de volgende paragrafen worden de verschillende planningsproblemen en invoergegevens kort
besproken. Daarna worden deze planningsproblemen in meer detail uitgewerkt en worden
verschillende oplossingstechnieken voorgesteld.
6.2 Beschouwde planningsproblemen voor geïnterconnecteerde WDM ringen
De planning van geïnterconnecteerde WDM ringen behelst 3 deelproblemen: ring
identificatie, ring dimensionering en ring routering.
Bij ring identificatie dienen de posities van de ringen bepaald te worden, en dit
zodanig dat het alle verkeer over deze ringen kan gerouteerd worden tegen een zo laag
mogelijke kost. Hierbij veronderstellen we dat de onderliggende kabeltopologie gegeven is,
maar dat nog geen ringen in het netwerk aanwezig zijn.
Bij ring dimensionering zijn de posities van de ringen gegeven, maar de ringen
dienen nog gedimensioneerd te worden, zodanig dat alle verkeer kan gerouteerd en
geprotecteerd worden. Ring dimensionering is eveneens een deeltaak van het ring
identificatieproces. Voor ring dimensionering beschouwen we DPRingen met een gegeven
aantal golflengten, wat kan verschillen van producent tot producent. Voor een gegeven type
DPRing met een vast aantal golflengten, houdt de dimensionering dus in dat op elke ring
positie dient bepaald te worden hoeveel van deze DPRingen er dienen geïnstalleerd te
worden.
Bij ring routering tenslotte, zijn zowel de posities als de dimensionering van de
ringen gegeven en vast. Binnen het gedimensioneerde ringnetwerk dient dan een zo groot
mogelijk deel van de verkeersmatrix gerouteerd te worden.
6.3 Invoergegevens
De verschillende invoergegevens voor het beschouwde planningsprobleem worden
hieronder kort besproken.
De netwerk topologie is gegeven onder de vorm van een stel knopen en
tussenliggende links. Voor ring dimensionering en routering is eveneens de ligging van de
verschillende ringen opgegeven, en in het laatste geval zelfs ook de dimensionering van de
ringen.
Het verkeer is uitgedrukt als een hoeveelheid golflengten (bidirectioneel) die tussen
elk knopenpaar gerouteerd en geprotecteerd dienen te worden.
SAMENVATTING
x
lviii
De ringen zijn van het DPRing type en worden gekarakteriseerd aan de hand van de
hoeveelheid golflengten op een ring. Alle nieuwe te plannen ringen binnen het netwerk zijn
verondersteld van dezelfde producent afkomstig te zijn en hebben zodoende dezelfde
karakteristieken. Ringen die eventueel reeds aanwezig zijn in het netwerk kunnen van een
andere producent zijn, of van dezelfde producent maar een ouder model, en kunnen dus
andere karakteristieken hebben.
De protectie gebeurt ring-per-ring. De verschillende segmenten van een connectie die
over verschillende ringen gerouteerd is, worden dus in elke ring afzonderlijk geprotecteerd.
De ringen zelf kunnen transparant of opaak zijn. In het eerste geval vindt geen regeneratie
plaats in de ring, terwijl in het laatste geval in elke OADM 3R regeneratie gebeurt. In het
geval van transparante ringen dienen ringen dus beperkt te worden in lengte en aantal knopen,
om aanvaardbare prestaties te garanderen. Deze restricties kunnen opgelegd worden tijdens
het ring identificatieproces. De interconnectie van verschillende ringen gebeurt steeds aan de
hand van transponders, zodat het signaal tussen 2 ringen kan geregenereerd worden. De
interconnectie tussen 2 ringen kan gebeuren via één knoop, of via 2 knopen met behulp van
drop & continue.
Wat de kostenstructuur betreft, wordt een generiek kostenmodel gehanteerd. Elke
ring heeft hierbij een vaste kost, onafhankelijk van de hoeveelheid verkeer die op de ring
gerouteerd wordt. Deze vaste kost omvat de vaste installatiekost van vezels, versterkers,
OADMs, enzovoort. Daarnaast is er ook een routeerkost voor elke golflengte die op de ring
gerouteerd wordt. Deze routeerkost behelst de modulaire componenten van de OADM en
lijnsystemen. Voor de interconnectie van een golflengte tussen 2 ringen wordt een
interconnectiekost aangerekend. Deze kost bevat bijvoorbeeld de kost van de transponder.
Naast kost, is ook de netwerkbeschikbaarheid een belangrijke parameter om een
netwerk te evalueren. Daarom worden aan de verschillende netwerkelementen
foutprobabiliteiten toegedicht die toelaten de netwerkbeschikbaarheid te berekenen.
6.4 Het principe van het equivalente netwerk
Het equivalente netwerk is een manier om de topologische relatie tussen
verschillende ringen op een compacte wijze voor te stellen. Gezien het equivalente netwerk
als hulpmiddel gebruikt wordt bij verschillende oplossingsmethodes voor de
planningsproblemen van geïnterconnecteerde ringen, wordt dit concept hier eerst
geïntroduceerd.
Het equivalente netwerk bevat een knoop voor elke knoop uit het originele netwerk
(zie Figuur 17). Daarenboven wordt ook elke positie van een ring voorgesteld als een knoop
in dit equivalente netwerk. De relaties tussen ringen en knopen, en ringen onderling, in het
originele netwerk worden nu voorgesteld als links in het equivalente netwerk. Indien een ring
een bepaalde knoop uit het originele netwerk bevat, wordt dit voorgesteld als een
toegangslink tussen de overeenkomstige knopen in het equivalente netwerk. Indien 2 ringen
met elkaar geïnterconnecteerd zijn, wordt dit voorgesteld als een interconnectielink tussen de
overeenkomstige knopen in het equivalente netwerk. Het equivalente netwerk is een nuttig
hulpmiddel bij het routeren van verkeer in het ringnetwerk. Verkeer tussen 2 knopen in het
originele netwerk kan gerouteerd worden door een pad te zoeken tussen de overeenkomstige
knopen in het equivalente netwerk. Dit pad dient te starten en eindigen met een toegangslink,
terwijl alle tussenliggende links interconnectielinks dienen te zijn. Het pad kan dan terug
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN
x
lix
vertaald worden in een sequentie van ringen. Teneinde gefundeerde routeringsbeslissingen te
nemen, dienen de kosten en capaciteiten van het originele netwerk ook vertaald te worden
naar het equivalente netwerk. Voor meer details over hoe de kost en capaciteit van een ring
kan toegewezen worden aan de aangrenzende links, verwijzen we naar de Engelse tekst.
ring 1n1
n4
n3
n2
n7
n6
n5
n8
ring 2
ring 3
ring 4
Interconnectie-
link
toegangs-
link
n1
n4
n3
n2
n7
n6
n5
n8
ring 1
ring 2
ring 3
ring 4
Ring netwerk Equivalente netwerk
Figuur 17: Ring netwerk en equivalente netwerk
6.5 Ring routering
Het maximaliseren van de hoeveelheid op te zetten verkeer, binnen een
gedimensioneerd ring netwerk, kan geformuleerd worden als een lineair programma. Voor elk
knopenpaar worden een aantal mogelijke paden beschouwd langs waar het verkeer kan
gerouteerd worden. Deze paden halen we uit het equivalente netwerk met behulp van een
algoritme om de K kortste paden te bepalen. Het lineaire programma beslist dan voor elk
knopenpaar uit de verkeersmatrix, hoeveel verkeer er langs elk pad loopt, teneinde de totale
hoeveelheid gerouteerd verkeer te maximaliseren. Hoe meer kortste paden er beschouwd
worden (hoe groter K), hoe optimaler de oplossing kan zijn, maar hoe meer rekentijd vereist is
om het lineaire programma op te lossen. Er werd echter aangetoond dat het in de meeste
gevallen mogelijk is om een zeer goede oplossing te bekomen met een beperkt aantal kortste
paden.
Voor grote netwerken, met een groot aantal ringen, kan het oplossen van het lineair
programma een langdradig proces worden. Daarom werd voor zulke netwerken een snelle
(maar minder optimale) heuristiek uitgewerkt. Deze heuristiek start met het routeren van alle
connecties die over één ring kunnen gerouteerd worden. Nadien worden alle connecties
gerouteerd die over twee ringen kunnen gerouteerd worden, enzovoort, tot alle ringen volledig
bezet zijn, of geen verkeer meer kan gerouteerd worden. Als dusdanig wordt eerst alle intra-
ring verkeer opgezet, omdat hiervoor enkel capaciteit verbruikt wordt op één ring. Nadien
wordt verkeer opgezet dat capaciteit verbruikt op meerdere ringen, zolang er nog capaciteit
vrij is.
Beide algoritmes werden toegepast op een aantal illustratieve netwerken. In Figuur
18 wordt het resultaat getoond voor een 16- en 32-knopen netwerk met tussen elk knopenpaar
een vraag van 0 of 1 golflengte. De resultaten tonen aan dat de heuristiek in de meeste
gevallen zeer goede resultaten boekt wanneer we vergelijken met de resultaten bekomen met
lineaire programmering. De resultaten van de heuristiek zijn meestal in lijn met de resultaten
van lineaire programmering met K=1. Bij K=1 wordt enkel het kortste pad gebruikt, maar
SAMENVATTING
l
dient nog steeds de beslissing genomen te worden, om al dan niet verkeer op te zetten
langsheen dit kortste pad. Voor grotere waarden van K, worden nog iets betere resultaten
geboekt met lineaire programmering, hoewel de verschillen niet zo groot zijn. Grotere
waarden dan K=2 blijken hierbij weinig zin te hebben. Dit toont aan dat bij ring routering,
inderdaad voorrang gegeven wordt aan korte paden, zodat de ringcapaciteit zo efficiënt
mogelijk gebruikt wordt, wat ook de werking van de heuristiek verantwoord.
0%
20%
40%
60%
80%
100%
16-knopen netwerk 32-knopen netwerk
Hoeveelheid opgezet verkeer
Heuristiek
ILP (K=1)
ILP (K=2)
ILP (K=3)
Figuur 18: Resultaten ring routering
6.6 Ring dimensionering
Indien niet alle verkeer kan opgezet worden met behulp van ring routering, is er een
gebrek aan capaciteit in het netwerk. Om dit op te lossen, dient de capaciteit op bepaalde ring
posities verhoogd te worden door een extra ring bovenop de bestaande ring te stapelen.
Teneinde de kost te minimaliseren, willen we het aantal nieuw te installeren ringen tot een
minimum beperken, terwijl toch alle verkeer dient gerouteerd te worden.
Ook hier is het mogelijk om dit ring dimensioneringsprobleem te formuleren als een
lineair programma. Het lineair programma kiest opnieuw voor elk knopenpaar uit de
verkeersmatrix, hoeveel verkeer langs elk pad gerouteerd wordt. Het aantal te beschouwen
paden per knopenpaar kan opnieuw vrij gekozen worden. Een klein aantal paden bevordert de
rekentijd, terwijl een groter aantal paden toelaat om een meer optimale oplossing te bekomen.
We gebruiken opnieuw de K kortste paden per kopenpaar in het equivalent netwerk. In dit
geval heeft K=1 weinig zin, omdat alle verkeer dient opgezet te worden en dit dus steeds
langs het kortste pad zou gebeuren (wat betekent dat er geen optimalisatie gebeurt). Te hoge
waarden voor K hebben ook weinig zin, omdat typische vrij korte paden verkozen worden
(om intra-ring verkeer te bevorderen). Daarom zullen we typisch K= 2 of 3 gebruiken.
Opnieuw heeft lineaire programmering zijn beperkingen wat betreft rekentijd en
grootte van de op te lossen problemen. Daarom is het ook aangewezen voor ring
dimensionering een heuristiek te gebruiken om voldoende snel resultaten te genereren of om
zeer grote problemen binnen een korte rekentijd op te lossen. We stellen een heuristiek voor
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN li
die begint met alle verkeer te routeren langs het kortste pad in het equivalent netwerk. Hieruit
kan bepaald worden hoeveel ringen er nodig zijn op elke ring positie. De ringen die hierbij het
minst efficiënt gebruikt worden, zullen we trachten te verwijderen, door een deel van het
verkeer op deze ring positie langs een alternatief pad te routeren. Als genoeg verkeer kan
geherrouteerd worden, kan een ring verwijderd worden op de beschouwde ring positie, indien
dit daadwerkelijk leidt tot een kostreductie.
We vergelijken de resultaten bekomen met de heuristiek en het lineaire programma
met resultaten die bekomen werden met behulp van eenvoudige kortste pad routering in het
equivalente netwerk (zie Figuur 19 voor het 32-knopen netwerk). Hieruit blijkt dat zowel de
heuristiek als het lineaire programma resultaten bekomen die beter zijn dan kortste pad
routering. De resultaten van de heuristiek zijn hierbij vergelijkbaar met de resultaten van het
lineaire programma. Het kostverschil tussen de dimensionering met en zonder drop &
continue is toe te schrijven aan de extra ring interconnectie kost en het feit dat in sommige
gevallen enkele connecties andere routes dienen te nemen langsheen de ringen, om drop &
continue mogelijk te maken.
100000
110000
120000
130000
140000
150000
160000
Kortste
pad
Heuristiek ILP (K=2) ILP (K=3) Kortste
pad
Heuristiek ILP (K=2) ILP (K=3)
Kost
Met drop & continue Zonder drop & continue
Figuur 19: Resultaten dimensionering voor 32-knopen netwerk
We beschouwen ook diverse ring combinaties in hetzelfde netwerk. Zo kunnen we de
kost van een netwerk met een groot aantal relatief kleine ringen vergelijken met de kost van
een netwerk met een klein aantal grote ringen. Hieruit blijkt dat kleine ringen toelaten een
veel goedkoper netwerk te bekomen. Voor een 16-knopen netwerk bijvoorbeeld, is een
netwerk met 6 kleine ringen 25% goedkoper dan hetzelfde netwerk met 3 grote ringen. In een
netwerk met veel grote ringen zijn de protectiepaden typisch vrij lang, en wordt de capaciteit
dus minder efficiënt aangewend.
Daarenboven is een netwerk met meerdere kleine ringen ook betrouwbaarder dan een
netwerk met een beperkt aantal grote ringen, omdat in het eerste geval meer meervoudige
fouten kunnen opgevangen worden. In een netwerk met meerdere kleine ringen is de kans
immers groter dat de verschillende fouten zich in verschillende ringen voordoen, zodat ze
onafhankelijk van elkaar kunnen opgevangen worden. Dit is vooral interessant indien
SAMENVATTINGlii
linkfouten de dominante fouten zijn. Indien ook knoopfouten veelvuldig voorvallen gaat een
groot gedeelte van het verkeer sowieso verloren. Bovendien bevat een netwerk met veel
kleine ringen veel ring interconnectiepunten, welke kwetsbare punten zijn indien geen drop &
continue gebruikt wordt. In het geval van knoopfouten blijkt dan ook dat het gebruik van drop
& continue een grote impact heeft op de netwerkbeschikbaarheid (zie Figuur 20).
0
10
20
30
40
50
60
70
80
90
100
12 ringen 12 ringen +
D&C
7 ringen 7 ringen + D&C
Jaarlijks te verwachten verlies (STM-16 uur/jaar)
voor vezel en link fouten
0
100
200
300
400
500
600
Jaarlijks te verwachten verlies (STM-16 uur/jaar)
voor alle (knoop + link) fouten
Enkel kabel
fouten
Volledige link
fouten
Alle (knoop +
link) fouten
Figuur 20: Netwerkbeschikbaarheid voor verschillende ringcombinaties in het 32-knopen netwerk
6.7 Ring identificatie
In de vorige paragraaf werd reeds het belang onderstreept van een goede ringkeuze.
In deze paragraaf stellen we oplossingsmethodes voor om in een netwerk waarin nog geen
ringen aanwezig zijn, de meest geschikte ring posities te bepalen. Wegens het grote aantal
mogelijke ringen, en nog meer mogelijke ring combinaties, is dit een zeer complex probleem.
Daarom opteren we voor een heuristische optimalisatiemethode voor dit probleem. Deze
heuristiek bestaat uit 3 fasen: ring generatie, ring selectie en ring optimalisatie. Elk van deze
fasen wordt hieronder in meer detail besproken.
6.7.1 Ring generatie
In deze fase worden alle mogelijke ring posities in de onderliggende netwerk
topologie bepaald. De ringen dienen hierbij aan bepaalde voorwaarden te voldoen. Een eerste
voorwaarde stelt dat de lengte van de ringomtrek beperkt is. Hiervoor zijn een aantal redenen:
het tegengaan van distortie van het optisch signaal (in transparante ringen), een beperking van
de tijd nodig om over te schakelen naar het protectie pad, en bieden van een hogere
netwerkbeschikbaarheid. Ook het aantal OADMs op de ring kan beperkt worden om dezelfde
reden.
Om binnen een bestaande netwerk topologie alle ring posities te bepalen, die voldoen
aan bovenvermelde voorwaarden, wordt gebruik gemaakt van een algoritme dat alle ringen
bepaald die een gegeven knoop bevatten. Alle ringen behorend tot een gegeven knoop kunnen
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN liii
bepaald worden door alle paden te zoeken die vertrekken vanuit deze knoop, en er ook
opnieuw aankomen, zonder dat dit pad een andere knoop uit het netwerk meer dan éénmaal
bevat. Deze paden kunnen bepaald worden door een zogenaamd diepte eerst zoekboom op te
bouwen vanuit deze knoop. Zo een diepte eerst zoekboom onderzoekt op recursieve wijze alle
knopen van de gegeven knoop tot de beginknoop opnieuw bereikt wordt. Een pad in deze
zoekboom kan vroegtijdig afgebroken worden indien het niet voldoet aan één van de
bovenvermelde voorwaarden.
Om tenslotte alle ring posities binnen een netwerk te bepalen, kan bovenstaand
algoritme (om alle ringen behorend tot een gegeven knoop te bepalen) toegepast worden op
elke knoop van het netwerk. Om het zoekproces te versnellen en te vermijden dat elke ring
positie evenveel keer gevonden wordt als hij knopen heeft, zal een knoop uit het netwerk
verwijderd worden na alle ringen behorend tot deze knoop bepaald zijn. Op deze wijze wordt
elke ring positie binnen het netwerk juist éénmaal gevonden.
6.7.2 Ring selectie
Tijdens de ring generatie fase worden ringen louter bepaald op basis van de
onderliggende netwerktopologie, zonder hierbij rekening te houden met de verkeersmatrix.
Indien de topologie zeer vermaasd is, zal een enorm groot aantal ringen gegenereerd worden.
Teneinde een handelbaar aantal ringen over te houden, zullen we deze enorme set van ringen
beperken aan de hand van de verkeersmatrix. We wensen enkel deze ringen te behouden die
goed geschikt zijn om de opgegeven verkeersmatrix op te vangen. De ringen die dus in staat
zijn om de grootste hoeveelheid intra-ring verkeer te transporteren (relatief ten opzichte van
het aantal knopen op de ring) krijgen hierbij de voorkeur. We wensen de hoeveelheid inter-
ring verkeer te beperken omdat dit verkeer capaciteit verbruikt op meerdere ringen, alsook
een extra interconnectiekost met zich meebrengt. Daarom verkiezen we per knoop de K
ringen met het meest intra-ring verkeer in relatie tot het aantal knopen op de ring. We
selecteren de ringen per knoop, en niet over het gehele netwerk, om er zeker van te zijn dat
elke knoop binnen het netwerk kan bereikt worden op één of meerdere manieren.
6.7.3 Ring optimalisatie
Nadat we het aantal mogelijke ringen beperkt hebben tot de beperkte set van meest
geschikte ringen, gaan we op zoek naar de beste combinatie van ringen, teneinde alle knopen
van het netwerk te bedekken, en alle verkeer op een zo kost-efficiënt mogelijke wijze te
transporteren. Het aantal mogelijke ring posities binnen zo een geïnterconnecteerd
ringnetwerk kunnen we beperken tot een minimum en maximum aantal, vanuit het oogpunt
van kost, betrouwbaarheid of netwerkbeheer. Teneinde een goede combinatie van ring
posities te vinden die hieraan voldoet, werden twee algoritmes uitgewerkt: een exhaustief
zoekalgoritme en een taboe zoekalgoritme. Beide algoritmes worden hieronder in meer detail
besproken.
Het exhaustief zoekalgoritme evalueert alle mogelijke ringcombinaties waarbij het
aantal ring posities zich bevindt tussen het opgegeven minimum en maximum. Aangezien dit
aantal enorm snel kan oplopen is dit algoritme enkel geschikt voor kleine netwerken met een
beperkt aantal mogelijke ring posities. De evaluatie van een ring combinatie gebeurt door het
verkeer te routeren en de kost van het gedimensioneerde ringnetwerk te berekenen. Voor deze
dimensionering gebruiken we de heuristiek die in paragraaf 6.6 beschreven wordt. Deze
SAMENVATTINGliv
heuristiek dient uiteraard enkel aangewend te worden voor ring combinaties die resulteren in
een geïnterconnecteerd netwerk dat alle knopen omvat.
Het taboe zoekalgoritme evalueert eveneens een aantal mogelijke ring combinaties,
maar niet alle mogelijkheden. Het zoekproces gebeurt op een meer intelligente wijze,
namelijk gebruik makende van het taboe zoeken (Engels: tabu search), een
oplossingsmethode die voor tal van optimalisatie problemen kan aangewend worden. Bij
taboe zoeken wordt vertrokken van een bestaande (niet optimale) oplossing van het probleem
en worden op iteratieve wijze de naburige oplossingen van de huidige oplossing geëvalueerd
door kleine veranderingen aan de huidige oplossing aan te brengen. De beste naburige
oplossing wordt dan behouden als nieuwe oplossing. Teneinde te vermijden dat het proces
halt houdt in een lokaal optimum, wordt een taboe lijst bijgehouden die alle recente
veranderingen opslaat die hebben geleid tot de huidige oplossing. Door te verbieden dat deze
veranderingen opnieuw gemaakt worden, kan ontsnapt worden uit lokaal optimale
oplossingen. Het zoekproces wordt gestopt nadat geen betere oplossing gevonden wordt na
een aantal iteraties.
Voor ring identificatie wordt de beginoplossing van het taboe zoeken bekomen door
ringen met een grote hoeveelheid intra-ring verkeer te combineren zodat ze het gehele
netwerk omspannen. De evaluatie van een oplossing gebeurt opnieuw door het
dimensioneringsalgoritme uit paragraaf 6.6 toe te passen. Een "verandering" tijdens het taboe
zoeken wordt gedefinieerd als het toevoegen en/of verwijderen van een ring positie aan de
huidige ring combinatie.
6.7.4 Resultaten
Het hierboven beschreven proces van ring identificatie bestaand uit ring generatie,
selectie en optimalisatie, kan nu toegepast worden op een aantal netwerken. Vooreerst wordt
een vergelijking gemaakt tussen de resultaten bekomen met exhaustief en taboe zoeken.
Algemeen blijkt dat tijdens een beperkt aantal iteraties van het taboe zoekalgoritme typisch
vrij goede resultaten bekomen worden. In sommige gevallen wordt de optimale oplossing
bekomen, maar wat belangrijker is, is dat in de meeste gevallen een zeer goede oplossing
bekomen wordt in de eerste iteraties. Hierdoor kan het taboe zoekalgoritme in korte tijd een
goede ring identificatie uitvoeren, terwijl het exhaustief zoekproces zelfs voor kleine
netwerken meerdere uren duurt.
Verder werd ook de impact van de belangrijkste parameters op de resultaten
bekeken. Gedurende de ring generatie fase kan de grootte van de te beschouwen ringen
gekozen worden. Uit simulaties voor ringen met verschillende groottes blijkt dat grote ringen
niet toelaten de netwerkkost te verlagen in vergelijking met kleinere ringen. Het aantal
beschouwde ringen dat kan gekozen worden tijdens de ring selectie fase werd eveneens
bestudeerd. Hoe meer ringen gekozen worden, hoe beter het resultaat. Toch zijn de
verschillen klein, omdat in het geval een klein aantal ringen gebruikt wordt in de ring selectie
fase, dit wel de best geschikte ringen zijn. Dit bewijst de geschiktheid van de gekozen ringen
tijdens de ring selectie fase. Ook het aantal ring posities dat toegelaten wordt in de ring
optimalisatie fase beïnvloedt het resultaat. Typisch worden betere resultaten bekomen
naarmate meer ringen toegelaten zijn, omdat meer verschillende ringen meer intra-ring
verkeer toelaten en de interconnectiekost drukken.
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN lv
6.8 Vergelijking van ring gebaseerde en vermaasde netwerk architecturen
Bovenstaande algoritmes kunnen nu gebruikt worden om een WDM netwerk
gebaseerd op OCh-DPRingen te vergelijken met een vermaasd netwerk gebruik makend van
optische cross-connects (OXC). De vermaasde netwerken worden ontworpen voor
padprotectie, padrestauratie en linkrestauratie. Voor de planning van deze netwerken wordt
gebruik gemaakt van bestaande algoritmes. De verschillende architecturen worden vergeleken
voor 2 netwerken: een netwerk met 16 knopen, en een netwerk met 32 knopen. We
vergelijken linkkost, knoopkost en netwerk beschikbaarheid. Voor de gebruikte gegevens
verwijzen we naar de Engelse tekst.
De architectuur die gebruik maakt van geïnterconnecteerde ringen heeft de hoogste
linkkost (zie Figuur 21 voor het 32-knopen netwerk). Deze kost is gemiddeld ongeveer 20-
25% hoger dan voor padprotectie. Alhoewel beide architecturen gebruik maken van protectie,
werken ringen meer lokaal, waardoor ze meer capaciteit nodig hebben, wat leidt tot een
hogere linkkost. Linkrestauratie laat toe om gemiddeld nog eens 25% te besparen ten opzichte
van padprotectie en padrestauratie zelfs 35%. Padrestauratie is dus duidelijk te verkiezen wat
de linkkost betreft.
De knoopkost is dan weer het laagst voor de geïnterconnecteerde ringen, wegens het
gebruik van goedkoop veronderstelde OADMs. OXCs vereisen een hogere interne
connectiviteit en zijn dus veel duurder. Het gebruik van drop & continue bij ringen leidt
gemiddeld tot een toename van ongeveer 10% in knoopkost, omdat verkeer dan in 2 knopen
uitgewisseld wordt tussen 2 ringen. De knoopkost is het hoogst voor padprotectie, waardoor
dit de duurste architectuur wordt wat betreft totale kost. Typisch is de total kost van
geïnterconnecteerde ringen 20% lager dan voor padprotectie, met de door ons gebruikte kost
gegevens. Wanneer linkrestauratie gebruikt wordt, kan de knoopkost met ongeveer de helft
gedrukt worden ten opzichte van padprotectie. Dit komt omdat restauratie geen vooraf
opgezette protectiepaden dient op te zetten en de knopen dus voornamelijk gedimensioneerd
zijn voor werkend verkeer. Padrestauratie behoeft op zijn beurt nog eens 15-30% minder
knoopkost, waardoor dit de goedkoopste architectuur wordt, gevolgd door linkrestauratie.
Geïnterconnecteerde ringen zijn iets duurder dan linkrestauratie, maar wel goedkoper dan
padprotectie, en zijn dus een goede tussenoplossing.
SAMENVATTINGlvi
0
50000
100000
150000
200000
250000
300000
350000
Pad protectie Link restauratie Pad restauratie OCh-DPRing OCh-DPRing
+D&C
Kost
Knoopkost
Linkkost
Figuur 21: Link- en knoopkost van de verschillende architecturen voor het 32-knopen netwerk
Naast de kostprijs is ook de netwerkbeschikbaarheid belangrijk. We vergelijken het
jaarlijks te verwachten verlies aan verkeer (door link- en knoopfouten) voor
geïnterconnecteerde OCh-DPRingen (met en zonder drop & continue) en voor padprotectie.
De restauratie architecturen werden niet beschouwd, omdat daar de beschikbaarheid te veel
afhangt van het gebruikte restauratiealgoritme en de hoeveelheid vrije capaciteit. Wanneer
enkel linkfouten beschouwd worden (zie Figuur 22 voor het 32-knopen netwerk), resulteert
padprotectie in de laagste netwerkbeschikbaarheid. Geïnterconnecteerde ringen kunnen beter
omgaan met dubbele linkfouten, omdat elke ring afzonderlijk van een linkfout in de ring kan
herstellen, terwijl bij padprotectie het herstel tussen de eindpunten gebeurt, gebruik makend
van 2 disjuncte paden. Wanneer de ringen ook nog eens verbonden worden met behulp van
drop & continue wordt zelfs een nog hogere netwerkbetrouwbaarheid bekomen. Indien naast
linkfouten eveneens knoopfouten beschouwd worden (zie Figuur 23 voor het 32-knopen
netwerk), resulteren geïnterconnecteerde ringen zonder drop & continue in de minst
betrouwbare netwerk architectuur. Aangezien de ringen slechts in één knoop met elkaar
verbonden zijn, tast zo een knoopfout alle verkeer tussen de ringen aan. Padprotectie gebruikt
2 knoopdisjuncte paden, waardoor het minder gevoelig is aan knoopfouten. Wanneer de
ringen verbonden worden met behulp van drop & continue, is de betrouwbaarheid veel hoger,
en zelfs nog iets beter dan voor padprotectie.
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN lvii
0
20
40
60
80
100
120
140
160
180
250000 275000 300000 325000 350000
Totale kost
Jaarlijks te verwachten verlies
(STM-16 uur/jaar)
OCh-DPRing + D&C
OCh-DPRing
Padprotectie
Figuur 22: Netwerkbeschikbaarheid versus kost voor linkfouten in het 32-knopen netwerk
0
200
400
600
800
1000
1200
1400
250000 275000 300000 325000 350000
Totale kost
Jaarlijks te verwachten verlies
(STM-16 uur/jaar)
OCh-DPRing + D&C
OCh-DPRing
Padprotectie
Figuur 23: Netwerkbeschikbaarheid versus kost voor link- en knoopfouten in het 32-knopen netwerk
Geïnterconnecteerde ringen, gebruik makend van drop & continue, lijken dus een
interessante netwerk architectuur, zowel vanuit het standpunt van kost als netwerk
beschikbaarheid.
7. Topologische planning van het toegangsnetwerk
7.1 Inleiding
De vorige hoofdstukken spitsten zich voornamelijk toe op planningsproblemen die
zich stellen in het transportnetwerk. In hoofdstuk 7 wordt in meer detail gekeken naar de
SAMENVATTINGlviii
planningsproblemen in het toegangsnetwerk. Het toegangsnetwerk is verantwoordelijk voor
de verbinding van de individuele gebruikers met het netwerk. Wegens het groot aantal
eindgebruikers, zijn de kosten die gepaard gaan met de uitbouw van het toegangsnetwerk dan
ook enorm hoog. De meeste kosten zijn hierbij vereist voor de installatie van transmissie
apparatuur. Bij de planning van toegangsnetwerken is één van de voornaamste
aandachtspunten dan ook de minimalisatie van de hoeveelheid te installeren transmissie
apparatuur.
In hoofdstuk 7 focusseren we op de planning van toekomstige toegangsnetwerken,
die toelaten om diensten met hogere bandbreedtes aan te bieden aan de eindgebruikers.
Hiertoe dient nieuwe apparatuur in het netwerk geïnstalleerd te worden, ter vervanging van de
bestaande koper of coax kabel infrastructuur die niet toelaat om grote bandbreedtes te
ondersteunen voor een groot aantal gebruikers of over grote afstanden. Uit economisch
oogpunt, zal in een eerste fase de bestaande infrastructuur enkel gedeeltelijk vervangen
worden. Voor de laatste kilometers naar de gebruikers toe (de zogenaamde last drop) kan de
bestaande apparatuur opgewaardeerd worden teneinde hogere bandbreedtes te vervoeren. Zo
kan me bijvoorbeeld met behulp van ADSL (Assymetric Digital Subscriber Line) enkele Mb/s
vervoeren over een klassieke koperen telefoondraad. Aangezien dit enkel mogelijk is over een
beperkte afstand, zal het toegangsnetwerk opgesplitst worden in eilandjes van beperkte
afstand waarbinnen de bestaande apparatuur kan hergebruikt worden. Zo een eiland omvat
typisch de gebruikers die met een lokale straatkast verbonden zijn. Om de verschillende
eilandjes te verbinden met de lokale centrale van het toegangsnetwerk dient nieuwe
infrastructuur geïnstalleerd te worden die toelaat om hoge bandbreedtes te transporteren over
grotere afstand. Voor dit gedeelte van het toegangsnetwerk (ook het feeder gedeelte genoemd)
is optische vezel uitstekend geschikt. We zullen in hoofdstuk 7 dan ook het scenario
veronderstellen waarbij optische vezel dient geïntroduceerd te worden tussen de lokale
centrale en de bestaande straatkasten in het toegangsnetwerk. Dit scenario wordt in het Engels
ook fiber-to-the-cabinet (FTTC) genoemd.
7.2 Toegangsnetwerk planning
Het planningsprobleem dat in hoofdstuk 7 behandeld wordt, beschouwt de installatie
van kabels met optische vezel, teneinde een aantal opgegeven straatkasten (Engels: street
cabinets of afgekort SC) te verbinden met de lokale centrale (Engels: local exchange of
afgekort LEX) op de meest kost-efficiënte manier, en dit binnen een gegeven stratenplan. De
kost optimalisatie streeft de minimalisatie van de installatiekost van optische vezel kabels na.
Deze installatiekost wordt in grote mate bepaald door de werken (zoals graven, boren, …) die
nodig zijn om de kabels te installeren. Het stratenplan is hierbij belangrijk, omdat kabels
enkel langs bestaande straten kunnen geïnstalleerd worden. Met elk straatsegment kan dan een
kost geassocieerd worden, die de installatiekost van de kabel in dit straatsegment weergeeft.
Deze kost kan verschillen van straat tot straat, en kan zeer hoog zijn in straten waar werken
moeilijk zijn (zoals drukke straten of in tunnels), terwijl deze kost ook laag kan zijn, indien
reeds wachtbuizen aanwezig zijn in een bepaald straatsegment. Teneinde dit soort
geografische informatie te modelleren wordt gebruik gemaakt van een geografisch informatie
systeem (GIS). Het GIS geeft de relevante geografische informatie door aan de
planningsalgoritmes en kan de door de algoritmes bekomen kabeltopologie dan opnieuw
weergeven.
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN lix
Wat de kabeltopologie voor het feeder gedeelte betreft, worden een aantal
mogelijkheden beschouwd (zie Figuur 24).
LEX
SC SC SC
a. Ster topologie
LEX
SC SC SC
b. Boom topologie
LEX
SC SC SC
c. Ring topologie
ON
LEX
SC SC SC
ONON
d. Hybride ring-boom topologie
Figuur 24: Topologieën voor het feeder gedeelte van het toegangsnetwerk
De meest eenvoudige kabeltopologie om alle SCs met de LEX te verbinden is een
ster topologie (zie Figuur 24a), waarbij er een individuele kabel loopt tussen de LEX en elke
SC. Deze topologie kan echter vrij duur uitvallen, en daarom kan de voorkeur gegeven
worden aan een boom topologie (zie Figuur 24b), die toelaat om de kabel te delen tussen
verschillende SCs. Het nadeel van een boomstructuur is echter de lage betrouwbaarheid,
omdat één kabelbeuk alle SCs afsnijdt die via deze kabel verbonden zijn met het netwerk.
Daarom kan geopteerd worden voor een ring topologie (zie Figuur 24c), die wel duurder is
maar die als voordeel heeft dat elke SC via beide kanten van de ring verbonden is met de
LEX, en dus bestand is tegen enkelvoudige kabelbreuken. Wanneer we de boom en ring
topologie met elkaar vergelijken komt het antagonistisch karakter van kost en
betrouwbaarheid duidelijk naar voor. Een tussenoplossing tussen de ring en boom topologie,
die een betere schaalbaarheid van kost en betrouwbaarheid toelaat, is de hybride ring-boom
topologie (zie Figuur 24d). Deze nieuwe topologie vereist de invoering van een nieuw type
netwerk element tussen de SC en de LEX, namelijk de optische knoop (Engels: optical node
of afgekort ON). Ringen worden in deze hybride topologie gebruikt om de ONs te verbinden
met de LEX. Vanuit elke ON wordt dan een boom uitgebouwd om enkele nabijgelegen SCs te
verbinden met deze ON. De grootste hoeveelheid verkeer is geconcentreerd in het gedeelte
dichtst bij de LEX, en daar is betrouwbaarheid dus kritisch. De ringen, die de LEX verbinden
met de ON bieden juist deze betrouwbaarheid. De netwerkgedeeltes verst van de LEX
concentreren veel minder verkeer (typisch verkeer naar één enkele SC), en daar kan dus
gekozen worden voor een iets mindere betrouwbaarheid. Daarom wordt gekozen voor een
boom topologie om SCs te verbinden met ONs.
SAMENVATTINGlx
Vanuit het oogpunt van netwerk planning, is de ster topologie het eenvoudigst te
plannen. Elke verbinding tussen SC en LEX kan immers afzonderlijk geoptimaliseerd worden
teneinde de kabel installatiekost te minimaliseren. De boom, ring en hybride structuur zijn
daarentegen veel moeilijker te plannen. In de volgende paragrafen worden een aantal
algoritmes voorgesteld die kunnen aangewend worden om deze topologieën te plannen, en
worden een aantal relevante resultaten samengevat.
7.3 Planning van de boom topologie
7.3.1 Probleemstelling
Bij de planning van de boom topologie zijn de locaties van de LEX en SC binnen het
stratenplan gekend, en dient de tussenliggende kabeltopologie bepaald te worden. Het
stratenplan kan hierbij voorgesteld worden als een ongerichte graaf waarbij elk hoekpunt op
het stratenplan een knoop van de graaf voorstelt, en elk straatsegment tussen 2 hoekpunten
een tak van de graaf voorstelt tussen de overeenkomstige knopen. De kost van zo een tak is
dan de installatiekost van kabel langsheen dit straatsegment. Het bepalen van de optimale
boom topologie voor de kabels komt dan neer op het bepalen van de Steiner tree die alle SCs
en LEX verbindt in de graaf die het stratenplan voorstelt. In de volgende paragraaf worden
drie technieken geschetst die toelaten dit probleem op te lossen. De technieken zijn gebaseerd
op heuristieken, omdat wegens de typisch grote omvang van de beschouwde problemen,
exacte technieken niet toelaten een oplossing te bekomen binnen een aanvaardbare rekentijd.
7.3.2 Oplossingsmethodes
Er werden 3 heuristieken uitgewerkt om de boom topologie zo optimaal mogelijk te
bepalen. Een eerste techniek, de zoom-in (ZI) techniek, houdt initieel enkel rekening met de
posities van de SCs en de LEX zonder het stratenplan in rekening te brengen. Binnen het
Euclidisch vlak wordt dan met behulp van een heuristiek een boomstructuur bepaald die de
SCs en LEX door middel van lijnstukken met elkaar verbindt. Deze lijnstukken worden dan in
een tweede fase vertaald in kabelroutes binnen het stratenplan. Nadat een route bepaald is
voor een bepaald lijnstuk, zal de kost van de gebruikte straatsegmenten tot nul gereduceerd
worden om het hergebruik van kabel aan te moedigen voor de andere lijnstukken die dienen
vertaald te worden in kabelroutes. Een laatste fase tenslotte optimaliseert bepaalde gedeeltes
van de bekomen boomstructuur binnen het stratenplan, door lokaal bepaalde gedeeltes van de
kabelroutes te herberekenen.
Een tweede techniek, de iteratieve padzoek (IPZ) techniek, tracht de langste
kabelroutes van de boom eerst te bepalen, zodat deze maximaal kunnen hergebruikt worden
voor de resterende kabelroutes. Dit gebeurt door eerst een kabelroute te bepalen die de twee
verst verwijderde SCs (of de LEX en een SC) met elkaar verbindt. Om te bevorderen dat deze
kabelroute ook andere SCs (of de LEX) passeert onderweg, wordt aan deze andere SCs (en de
LEX) een negatieve “bonus kost” gegeven, zodat zij aangemoedigd worden om in de
kabelroute opgenomen te worden die beide verst verwijderde SCs (of de LEX en een SC) met
elkaar verbindt. Nadat de kabelroute bepaald is, wordt de kost van de gebruikte
straatsegmenten tot nul gereduceerd om het hergebruik van kabel aan te moedigen voor de
volgende twee SCs (of de LEX en een SC) die verst van elkaar verwijderd zijn. Naarmate het
algoritme vordert wordt dus typisch minder en minder kabel geïnstalleerd. Nadat alle SCs en
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN lxi
LEX met elkaar verbonden zijn, wordt opnieuw een laatste lokale optimalisatie fase
doorgevoerd om bepaalde gedeeltes van de boom te trachten verbeteren.
Een derde techniek, de artificiële minimum spanning tree (AMST) techniek, neemt
het stratenplan van bij aanvang in rekening. Er wordt eerst een hulpgraaf opgebouwd die als
knopen alle SCs en LEX bevat. De hulpgraaf is volledig vermaasd, en de kost van de takken
tussen twee knopen is de kost van het kortste pad in de graaf overeenkomstig met het
stratenplan. Nadien wordt een minimum spanning tree berekend in deze hulpgraaf. De takken
van deze minimum spanning tree worden dan vertaald in kabelroutes binnen het stratenplan.
Nadat een route bepaald is voor een bepaalde tak uit de minimum spanning tree van de
hulpgraaf, zal de kost van de gebruikte straatsegmenten opnieuw tot nul gereduceerd worden
om het hergebruik van kabel aan te moedigen voor de andere takken die dienen vertaald te
worden in kabelroutes. Finaal wordt opnieuw een lokale optimalisatie fase doorgevoerd om
bepaalde gedeeltes van de boom trachten te verbeteren.
7.3.3 Vergelijking van de methodes
Teneinde de verschillende heuristieken met elkaar te kunnen vergelijken worden
deze toegepast op een aantal voorbeeld netwerken. In Figuur 25 wordt een typisch resultaat
getoond voor een boomtopologie.
Figuur 25: Voorbeeld van een boom topologie
De resultaten van de verschillende heuristieken worden vergeleken op basis van de
kost van de bekomen boomstrucuur en benodigde rekentijd. Een eerste algemene opmerking
is dat de bekomen resultaten met de 3 heuristieken voor alle beschouwde netwerken zeer dicht
bij elkaar liggen wat betreft kost (typische verschillen van 1%). Dit laat toe te besluiten dat
alle 3 heuristieken zeer goede resultaten bekomen. De ZI techniek geeft over het algemeen de
beste resultaten, maar heeft ook de langste rekentijd nodig. De IPZ en AMST techniek zijn
SAMENVATTINGlxii
zeer gelijkaardig wat betreft rekentijd. De AMST techniek is daarenboven beter bestand tegen
kostmodellen waarbij de kost van een straatsegment niet lineair is met de Euclidische afstand.
7.4 Planning van de ring topologie
7.4.1 Probleemstelling
Net als bij de planning van de boom topologie, zijn de posities van de SCs en de
LEX gegeven binnen het stratenplan. De vraag is nu om een zo economische mogelijke kabel
topologie te vinden die de SCs met de LEX verbindt met behulp van één of meerdere ringen.
Elke SC behoort tot juist één ring, en elke ring is verbonden met de LEX. Dit kan het
eenvoudigst gerealiseerd worden door één grote ring die alle SCs verbindt met de LEX. In de
meeste gevallen is dit echter niet toegelaten, omdat de ringen moeten voldoen aan bepaalde
voorwaarden wat betreft lengte en aantal knopen op de ring. Deze restricties geven de
technische limieten van de ring weer (zoals het beschikbare optische vermogen of restricties
inherent aan het protectiemechanisme), maar kunnen ook ingevoerd worden om de
betrouwbaarheid op te drijven (kortere ringen met een beperkt aantal knopen zijn
betrouwbaarder dan lange ringen met een groot aantal knopen).
7.4.2 Oplossingsmethode
We verkiezen opnieuw een heuristische optimalisatiemethode, teneinde grote
problemen relatief snel te kunnen oplossen. De heuristiek die we gebruiken is geïnspireerd op
oplossingsmethodes die met succes toegepast werden op het vehicle routing probleem (VRP).
Hierbij dienen de optimale routes bepaald te worden die een aantal vrachtwagens moeten
afleggen, vertrekkend en eindigend in één centraal depot, om een aantal klanten te bedienen.
De analogie is duidelijk: het depot komt overeen met de LEX, de klanten met de SCs, en de
routes van de vrachtwagens met de kabeltopologie van de ring.
We lossen het probleem op in een aantal stappen. In een eerste stap worden clusters
van SCs opgebouwd, waarbij alle SCs in een cluster tot dezelfde ring behoren. Initieel kiezen
we één SC per cluster. Hiertoe kiezen we een aantal SCs zodanig dat zij zo ver mogelijk van
elkaar verwijderd zijn binnen het stratenplan. Daarna worden aan deze clusters andere SCs
toegevoegd, en op die manier worden verschillende ringen uitgebouwd in verschillende
gebieden van het netwerk. Om te kiezen welke SC bij welke andere cluster van SCs wordt
gevoegd, wordt gebruikt gemaakt van een aangepast versie van de heuristiek van Clarke en
Wright. Voor we een SC toevoegen aan een cluster dient ook gecontroleerd te worden of de
ring binnen deze cluster nog wel aan de vooropgestelde restricties voldoet. Wanneer alle SCs
toegewezen zijn aan clusters, houden we een aantal ringen over. Elke ring wordt dan in een
laatste fase nog eens afzonderlijk geoptimaliseerd door andere volgordes van de SCs in de
ring uit te proberen.
7.4.3 Resultaten
Om de kwaliteit van de resultaten bekomen met de hierboven beschreven heuristiek
te onderzoeken, vergelijken we deze met optimale resultaten. Deze optimale resultaten
werden bekomen met behulp van lineaire programmering, op een aantal problemen van kleine
omvang. Te grote problemen kunnen immers niet opgelost worden met behulp van lineaire
programmering. Gemiddeld genomen bleken de resultaten bekomen met de heuristiek enkele
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN lxiii
procenten duurder te zijn dan de optimale oplossingen, maar de verschillen waren nooit zeer
groot. De rekentijd daarentegen verschilde wel dramatisch. Daar waar lineaire programmering
vele minuten vergt, geeft de heuristiek een oplossing binnen enkele seconden. In Figuur 26
wordt een typische voorbeeld gegeven van een ring topologie.
Figuur 26: Voorbeeld van een ring topologie
Gebruik makend van de algoritmes ontwikkeld voor de planning van boom
topologieën, kunnen we nu ook de kostprijs van een ring topologie vergelijken met deze van
een boom topologie. Hiertoe werd voor een aantal problemen zowel de boom als ring
topologie bepaald, en de procentuele meerkost van de ring topologie ten opzichte van de
boom topologie werd uitgemiddeld over een groot aantal problemen (zie Figuur 27). Hieruit
bleek dat de ring topologie minstens 25 tot 30% duurder is dan de boom topologie, wanneer
alle SCs op 1 ring kunnen geplaatst worden. Wanneer strengere voorwaarden opgelegd
worden voor het aantal SCs op de ring of de maximale lengte van de ring, stijgt deze meerkost
(en kan deze zelfs oplopen tot bijna 2 maal de kost van de boom topologie). Uit de
berekeningen voor verschillende opgelegde beperkingen, kan dan de meest geschikte ring
topologie bepaald worden, afhankelijk van de meerkost die de operator bereid is te betalen
voor de betrouwbaarheid geboden door de ring structuur.
SAMENVATTINGlxiv
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
50
50,000
20
50,000
10
50,000
50
20,000
20
20,000
10
20,000
50
10,000
20
10,000
10
10,000
50
5,000
20
5,000
10
5,000
Aantal toegelaten knopen op de ring
Maximaal toegelaten lengte van de ring
Gemiddelde extra kost van ring t.o.v. boom topologie
Figuur 27: Gemiddelde extra kost van een ring topologie versus boom topologie
7.5 Planning van de hybride ring-boom topologie
7.5.1 Probleemstelling
Bij de planning van de boom of ring topologie dient enkel de meest economische
kabeltopologie bepaald te worden voor een gegeven stel SCs en LEX. Voor de hybride ring-
boom topologie daarentegen, is het planningsprobleem veel complexer. Het probleem bestaat
hierbij uit de bepaling van het optimale aantal ONs en de optimale posities van deze ONs.
Daarnaast dient ook beslist te worden welke clusters van SCs met welke ON verbonden
worden. Tenslotte dient de optimale boom topologie berekend te worden die elke ON met de
hem toegemeten SCs verbindt, alsook de optimale ring topologie die de ONs met de LEX
verbindt.
Daarenboven is bij de hybride ring-boom topologie het minimaliseren van de
installatiekost op zich niet primordiaal, maar is het belangrijk om een goede afweging van
kost en betrouwbaarheid te realiseren. In de praktijk zullen we dit doen, door bepaalde
restricties op te leggen, die een bepaalde betrouwbaarheid garanderen, en rekening houdend
met deze voorwaarden de kost minimaliseren. Voorbeelden van beschouwde restricties zijn:
de maximale hoeveelheid verkeer binnen een cluster van SCs, het aantal SCs binnen zo een
cluster, of de maximale afstand tussen 2 SCs in dezelfde cluster. De betrouwbaarheid kan
gemeten worden aan de hand van de hoeveelheid verkeer die over niet-beschermde
kabelroutes loopt (dit zijn dus de boomstructuren). Dit betekent dat de probleemstelling in dit
geval ook een verkeersmatrix bevat, die weergeeft hoeveel verkeer (in Mb/s) elke SC naar de
LEX toevoert. Hoe de (on)betrouwbaarheid juist berekend wordt, staat in de Engelse tekst
beschreven. De kostprijs tenslotte, bevat niet enkel de kost van de te installeren kabelroutes,
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN lxv
maar ook de kost van de op te leveren optische knopen (ON). De kost van de ON bevat een
vaste kost, en een kost afhankelijk van het aantal interfaces nodig om SCs aan te sluiten. Voor
een meer gedetailleerde bespreking van de kostgegevens verwijzen we ook naar de Engelse
tekst.
7.5.2 Oplossingsmethode
Voor de planning van de hybride ring-boom topologie stellen we een algoritme voor
dat bestaat uit 4 opeenvolgende fasen die iteratief herhaald worden (zie Figuur 28).
Cluster
optimalisatie
Kost
aanpassing
Kost
aanpassing
Ring
optimalisatie
Figuur 28: Schematisch overzicht van het algoritme
In een eerste fase worden clusters van SCs bepaald die met eenzelfde ON verbonden
worden, alsook de boom topologie binnen elke cluster wordt bepaald. De bepaling van de
clusters gebeurt op basis van de geografische posities van de SCs en op basis van de
hoeveelheid verkeer die elke SC genereert. De SCs die dichtst bijeen gelegen zijn (gerekend
langs het kortste pad in het stratenplan) worden samengenomen in clusters voor zover hierbij
geen restricties overtreden worden. Voor de topologie van de boom binnen een cluster wordt
gebruik gemaakt van gelijkaardige algoritmes als ontwikkeld voor het plannen van een boom
topologie.
De tweede fase voert een kost aanpassing door van de straatsegmenten die gebruikt
worden door kabels van de boomstructuur. Dit laat toe dat de te bepalen ring bepaalde
kabelroutes hergebruikt die reeds voorzien werden voor de boomstructuur. De kost
aanpassing gebeurt door de kost van alle straatsegmenten die gebruikt worden door kabels van
de boomstructuur te vermenigvuldigen met een factor kleiner dan 1.
De derde fase bepaalt de optimale plaatsing van de ONs en van de ring die deze ONs
met de LEX verbindt, op basis van de bestaande clusters en de aangepaste kostenstructuur.
Voor de plaatsing van de ON binnen een cluster worden per cluster alle mogelijke
vertakkingspunten van de boom geëvalueerd als mogelijke positie van de ON. Voor
vastgelegde posities van de ONs kan de positie van de ring bepaald worden aan de hand van
het algoritme ontwikkeld voor planning van de ring topologie. Als alternatief kan ook een
ander algoritme gebruikt worden, dat de ONs verbindt in de volgorde die bepaald wordt door
hun hoek ten opzichte van elkaar en de LEX. Dit algoritme laat ook toe om bepaalde
gedeeltes kabel binnen de ring te hergebruiken (en dus een deel van de ring te reduceren tot
boomstructuur) om een lagere ringkost te bekomen.
Vervolgens wordt opnieuw een kost aanpassing doorgevoerd, dit keer op basis van
gebruikte kabels voor de ring. Deze kost aanpassing gebeurt door de kost van alle
straatsegmenten die gebruikt worden door kabels van de ringstructuur te vermenigvuldigen
SAMENVATTINGlxvi
met een factor kleiner dan 1. Daarna wordt het algoritme herhaald, waarbij nu de aangepaste
kostenstructuur uit de laatste fase van de vorige iteratie gebruikt wordt. Dit laat toe om tijdens
het bepalen van de clusters en de boomstructuur de toekomstige positie van de ring te
anticiperen en in rekening te brengen.
7.5.3 Resultaten
In de hieronder besproken resultaten illustreren we de afweging tussen kost en
betrouwbaarheid, die kan ingesteld worden door de parameters van het algoritme passend te
kiezen.
In de boomstructuren die de SCs met de ON verbinden kunnen we de
betrouwbaarheid regelen door bijvoorbeeld het maximaal aantal SCs per cluster te beperken,
of de maximale afstand tussen twee SCs in dezelfde cluster. Uit de resultaten blijkt inderdaad
dat topologieën waarbij een klein aantal SCs per cluster toegelaten worden en waarbij de
afstand tussen SCs beperkt is, resulteren in een hoge kost maar grote betrouwbaarheid. Deze
hoge betrouwbaarheid kan intuïtief verklaard worden omdat strenge restricties resulteren in
kleine clusters, waardoor een kabelbreuk typisch minder verkeersverlies met zich meebrengt.
In de ringstructuren kan kost ingeruild worden tegen betrouwbaarheid door
hergebruik van kabelroutes binnen dezelfde ring toe te laten om ONs te verbinden. Als
dusdanig ontstaan boomstructuren op de ring, waarbij we wel vereisen dat per boomstructuur
op de ring slechts één ON toegelaten is, zodat een enkele kabelbreuk op de ring maximaal één
cluster kan loskoppelen van het netwerk. Uit de resultaten blijkt inderdaad dat door hogere
gradaties van hergebruik toe te laten de kostprijs van de ring daalt, maar ook de
betrouwbaarheid (zie Figuur 29).
Figuur 29a: Ring met hoge kost en grote
betrouwbaarheid
Figuur 29b: Ring met lagere kost en een mindere
betrouwbaarheid
Finaal wordt ook de interactie tussen de boom en ring optimalisatie fase (en vice-
versa) onder de loep genomen. Deze interactie houdt een kostaanpassing in die toelaat beide
fasen beter op elkaar af te stemmen. Zonder deze kostaanpassing, blijkt dat de resultaten
beduidend minder goed zijn dan wanneer deze kostaanpassing wel doorgevoerd wordt. Er
blijkt eveneens dat de kostaanpassing die gebeurt bij overgang van de boom optimalisatie naar
de ring optimalisatie een grotere invloed heeft dan de kostaanpassing in de omgekeerde
richting.
PLANNING VAN RING GEBASEERDE TELECOMMUNICATIENETWERKEN lxvii
8. Finaal overzicht
Dit werk bundelt een aantal planningsproblemen voor telecommunicatienetwerken
gebaseerd op ringstructuren. Hierbij werd de toepassing van ringen in diverse gebieden van
het netwerk aangetoond.
Een eerste toepassing beschouwde een SDH gebaseerd netwerk, waar geïnter-
connecteerde MS-SPRingen vergeleken werden met SNCP padprotectie. Hierbij werd in het
bijzonder aandacht besteed aan een realistische modelering van verschillende mogelijke
scenario's om netwerkapparatuur te configureren in de knopen van het netwerk. De invloed
van de diverse scenario's op de planning werd hierbij bestudeerd, en de voor- en nadelen van
geïnterconnecteerde MS-SPRingen ten opzichte van SNCP padprotectie werden besproken
voor verschillende knooparchitecturen.
Een tweede bestudeerde toepassing van ringen spitste zich toe op een enkele WDM
ring. Zulke ringstructuren vind men bijvoorbeeld terug in het regionale netwerk. Diverse
plannings- en configuratieproblemen voor zo een ring werden geïdentificeerd, en gepaste
oplossingstechnieken werken voorgesteld die toelaten een ring op een optimale wijze uit te
bouwen.
Een derde toepassing beschouwde geïnterconnecteerde WDM ringen, die men
typisch terugvindt in het kernnetwerk. Er werden gepaste methodes uitgewerkt om verkeer
optimaal te routeren en om ringen op kost-efficiënte wijze te plannen. Bovendien werden
geïnterconnecteerde WDM ringen vergeleken met vermaasde WDM architecturen, zowel
vanuit het oogpunt van kost en betrouwbaarheid.
Ten laatste werd ook de toepassing van ringen in het toegangsnetwerk bestudeerd.
Aan de hand van gepaste planningstechnieken werden vergelijkingen gemaakt met boom
topologieën, en werd eveneens een hybride ring-boom topologie naar voor gebracht die de
voordelen van beide architecturen combineert.
SAMENVATTINGlxviii
CHAPTER 1
Introduction
1.1 Ring, ring
Just a minute, that's my GSM ringing. Hello! No, I'm just about to start writing my
Ph.D. Just send me an e-mail, ok!
GSM? E-mail? Services that were unknown ten years ago, but that now have
proliferated widely and became part of our everyday lives. Indeed, during the last decade
great shifts have been witnessed in the telecommunications arena, which provided us much
more powerful communications tools than ever before. Over the edges of time and place we
live, always connected, always on. To enable this, new technologies and services have been
created, and new networks have been built. Many operators are competing for a piece of the
pie, and want to see their networks optimized for cost, flexibility and reliability. This is where
the network planner comes in. He is the little grey mouse that no one sees, but performs
evaluations of different network technologies and architectures that are used to make million-
Euro investment decisions. In this thesis we present a number of planning cases that are of
interest to networks being deployed today. As the title of this paragraph already stated, the
leitmotif in this thesis is the use of rings in the different planning problems. Rings are a
simple, yet very efficient means to construct reliable networks. In this thesis we give an
overview of different planning problems involving rings, and we develop suitable planning
methods for these problems. This allows to evaluate the use of rings in the different parts of
the network.
1.2 Contents of this work
The remainder of this work is structured in different chapters, of which the content is
summarized below. Chapters 2 and 3 are introductory chapters, which lay the foundations on
which our research is based. Chapters 4, 5, 6 and 7 describe our own research, focussing on
different network planning problems and solution methods.
Chapter 2 serves as a general introduction for readers who are new to the
telecommunications world, and the problems of network planning. In this chapter, we
describe the current telecommunications landscape, and the nature of the planning problems
that come forward. The telecommunications environment is driven by technological
innovations, new services, and increasing competition. The different players in the
telecommunications market are all affected differently by this. While end-users have
benefited substantially from the increased competition, service providers and network
operators face increased network planning needs to increase their profit margins. Besides the
aspect of cost minimization, other facets of network planning are also highlighted. The
CHAPTER 12
different planning problems are further classified according to the considered time frame, the
geographical or functional area of the network and the requirements and objectives of the
network planner. Finally, Chapter 2 is concluded by giving a high level overview of different
planning techniques. On the one hand, some investment decision criteria are discussed to cope
with strategic planning problems. On the other hand, several optimization techniques
stemming from operations research are proposed in order to solve network dimensioning
problems.
Chapter 3 bundles the different aspects of transport network technologies and
architectures, which are of relevance for the subsequent chapters. After a general overview of
the structure of a transport network, the different functional requirements of such a network
are described. These requirements involve both methods to manipulate individual traffic
streams, as mechanisms to operate the entire network. Afterwards, the different transport
network technologies providing these functionalities are described. We hereby focus on the
SDH (Synchronous Digital Hierarchy) and WDM (Wavelength Division Multiplexing)
technologies, since these are the prominent technologies for the planning problems considered
in this thesis. For completeness, we also give an overview of packet switched technologies,
and the integration of packet switched networks in transport networks. The last part of this
chapter focuses on recovery techniques, which enable to reroute traffic in case of failures in
the network. The various recovery techniques that are used in SDH and WDM are
highlighted, including both ring-based architectures as protection and restoration techniques
in meshed networks. As already stated, in this thesis the ring-based mechanisms will prevail.
Chapter 4 deals with a first planning problem, considering a long distance network
based on SDH and WDM technology. A framework is developed to dimension and compare
different network architectures (based on interconnected rings and/or end-to-end path
protection), as well as different equipment configurations in the nodes of the network. The
main focus is on assessing the impact of the different equipment configurations on the
network design. Therefore the network design is split up in a routing phase and a
dimensioning phase. While the routing phase is kept relatively simple, for the dimensioning
phase detailed mathematical models are developed to take the characteristics and limitations
of equipment into account. Using this methodology, we then compare the different network
architectures and equipment configurations on a number of network topologies and traffic
matrices and present relevant conclusions.
In Chapter 5, we discuss the dimensioning and configuration problems that have to
be faced when deploying a WDM ring. Although the focus is on one single ring, many
challenging problems need to be solved, both in the long term as in the short term. For
optimally routing connections on a shared protection ring (i.e., the ring loading problem), we
present mathematical bounds for special demand patterns, as well as an exact optimization
technique for any demand pattern. Furthermore, a wavelength assignment algorithm for the
shared protection ring is developed, that allows to compare the performance of rings with and
without wavelength conversion. These algorithms also allow to compare the capacity
requirements of a shared protection ring with a dedicated protection ring. In addition, the
concept of a hybrid ring architecture is introduced, combining both ring types, and a
dimensioning algorithm for this hybrid architecture is developed. Finally, we also look at
rings in a multi-technology environment. More specifically, we target the design of SDH
rings, supported by a WDM ring. We discuss how to obtain cost reductions by eliminating
ADMs in the SDH rings, and present an algorithm to optimally design such networks.
INTRODUCTION 3
Whereas Chapter 5 deals with a single WDM ring, Chapter 6 goes one step further
by considering multiple interconnected rings in order to cover large networks. Design of such
a network involves determining the topology of the rings, and dimensioning of the rings by
routing traffic across the interconnected rings. We consider three different planning problems
that come forward when dealing with an interconnected ring network. For routing traffic in a
dimensioned ring network, we present an optimal algorithm and a heuristic that maximizes
the overall network utilization. For dimensioning a given ring topology with the aim of
minimizing the total installation cost, we also present an optimal algorithm and a heuristic.
Lastly, for determining the optimal ring topology, a three-phased heuristic is developed. The
performance of the developed algorithms is analyzed through results generated on different
sample networks. Finally, we also use these algorithms to compare network designs using
interconnected rings, with mesh-based designs.
Chapter 7 considers a planning problem specific to access networks. We consider a
migration scenario in which fiber is gradually introduced in the access network, reaching up
to the existing street cabinets. The planning problem then involves the minimization of the
fiber installation cost, taking into account geographical restrictions for the fiber topology
(which are imposed by the street map of a city). Several topologies are considered for this
fiber roll-out, such as star, tree and ring topologies. In addition a new hybrid ring-tree
topology is engineered, as an intermediate solution between the tree and ring topology,
combining the best of both worlds. For planning the different topologies, several heuristic
algorithms are developed that allow to cope with large-scale problems. Using these design
algorithms, the different network topologies can be compared with each other.
Finally, Chapter 8 summarizes the main results and conclusions obtained in this
thesis.
1.3 Published results
Most of the research work performed for this thesis, has also been published in
papers in international journals or proceedings of a large number of conferences. An overview
of these publications is given below.
Journal papers:
• P. Arijs, M. Gryseels, P. Demeester, "Planning of WDM ring networks", Photonic
Network Communications, Special issue on WDM transport networks: Key elements and
architectures, Vol. 2, No. 1, pp. 33-51, January-March 2000.
• K. Struyve, P. Arijs, N. Wauters, D. Colle, P. Demeester, P. Falcao, P. Lagasse,
"Application, design and evolution of WDM in GTS’s pan-European transport network",
IEEE Communications Magazine, Feature topic on WDM optical networks: A reality
check, Vol. 38, No. 3, pp. 114-121, March 2000.
• P. Arijs, W. Van Parys, B. Van Caenegem, P. Achten, P. Demeester, "Design of ring and
mesh based WDM transport networks", Optical Networks Magazine, Vol. 1, No. 3, July
2000.
• P. Arijs, R. Meersman, W. Van Parys, E. Iannone, A. Tanzi, M. Pierpaoli, F. Bentivoglio,
P. Demeester, "Architecture and design of optical channel protected ring networks", to
appear in Journal of Lightwave Technology, January 2001.
CHAPTER 14
Conference papers:
• P. Arijs, M. Claeys, P. Demeester, "The design of SDH ring networks using Tabu Search
and Simulated Annealing", Proceedings of the 5th International Conference on
Telecommunication Systems: Modeling and Analysis, Nashville (TN), March 1997.
• F. Poppe, M. Pickavet, P. Arijs, P. Demeester, "Design techniques for SDH mesh-
restorable networks", Proceedings of the 2nd European Conference on Networks and
Optical Communications (NOC'97), Antwerpen (Belgium), June 1997.
• F. Yin, M. Pickavet, P. Arijs, M. Gryseels, P. Demeester, "Three heuristic techniques for
topological access network design", Proceedings of the 2nd European Conference on
Networks and Optical Communications (NOC'97), Antwerpen (Belgium), June 1997.
• P. Arijs, P. Demeester, "Topological design of rings in the local access network",
Proceedings of the 6th International Conference on Telecommunication Systems:
Modeling and Analysis, Nashville (TN), March 1998.
• P. Arijs, M. Gryseels, M. Pickavet, F. Yin, P. Demeester, "Design of large-scale tree
networks: The case of urban access networks", Proceedings of the 4th Informs
Telecommunications Conference, Boca Raton (FLA), March 1998.
• P. Arijs, D. Colle, M. Gryseels, M. Pickavet, P. Demeester, N. Wauters, M. Groisman, K.
Struyve, P. Falcao, "SDH protection in long-distance networks: A practical case study",
Proceedings of the 1st International Workshop on the Design of Reliable Communication
Networks (DRCN'98), Brugge (Belgium), May 1998.
• P. Arijs, P. Demeester, "Efficient design of stacked SDH multiplex section shared
protection rings", Proceedings of the 3th European Conference on Networks and Optical
Communications (NOC'98), Manchester (UK), June 1998.
• P. Arijs, F. Yin, P. Demeester, "Geoplan: A tool for geographical planning of local access
networks", Proceedings of the 3th IEEE Symposium on the Planning and Design of
Broadband Networks, Mont-Tremblant (Canada), October 1998.
• B. Van Caenegem, P. Arijs, S. Declercq, P. Demeester, "Design of interconnected ring
WDM transport networks", Proceedings of the 3th IEEE Symposium on the Planning and
Design of Broadband Networks, Mont-Tremblant (Canada), October 1998.
• P. Arijs, F. Yin, P. Demeester, "A GIS-based planning environment for optical access
networks", Proceedings of the 3th Annual Symposium of the IEEE/LEOS Benelux
Chapter, Gent (Belgium), December 1998.
• P. Arijs, P. Demeester, "Design issues for WDM rings", Proceedings of the 6th
International Conference on Optical Communications and Networks, Paris (France),
January 1999.
• P. Arijs, M. Pickavet, N. Nelissen, S. Vanderdonckt, P. Demeester, "Design of a hybrid
ring-tree architecture for optical access networks", Proceedings of the 16th International
Teletraffic Congress (ITC16), Edinburgh (UK), June 1999.
• P. Arijs, P. Demeester, "The merit of shared and dedicated protection WDM rings in a
hybrid network design", Proceedings of the 25th Optical Fiber Communication conference
(OFC 2000), Baltimore (MD), March 2000.
• W. Van Parys, P. Arijs, P. Demeester, "Cost boundaries for economical OXCs",
Proceedings of the 25th Optical Fiber Communication conference (OFC 2000), Baltimore
(MD), March 2000.
INTRODUCTION 5
• P. Arijs, D. Colle, P. Demeester, "Optimization models for cost savings in stacked ring
network design", Proceedings of the 5th Informs Telecommunications Conference 2000,
Boca Raton (FLA), March 2000.
• P. Arijs, P. Demeester, P. Achten, W. Van Parys, "Dimensioning of non-hierarchical
interconnected WDM ring networks", Proceedings of the 4th conference on Optical
Networks Design and Modeling (ONDM 2000), Athens (Greece), February 2000.
• P. Arijs, W. Van Parys, P. Demeester, "Comparison of ring and mesh based WDM
networks", Proceedings of the 2nd conference on the Design of Reliable Communications
Networks (DRCN 2000), Munich (Germany), April 2000.
• W. Van Parys, P. Achten, P. Arijs, P. Demeester, E. Iannone, F. Bentivoglio, "Tools for
the design of meshed WDM networks using optical cross-connects", Proceedings of the
2nd conference on the Design of Reliable Communications Networks (DRCN 2000),
Munich (Germany), April 2000.
• D. Colle, P. Arijs, P. Demeester, K. Struyve, "Comparison of architectures for stacked
ring networks featuring compact add/drop multiplexers", Proceedings of the 2nd
conference on the Design of Reliable Communications Networks (DRCN 2000), Munich
(Germany), April 2000.
• P. Arijs, W. Van Parys, P. Demeester, "Ring and mesh architectures for reliable WDM
transport networks", Proceedings of the LEOS Benelux Workshop on Advanced Optical
Networks, Antwerp (Belgium), May 2000.
• P. Arijs, P. Demeester, W. Van Parys, R. Meersman, P. Achten, "Planning of survivable
WDM transport networks based on interconnected optical channel protected rings",
Proceedings of the 5th European Conference on Networks and Optical Communications
(NOC 2000), Stuttgart (Germany), June 2000.
• W. Van Parys, R. Meersman, P. Arijs, P. Demeester, E. Iannone, F. Bentivoglio,
"Network design tools for the optical layer and their value for operators and suppliers",
Proceedings of Networks 2000, Toronto (Canada), September 2000.
CHAPTER 1
6
CHAPTER 2
Network planning for the new millennium
2.1 Introduction
In this chapter we present some perspectives on the evolution of the
telecommunications market, that have intensified the need for network planning. Within this
context, an overview is given of the goals and application areas of network planning. Finally
some planning techniques, both from an economical and mathematical point of view, are
highlighted.
2.2 Evolution of the telecommunications sector
As we step into the third millennium, telecommunications and information
technology have converged in one of the largest and fastest growing industries around the
globe. This stunning growth has been fueled by an increased exchange of information on a
much more global scale as ever before. The main catalysts behind this evolution are the
successful introduction and adoption of new services, the leapfrogging advances on the
technological level and the liberalization and globalization of the telecommunications market.
2.2.1 Services
While traditional telecommunications services, such as fixed telephony, are still
growing approximately 10% a year in volume, new services are being introduced with much
more explosive growth figures. Data traffic, both on the public internet, as on the data-
optimized backbone networks of operators, is growing at triple-digit percentages each year [1]
(see also Figure 1). Another example is the subscriber growth for mobile telephony, which
has skyrocketed over the past few years (Figure 2). Other emerging services, such as virtual
private networks (VPNs) and unified messaging, are expected to gain popularity in the
forthcoming years [2], thereby sustaining the future growth. In order to migrate existing
voice-optimized networks towards a next generation network, capable of managing,
provisioning, transporting the plethora of new data-centered services, adoption and integration
of new technologies is of crucial importance for a competitive network operator.
CHAPTER 28
0
20,000
40,000
60,000
80,000
100,000
120,000
1999 2000 2001 2002 2003
Gbps
Voice Multiservice data Internet
Source: RHK, Inc. 1999
Figure 1: Traffic growth 1999-2003
Figure 2: Mobile subscriber growth in Europe
2.2.2 Technology
Over the past years, the speed of technological innovation has continued to increase
at a breathtaking pace. Governments (such as the European Commission) have provided
strong incentives to support research and development in the telecommunications arena.
Operators are eager to deploy the new technologies to expand their service portfolio, shorten
service provisioning cycles and reduce investment and operational costs. One example of such
technological advances is the increase in transmission capacity obtained on a single optical
fiber through the use of Wavelength Division Multiplexing (see Chapter 3). As can be seen in
Figure 3, the throughput on a fiber is ever increasing, while the transmission cost per Gb/s is
NETWORK PLANNING FOR THE NEW MILLENNIUM 9
continuously dropping [3]. Other examples of technological innovations can be mentioned as
well. Digital subscriber line (xDSL) technologies enable to reuse the existing copper
infrastructure for broadband services [4]. High-end switch-routers can forward billions of bits
per second through the internet [5]. Quality of service enhancements, complementing best
effort data transport, will soon enable to transport voice calls over this internet [6].
0
50
100
150
200
250
300
350
400
450
1994 1995 1996 1997 1998
Throughput (gbps)
$0
$50
$100
$150
$200
$250
$300
$350
$400
$450
Bandwidth Cost
$(000) per Gbps
Throughput Cost/Gbps Source: RHK, Inc. 1999
Figure 3: Advances in throughput and cost per Gb/s on a single optical fiber
2.2.3 Liberalization
After the US and the UK set the example more than a decade earlier, the European
telecommunications market was liberalized in 1998 [7], making it possible for new entrants to
compete with the incumbent operators. This new marketplace attracted a lot of worldwide
capital and soon numerous new operators saw the light of day, attacking the incumbents often
on the most profitable and promising market segments. This competition has lead to an
increased choice of services for the customer at much lower tariffs than ever before. While
end-users have benefited substantially from the continuous deflation of tariffs, it has had a
negative impact on the profit margins of operators. This decrease in margins is however
largely compensated by the increase in traffic volume (and thus revenue). Nevertheless,
operators are seeking new ways to further improve their profit margins. Cost reductions can
be obtained by increasing the internal efficiency (e.g. reducing overhead and staff) and
adopting new technologies. On the other hand, numerous mergers and acquisitions between
operators have been witnessed, proving the importance of taking full advantage of economies
of scale to reduce overall costs. Finally, an efficient network planning can play an important
role in an environment with high investment costs and decreasing margins, as will be shown
further in this chapter.
CHAPTER 210
2.3 Overview of the telecommunications market
The impact of the above described paradigm shift in the communications industry on
the global economy can not be underestimated. The information and telecommunications
market has radically changed the way of doing business [8] and thereby created a 'new
economy', which produces mainly immaterial goods and is characterized by a continuous high
growth and very low inflation [9]. Due to this 'goldilocks scenario', the new economy has
started dominating the more traditional economy. In April 2000, the list of 50 largest
companies in the world (based on market capitalization) contained 12 network operators and
service providers (top 3: NTT, Deutsche Telecom and AT&T), 8 telecommunications
equipment vendors (top 3: Cisco, Nokia and Lucent) and 10 companies from the computer
and software industry (top 3: Microsoft, Intel and Oracle) [10].
In the telecommunications market, 6 major players can be identified: the end-users,
the service providers, the network operators, the equipment vendors, the standardization
bodies and the regulating authorities.
The end-users or customers are unmistakably the most important players. These are
the consumers of the immaterial goods that the telecommunications market produces. Without
end-users there is simply no market: no services can be sold, no networks are needed and no
equipment is vended. Therefore it is important for service providers and operators to fully
understand the customers' needs and offer the right services to the right customers. Service
providers often identify three customer segments, requiring different services: residential
users, small and medium enterprises (SMEs) and large corporate customers.
The service providers are developing, marketing and selling the services for the
customers. Residential users typically require a single fixed and/or mobile telephone
connection, and dial-up internet access. SMEs on the other hand, require multiple fixed
telephones connected to a private automatic branch exchange (PABX), a fax service, a local
area network (LAN) with a permanent connection to the internet, web-hosting, unified
messaging, etc. Large corporate customers have the same needs as SMEs, but on a much
larger scale. Typically large corporations have multiple branches in different countries,
requiring virtual private networks interconnecting the networks of the individual branches. In
addition, large corporations often require call-centers, to offer support to their customers. As
communications is an essential part of any company's business, SMEs and large corporate
customers typically demand stringent service level agreements (SLAs) to ensure the quality of
the communications service. Thus, besides a differentiated service portfolio and attractive
pricing, service providers can also use quality of service (through SLAs) as a competitive
edge.
While some service providers offer services with a minimal amount of owned
infrastructure and lease network connectivity from operators (e.g. voice resellers, internet
service providers), other service providers are also network operators themselves.
The network operators or carriers own the network and sell network connectivity to
their customers. If the network operator is also a service provider (e.g. incumbent operators)
the customers of the network operator are the same end-users as mentioned above.
Alternatively, the operator can also sell capacity or infrastructure (e.g. so-called 'dark' fiber) to
service providers or even to other operators. Depending on the geographical area and
historical background of the operator, we can distinguish different kinds of operators. In
Europe the incumbent operators still have a strong foothold in their country of origin, because
NETWORK PLANNING FOR THE NEW MILLENNIUM 11
they own the infrastructure that connects the individual subscribers (i.e., the access network).
Therefore, the new operators are not mainly targeting the residential market, but other highly
profitable market areas with low up-front investment costs and high potential revenues.
Wholesale carriers are emerging on the pan-European market, offering international capacity
to service providers and international corporations. Also competitive local exchange carriers
(CLECs) are entering the market by installing access infrastructure (e.g. optical fiber, wireless
solutions) in large business cities. In the US, the monopoly of the incumbent operator was
brought to and end in 1984, when 'ma-Bell' AT&T was forced to divest all other holdings,
except for its long-distance network. At that point the market was also opened for other so-
called interexchange carriers (IXC) offering long-distance telephony, such as Sprint and MCI.
Over the past few years, these IXCs have been optimizing their network for data traffic to tap
into new revenue streams. At that point the IXCs also face fierce competition from the new
long-distance players, such as Qwest and Level 3, building a purely data optimized backbone
network. The former regional and access network of AT&T was divided amongst several
regional bell operating companies (RBOCs), the so-called 'baby-Bells' (e.g. Bell-South, Bell-
Atlantic). The RBOCs are in turn constituted of several incumbent local exchange carriers
(ILEC). This local exchange market was also opened to competitive local exchange carriers
(CLECs) and cable operators.
The equipment vendors are delivering the building blocks of the network to the
operators. Due to the large diversity of networking technologies, the vendor community
consists of a large amount of players. The best known vendors (e.g. Lucent, Alcatel, Nortel)
are those that develop products for almost all technologies, which allows them to offer
turnkey solutions to the operators. Besides these large players a lot of vendors focus on only
one technology or even one niche product (such as the start-up companies). The large
selection in technology and vendors widens the possibilities for the network operator, but on
the other hand it complicates the network planning process. It is important for the equipment
vendors to thoroughly understand the needs of the operators and to have a strong research and
development team (or good acquisition strategy for that matter) to turn innovative ideas into
real products. Besides, an attractive pricing of the products and a well trained post-sales
support team, is equally important.
Because of the myriad technologies and large amount of equipment vendors,
standardization bodies have an important role to play. Products must function and be
accepted in differing cultures, value systems and environments. By harmonizing
developments and reaching common agreements, the quality and interoperability of products
is ensured. Standards are of mutual benefit to vendors and operators: if adopted throughout
the world, standards create a large market instead of many fragmented markets.
It is a common misunderstanding that liberalization and deregulation go hand in
hand. Certainly in the initial phase of a liberated telecommunications market, rules set out by
regulatory bodies should smoothen the transition in order to protect the customer, new players
and incumbent operators while creating a competitive marketplace. In later stages, the market
can operate more freely, although the regulators should stay vigilant in ensuring that no
players with significant market power dominate the market.
CHAPTER 212
2.4 Importance of network planning
Network planning can be seen as the decision process involved with timely,
efficiently and cost-effectively building out a telecommunications network that meets the
business plan of the network operator. Within this definition, three major objectives can be
observed:
• The timely delivery of services, procurement and installation of equipment, and other -
also more strategic - decisions is of paramount importance to any network operator. By
closely supervising the network evolution, predicting future requirements, and using
accurate planning methods, the operator tries to take the right decision at the right time.
Decisions taken at a too early stage may result in stranded assets, while decisions taken
too late may result in customer dissatisfaction and loss of revenue.
• The efficient build out of the network involves choosing the technology and architecture
capable of offering the service-mix the operator wants to offer, while meeting quality of
service guarantees (such as reliability and delay). At the same time the operator wants to
minimize floor-space and operational overhead by choosing for proven equipment,
which is simple to operate, easy to manage, and accommodates future growth of traffic
and service types.
• A cost-effective network construction can have a serious impact on the business plan of
the operator. Even if the relative cost savings obtained through a cost-effective network
planning might be limited, the absolute cost savings might be substantial because of the
high investments involved in the network roll out. Furthermore, a reduction in
investment costs often also results in operational cost savings. Typically, the process of
network planning and resource and service provisioning, represents more than 50% of
the network operating expenses [11]. But perhaps the most important benefit of network
planning from a cost viewpoint, is the so-called 'gearing effect' it can have on the profit
of an operator. This can be illustrated by the following example. Consider an operator
with a profit margin of 10%. In essence, this means that a service which costs the
operator 100 units, can be sold at 110 units. Now, assume that a more efficient network
planning can save an operator just 1% on its total expenditures (which also includes
non-directly network related costs such as salaries, marketing costs, etc.). This means
that the cost of the service drops to 99 units and the profit margin increases to 11% (thus
a 10% relative increase). As such, savings of a mere 1% in expenditures through a more
efficient network planning, can lead to a 10% profit increase! If we consider lower
profit margins and higher savings the increase can be much higher. This clearly
illustrates the benefits of network planning in an environment with increasing
investments and decreasing profit margins.
The two major applications of network planning - certainly within this thesis - relate
to the optimization of a single network architecture and the comparison of different network
architecture alternatives.
Cost minimization is often the determinative objective when the planning problem
consists of dimensioning a single, well-defined network architecture and all strategic and
technological decisions have already been taken. At this point the planning problem can be
specified in full detail, and an optimization method tailored for the specific problem can be
worked out. Other than minimizing the network cost, the network planning also assists in the
provisioning process by giving input regarding the configuring of network equipment and
NETWORK PLANNING FOR THE NEW MILLENNIUM 13
routing of connections in the network. The network configuration process can be further
automated and optimized by integrating the network planning functionalities in the
management system of the network, to speed up service deployment cycles and reduce
operational costs.
Besides the optimization of a single network architecture, network planning also
plays a critical role in evaluating alternative architectures and making technology
development and deployment decisions [12]. These roles become particularly important in an
environment with a wide variety of fast developing technological alternatives and a large
amount of telecommunications equipment suppliers. Combined with the proliferating number
of services and uncertain traffic growth, it becomes much harder to decide how the network
should evolve over time. The use of network planning tools, which allow to evaluate different
network architecture and technology deployment scenarios, helps to make early selections
amongst alternatives and create a robust and future proof network.
2.5 Different planning problems
The nature of the planning problem can differ depending on the considered time
frame, the geographical or functional area of the network and the requirements and objectives
of the network planner. These different planning domains are depicted in Figure 4.
time
network
planner
area
operational tactical strategic
access
switching/routing
transport
incumbent operator
new operator
equipment vendor
Figure 4: Network planning domains
2.5.1 Planning time frame
As mentioned in paragraph 2.4, network planning is related to the build out of the
network over time. Depending on the considered time frame, different network planning
problems, with distinct goals and requirements, can be identified:
• Long term 'strategic' planning deals with the planning for the next 5 to 10 years and is
not quantitatively precise because the planning requirements and objectives are still
vague at this point and only few details of the planning problem are available. Long
CHAPTER 214
term planning concentrates on high level issues such as definition of business
objectives, corporate strategy, service development, demand forecasting, technology
choice and network architecture.
• Medium term 'tactical' planning is concerned with a planning horizon of the next few
years, for which relatively accurate forecasts are available regarding traffic evolution,
equipment prices and technological advances. Also planning requirements and
objectives can be specified in more detail, which are related to dimensioning,
procurement and installation of equipment for building and extension of the network.
• Short term 'operational' planning is closely coupled with day-to-day operations,
administration and management tasks. At this point the network architecture is well-
defined and detailed requirements and objectives can be specified. Short term planning
tasks typically relate to setting up network connections, modular equipment upgrades
and rearrangement of facilities to cope with unexpected failures or demand fluctuations.
These different planning tasks are not one time events, but should be performed
periodically, in an iterative process that adapts to the changing environment. As such, changes
in network architecture, technology advances, customer requirements, traffic growth, service
evolution and general business directions can be taken into account [13].
The planning problems considered in this thesis are mainly related to tactical and
strategic planning.
2.5.2 Network planning area
Operator's networks are typically divided in different geographical and functional
areas. The three most commonly considered areas of the network are the access network (also
called local loop), the switching network and the transport network (also called backbone
network):
• The access network connects the individual subscribers to the switch located in that
access area. Because of the large amount of equipment (mainly cabling) required in the
access network and the high labor cost associated with it, the access network represents
around 70% of the existing and new network investment. This implies that an efficient
planning of the access network can lead to serious overall cost reductions. Some of the
main planning activities in access networks relate to the choice of the access technology
and architecture (i.e. strategic planning), topological optimization of the cable roll-out,
and optimization of the switch locations and configurations (i.e. tactical planning).
• The switching network is involved with establishing dynamically switched connections
between different users of the network. Two main types of switched networks can be
distinguished [14]. The voice switched network deals with setting up, maintaining and
tearing down voice circuits. Because of this explicit connection establishment (through
signalling), such networks are referred to as connection-oriented networks. These
networks have the advantage that once a connection is set-up, reliable and low-latency
communications is guaranteed for the reminder of the connection. In the packet
switched network, data is transported in the form of packets and each packet is stored at
each intermediate packet switch (also called router) and forwarded to a neighboring
router until it reaches its final destination. The forwarding is based on the destination
address of the packet, which is contained in the header of the packet. As such, each
packet can find its own way through the network, and no explicit connection
NETWORK PLANNING FOR THE NEW MILLENNIUM 15
establishment is required. Such networks are referred to as connection-less networks. In
contrast to voice switched networks, current packet switched networks do not have the
same level of quality of service. Planning of the (voice or packet) switching network
involves dimensioning of the switches/routers and provisioning enough capacity on the
interconnection links between the switches/routers, such that all traffic can be
transported at each hour of the day (i.e. tactical planning). It might also involve
developing routing algorithms or signaling protocols to be implemented in the
switches/routers to dynamically set up the connections (i.e. operational planning).
Although there are some major differences between the packet and voice switched
network, which also impact the planning (e.g. related to quality of service requirements
of the different services), these are not within the scope of this initial overview and will
be discussed at a later stage.
• The transport network is responsible for establishing semi-permanent high capacity
connections between points of presence in the network. The transport network acts as a
server layer to client layers such as the voice and packet switched network for which it
provides interconnection links between the switches or routers. Due to the high
bandwidth requirements, optical fiber is the preferred transmission medium for the
transport network and reliability of the network is of paramount importance. Planning of
the transport network involves the choice of transmission technology, network topology
and architecture, recovery strategy (i.e. strategic planning) and capacity optimization on
the links, equipment dimensioning and configuration and routing of connections (i.e.
tactical planning).
Although the different areas all relate to each other, the functionalities are very
distinct as are the related planning problems. As such, a planning problem typically relates to
only one area. By using the result of the planning problem of one area as the input of the
planning problem of another area, relations between different planning problems can be
considered (e.g. interconnection links of the switching network, serve as demand input to the
transport network). Moreover some of the more competitive operators only have to solve one
of these planning problems because they focus on only one area, such as carrier's carriers
(focus on the transport network), voice resellers (focus on the voice switched network),
internet service providers (focus on the packet switched network) and competitive local
exchange carriers (focus on the access network).
The planning problems considered in this thesis are mainly related to the transport
and access network.
2.5.3 Network planner
Network planning can have very different goals or requirements depending on the
network planner himself. Network planning is often regarded as a process executed by a
network operator. But even then, different network operators have different needs.
An incumbent operator must take into account his legacy network and is concerned
with upgrades of the existing network. These upgrades involve capacity extensions in the near
term and technology changes in the longer term. For the capacity extensions, network
planning tools are required that take into account the current network situation and make pro-
active decision based on the future traffic forecast. For technology upgrades, tools are
required that model the different technologies and design networks based on these
CHAPTER 21
6
technologies to compare the different alternatives under different cost and traffic scenarios.
Such technology changes might be precipitated by the escalating operating cost for the older
equipment or because the older equipment cannot deliver the desired services.
Where an incumbent operator has an existing network, this is not the case for a new
operator. Therefore, one of the main planning decisions for a new operator is to determine the
network topology. This involves deciding in which locations to offer connectivity (points of
presence) and between which points of presence to install transmission facilities. The location
of points of presence often involves demographic and economic market studies, taking into
account traffic forecast of the different geographical areas. Once the points of presence are
determined on, the topology of the transmission facilities can be determined based on
connectivity requirements and installation costs. Next to the actual network topology, also the
network architecture and technology must be decided on. In contrast to an incumbent
operator, a new operator is not constrained by technology already in place. An investment
analysis (see section 2.6.1) for different network technologies over a sustainable period of
time should point out the most suitable technology for the future. Different network
architectures based on these technologies should be evaluated taking into account different
equipment configuration possibilities and disaster recovery strategies. Therefore, tools are
required that allow different network designs to be compared to each other.
Network planning is also of strategic importance to equipment vendors. The
availability of an integrated network planning framework gives an added value to the
equipment vendor, because it helps him in providing turnkey solutions to his customers (the
network operators). This can be either in the format of strategic and tactical consulting
activities executed for network operators. Also some operational planning might be integrated
in the management system of the network equipment to optimize the network usage of the
operator. In addition, network planning tools can also help the vendor to compare network
designs based on his product, with designs based on competing products. From this, the
operator can learn important lessons that can help him in product pricing and marketing. This
is becoming increasingly important, as consolidations among network operators has given the
operators much more power to 'squeeze' vendors into bringing better and cheaper products to
the table. Therefore, network planning also becomes a competitive instrument for vendors, in
order to determine an optimal pricing of their products.
Most of the planning problems considered in this thesis are typical network operator
planning problems, although some of the work also relates to the planning problems of an
equipment vendor.
2.6 Network planning techniques
Although network planning can be done manually on a case by case basis, computer
aided network planning tools, based on sophisticated planning techniques, help in reducing
network planning cycles. By using such tools, the labor-intensive process of manual
calculations can be automated and more scenarios can be evaluated in shorter timeframes. In
this paragraph we will focus on some of the financial and mathematical optimization
techniques used in such tools.
NETWORK PLANNING FOR THE NEW MILLENNIUM 1
7
2.6.1 Strategic network planning: making investment decisions
In strategic planning, as for the technology choice, a detailed network model is not
needed. A high level description of the network architecture is sufficient. Because strategic
planning considers a long time horizon, it is important to know the evolution of the key
parameters influencing the strategic decisions. Such parameters include: number of
customers, bandwidth requirements, equipment prices, interest rates, etc.
While some aspects of strategic planning require human expertise (e.g. for
evaluating and testing future technologies), the main planning techniques used for strategic
planning deal with evaluations of the economic viability of the project under study. Simple
metrics, such as capital costs, revenue, profit, etc. are often not usable, because they are all
differently effected by factors such as income tax and depreciations. In addition such metrics
do not take into account the time value of money. Therefore, other investment criteria have to
be used [15][16]. An example is the net present value (NPV), which expresses the difference
between an investment's market value and its cost, taking into account the time value of
money for future investments and returns (i.e. using discounted cash flows). Also the pay-
back period can be an interesting metric, because is expresses the amount of time required for
an investment to generate cash flows to recover its initial costs. Thus, the shorter the pay back
period, the more interesting the investment. Although the payback period does not use
discounted cash flows, nor does it take into account cash flows after the payback period, this
method can be used as a rule of thumb to get a rough idea of the liquidity and risk of an
investment. Another criterion, is the internal rate of return (IRR), which is the discount rate
that makes the NPV of an investment zero. Based on the IRR, an investment is acceptable if
the IRR exceeds the required return. Other economic parameters, such as present worth of
capital expenditures (PWCE) and present worth of annual charges (PWAC) help determining
the best technical alternative, assuming revenues are constant.
In order to make a more judicious investment decision, it is important to have these
multiple investment criteria. In addition, it is an important task of strategic planning to
consider different scenarios. Using different scenarios, with varying evolutions of the key
parameters, it is possible to perform what-if analysis and sensitivity studies, to assess the
degree of forecasting risk and to identify the key components that are most crucial to success
or failure of future network investments.
2.6.2 Tactical network planning: solving dimensioning problems
Once the network architecture and technology have been chosen, the network
dimensioning process can be invoked. Given a certain traffic forecast, network dimensioning
tries to answer the following questions related with the network roll out:
• Which points of presence to serve (i.e. where to locate the nodes of the network)?
• Between which points of presence to install transmission facilities?
• Routing of connections in the network.
• Dimensioning of the equipment required in the points of presence and of the
transmission equipment.
During the network dimensioning process, one of the main objectives is
minimization of the total cost. The cost structure of the network and appropriate cost values
for the equipment are thus required as input parameters for the design process. The network
CHAPTER 218
design problem can then be modelled as an optimization problem or as a sequence of
optimization problems. When modelling the problem as a sequence of subproblems, the
interaction between the different subproblems is lost, which might result in a sub-optimal
solution. An iterative approach with feedback between the different sequential phases can
partly overcome the loss of optimality incurred with decoupling the subproblems. While a
sequential approach can handle the problem within acceptable computational effort, an
integrated approach can yield a better overall result.
The network design process can be completed by an evaluation phase. In this phase
network simulations are performed to evaluate the blocking performance for dynamic traffic
or to evaluate the network availability. The evaluation phase can in turn give feedback to the
design process.
Most of the encountered problems in network dimensioning are combinatoric
optimization problems for which techniques from operations research can be used. As some
techniques are only applicable for specific problem types, it is important to recognize the
characteristics of the problem, in order to find an appropriate solution technique. Different
types of optimization problems can be identified, depending on the amount of control
variables, the type of the control variables (i.e. real numbers, integers or binary variables) and
the nature of the objective function and constraints (e.g. linear or non-linear function of the
control variables). Most of the problems related to network dimensioning are optimization
problems with multiple integer or binary variables. In many cases the objective function and
constraints can be formulated as linear functions of the control variables, although in some
cases this is not straightforward or possible at all.
For small problems, with a limited number of variables of discrete values, it can be
possible to enumerate all possible solutions and retain the best one. However, for large-sized
problems typically encountered in network dimensioning this is not a feasible method because
the number of possible solutions (and thus calculation time) explodes very rapidly. Therefore
more intelligent optimization techniques are required that use a minimal amount of
calculation time and computation memory. It should be noted that some problems are
inherently difficult, such as the class of NP-complete problems [17]. NP-complete problems
can not be solved to optimality in a time that is proportional to a polynomial function of the
control variables. For such problems, which are typically encountered in network
dimensioning, we often have to fall back on sub-optimal solution methods.
When the objective function and constraints are linear functions of the control
variables, and the control variables are real numbers, the optimization problem is called a
linear program (LP) [18]. Such linear programs can be solved to optimality in polynomial
time. A well know method for such problems, is the Simplex algorithm [18]. When the
control variables are no longer real numbers, but integers or a set of discrete values, we are
dealing with an integer linear program (ILP), which can no longer be solved in polynomial
time. In this case the Simplex algorithm has to be solved multiple times in a so-called branch
and bound tree [18], in which all variables have to be iteratively fixed to integer values.
Sometimes the branch-and-bound tree is 'cut off' at an early stage after a near-optimal solution
is found, or to limit the calculation time. The Simplex algorithm, and many branch-and-bound
techniques are available in commercially available (I)LP packages, such as CPLEX [19].
When the objective function and/or constraints are non-linear functions of the control
variables, other techniques besides (I)LP have to be used. Likewise, when the problem
complexity does no longer allow to solve linear problems using acceptable computation
NETWORK PLANNING FOR THE NEW MILLENNIUM 19
resources, we have to resort to sub-optimal heuristic techniques. These heuristic techniques
can be either generic search heuristics, suitable for any kind of optimization problem, or
problem-specific heuristics. In the class of generic search techniques we find methods such as
simulated annealing [20], tabu search [21] and genetic algorithms [22]. These are basically
intelligent search techniques, that explore the solution space by trying to move from one or
more solutions to better closely related solutions. Although the methods themselves are
generic, they can still be tailored by embedding some intelligence related to the specific
problem into the search process. Problem specific heuristics on the other hand do not quasi-
randomly explore the solution space, but build up a solution based upon historical knowledge
and insight in the problem. This often allows to obtain a good solution in a limited amount of
computation time.
In this thesis, dependent on the characteristics and complexity of the considered
planning problems, we will use both (integer) linear programming techniques as well as
problem specific and generic search heuristics.
2.7 Conclusion
In this introductory chapter, we started by giving an overview of the
telecommunications sector. Great shifts in services, technologies and regulatory environment,
have stimulated high growth and created a competitive marketplace. The different players in
this market have been described. While end-users have benefited substantially from the
increased competition, service providers and network operators face increased network
planning needs to increase their profit margins. Other facets of network planning besides cost
minimization have been discussed as well.
Different planning problems have been identified, depending on the considered time
frame, the geographical or functional area of the network and the requirements and objectives
of the network planner. The main applications of network planning are twofold: the optimal
design or extension of a given network architecture on one hand, and the comparison of
different network architectures and technologies on the other hand. Besides incumbent or new
operators, also equipment vendors see network planning as an important instrument to help
them determine an optimal pricing for their products.
Finally a high level overview of some planning techniques has been given. For more
strategic decisions, a detailed network model is less important, and the planning process
focuses on the comparison of different investment options from a more economical
perspective. Some elementary investment criteria have been described. Once the strategic
decisions have been taken, network planning can focus on dimensioning of the network. The
main aspects of network dimensioning have been described shortly, and an overview of
possible optimization techniques has been given.
Before describing the considered planning problems in more detail, and applying the
above described planning methods, we will first give an overview of the relevant network
technologies and architectures in the next chapter.
CHAPTER 220
2.8 References
[1] L. G. Roberts, "Beyond Moore's law: Internet growth trends", IEEE Computer, Vol. 33, Nr. 1, pp. 117-
119, January 2000.
[2] D. Jennings, "Next generation networks: The vision, the need and the challenge", Proceedings of
NFOEC’99 (Chicago, IL), September 1999.
[3] P. Lagasse et al., "Photonics networks in Europe", Horizon-Infowin (ACTS thematic issue), Telenor AS
R&D, 1998.
[4] A. R. Pach and Z. Papir (eds.), "Broadband access copper technologies", special issue of IEEE
Communications Magazine, Vol. 37, No. 5, May 1999.
[5] D. Greenfield, "Terabit routers: A lesson in carrier-class confusion", Network Magazine, Vol. 15, No. 3,
pp. 78-84, March 2000.
[6] X. Xiao, L. M. Ni, "Internet QoS: A big picture", IEEE Network, Vol. 13, No. 2, pp. 8-18, March-April
1999.
[7] K. Van Miert, "Commission directive 96/19/EC of 13 March 1996 amending directive 90/388/EEC
with regard to the implementation of full competition in telecommunications markets", March 1996.
[8] B. Gates, "Business @ the speed of thought: Using a digital nervous system", Warner Books, 1999.
[9] L. Soete, "Immateriële economie", Post-academic course on Telecommunications strategy and
regulation (in Dutch), Ghent University (Belgium), 1998-1999.
[10] http://screen.yahoo.com/yc/stocks?cc=stocks&n=50&ce_mktcap=C &cf=1&b=1&cs=mktcapt+2
[11] G. Rainey, "Configuration management", in Telecommunications management in the 21st century, pp.
234-267, IEEE Press, 1993.
[12] B. Doshi, P. Harshavardana, "Broadband network infrastructure of the future: Roles of network design
tools in technological deployment strategies", IEEE Communications Magazine, Vol. 36, No. 5, pp. 60-
71, May 1998.
[13] E. Drakopoulos, "Enterprise network planning and design: Methodology and application", Lucent
(Optical networking group) White Paper, 1999.
[14] P. Demeester, "Telecommunicatienetwerken A", course at the Ghent University, 1999-2000.
[15] S. A. Ross, R. Westerfield, B. D. Jordan, "Fundamentals of corporate finance", Irwin/McGraw-Hill,
1997.
[16] E. L. Grant, W. G. Areson, W. G. Ireson, "Principles of engineering economy", John Wiley & Sons,
1990.
[17] M. R. Garey, D. S. Johnson, "Computers and intractability : A guide to the theory of NP-completeness",
W. H. Freeman & Co, 1979.
[18] G. Nemhauser, L. Wolsey, "Integer and combinatorial optimization", John Wiley & Sons, 1998.
[19] ILOG Inc., CPLEX Division ,"Using the CPLEX callable library", 1997.
[20] S. Kirkpatrick, C.D. Gelatt, Jr. and M.P. Vecchi, "Optimization by Simulated Annealing", Science, Vol.
220, No. 4598, May 1983.
[21] F. Glover, M. Laguna, "Tabu Search", Kluwer Academic Publishers, 1997.
[22] D. Goldberg, "Genetic algorithms in search, optimization and machine learning", Addison-Wesley,
1989.
CHAPTER 3
Transport network technologies and
architectures
3.1 Introduction
In this chapter, we give an overview on the technologies and architectures that are
adopted in current and future transport networks, and for which different planning problems
have been studied in the subsequent chapters. We start by giving an overview of the general
structure and required functionalities of a transport network. Afterwards the enabling
technologies are introduced. The focus of this chapter is on transport network technologies,
for which the protagonists are the synchronous digital hierarchy (SDH) and the wavelength
division multiplexing (WDM) technologies. After an in-depth discussion on SDH and WDM,
the emphasis is shifted temporarily towards packet-switched technologies. We give a short
overview of two main packet-switched technologies, the asynchronous transfer mode (ATM)
and the internet protocol (IP), without going in much detail, because the planning problems in
this thesis do not deal with packet-switched networks. Nevertheless, we feel it is important to
mention these technologies, because an integration of packet switching in transport networks
is being witnessed in the new technologies that are entering the market. An overview of this
convergence between both technologies is given, with the focus on IP-over-WDM enabling
solutions. The last part of this chapter deals with reliable network architectures. Recovery
mechanisms, for different topologies (both ring and meshed networks) and technologies (SDH
and WDM) are described and compared.
3.2 Transport network structure
The goal of every communications network is to transport information (voice, video,
data, …) between the different users of the network. Hence, there is a need for a transmission
medium between all users of the network. For very small networks, a direct transmission line
between each two users of the network can be established, such that all users can directly
transfer information between each other. However, for large-scale networks, this is not a very
scalable solution. Each time a new user has to be connected to the network, a large number of
extra transmission lines (equal to the number of current users) has to be provisioned, resulting
in a very complex and expensive architecture. In addition, since not all users interact at the
same time, a large number of transmission lines does not transport any information, resulting
in a very low network utilization. To overcome these drawbacks, the network has been
structured in a more hierarchical way as depicted in Figure 1. Instead of connecting each user
directly to another user, each user is connected to a voice switch or data router in the local
CHAPTER 322
exchange of the access network [1][2]. The switch or router acts as a flexibility point, which
enables to exchange information (i.e. set up connections or forward data packets) between the
different users that are connected to it. To reduce the cost of the access network (which is
dominated by the installation cost of the transmission lines between the users and the local
exchange), several local exchanges are created in different geographical regions, such that the
distance between the user and the nearby local exchange is limited. In order to allow
interaction between users in different geographical regions (connected to different local
exchanges), the different local exchanges have to be connected to each other. This is done by
providing interconnection links between local exchanges and between a local exchange and a
so-called transit exchange. A transit exchange acts as an intermediate flexibility point which
can switch traffic between the different local exchanges connected to it. The network
responsible for providing the interconnection links is referred to as the transport network
[1][3]. The local exchanges and transit exchanges form the nodes of the transport network,
while the transmission facilities interconnecting the different exchanges are the links of the
transport network.
access network
metropolitan
network
regional
network
core
network
tier 1
tier 2
tier 3
Transport Access
LEX LEX
Figure 1: Network structure
The main function of the transport network is thus to provide interconnection links.
The transport network as such acts as a server layer, providing high-speed trunks to its client
layers such as the telephony and data network. To reduce cost and complexity, also the
transport network often has a structure consisting of several hierarchical levels or so-called
tiers. The different local exchanges in a city can be connected to a transit exchange in what
we call the metropolitan transport network. Such ‘metro networks’ are only found in large
business cities, and not in smaller cities, in which typically only one or two local exchanges
are present. Metropolitan transport networks of cities within the same geographical region can
be interconnected through the regional transport network. Different metropolitan or regional
transport networks can again be interconnected through the core transport network. In this
example we recognize a 3-tier structure, enabling all users to interconnect to each other.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 23
The main requirements of the transport network depend on the client layer(s) it
serves and on the hierarchical levels (tiers).
Interconnection trunks offered by the transport network between voice switches
should be subject to low delay and high reliability in case of network failures. In addition, the
transport network for voice telephony has to take into account the historical multiplexing
structure in discrete steps, going from a basic 64 Kb/s all the way up to 2.5 or even 10 Gb/s
(see section 3.4.2). For the transport network that interconnects the routers of the data
network, the requirements are somewhat different. There is no specific need for a rigid
multiplexing structure. Interconnection trunks between routers should simply be flexible high
capacity pipes with low cell tax (i.e. low overhead capacity). In addition, transport network
reliability is less critical for some classes of data traffic (even more, the routers can auto-
detect topology changes due to failures and thus reroute traffic themselves). In order to reduce
the cost and management overhead, these different requirements of the client layers should be
available within the same transport network, rather than implementing separate transport
networks for the different client layers.
Besides the different client layers, also the hierarchical levels of the transport
network can have different requirements. The higher level tiers (i.e. closest to the end users)
should typically be highly flexible and manageable with respect to configuration possibilities,
in order to adapt to sudden changes in customer locations or traffic patterns. In addition, since
the higher tiers deliver a large contribution to the total transport network cost, low-cost
implementations and adequate network planning are of paramount importance. Lower tiers on
the other hand require less configuration flexibility, as traffic patterns change less suddenly,
but should be easily upgradeable to deal with the continuous traffic growth. Lower tiers
typically require architectures capable of delivering very high speed interconnection capacity.
Because of the tremendous amount of traffic passing though the lower tiers, reliability is also
a critical issue. Due to the large distances typically spanned for lower tiers, physical
impairments of transmission systems should also be taken into account. Since all the different
tiers represent hierarchical separated networks, there is no need for an integrated transport
network technology fulfilling the needs of all the different tiers, but in contrast different
technologies and architectures can be deployed as suited (e.g. SDH ring architectures in
metropolitan networks and WDM mesh architectures in core networks).
3.3 Transport network functionalities
While the previous paragraph described the structure and architecture of the transport
network, this paragraph will focus on the required functionalities that need to be implemented
in the network. We will start with the layering and partitioning concept, because it allows to
simplify the description of a network consisting of multiple technologies and domains. Then
we describe some generic functional requirements, such as multiplexing, cross-connecting,
consolidation, segregation and grooming. We also treat some more operational requirements,
such as the need for network management, monitoring and reliability.
3.3.1 Layering and partitioning
In order to simplify the description of a transport network consisting of multiple
technologies, we can use a model based on the concepts of layering and partitioning, as
CHAPTER 324
described in G.805 [4]. One of the essential features of G.805 is that it has identified the
generic functionality of a transport network, which is independent of the implementation
technology. As such, a model of a transport network based on different technologies can be
described with a high degree of recursiveness. Each technology within the network can be
described using one or more layers, each with its own topology and functionality.
3.3.1.1 Partitioning
In general each layer network may be decomposed into subnetworks and links
between them. Each subnetwork may be further decomposed into smaller subnetworks
interconnected by links, revealing the connectivity supported by the containing subnetwork,
in a recursive way until the desired level of detail is revealed. At the limit, the subnetwork can
be decomposed into atomic subnetworks called matrices, which represent the functional
equivalent of a cross-connect or add-drop multiplexer (see section 3.3.3). This process of
decomposition is referred to as partitioning. Examples of subnetworks are the different tiers
in a layer network. A representation of a layer network, partitioned in subnetworks, is shown
in Figure 2.
Figure 2: Partitioning of a layer network in subnetworks
According to G.805, the partitioning concept is important as a framework for
defining:
- the network structure within a layer network;
- administrative boundaries between network operators jointly providing connections
within a single layer network;
- domain boundaries within a layer network of a single operator to allow the
apportioning of performance objectives to the architectural components;
- routing domain boundaries within the layer network of a single operator;
- the part of a layer network or subnetwork that is under the control of a third party for
routing purpose (e.g. customer network management).
3.3.1.2 Layering
A transport network can be decomposed into a number of independent layer
networks with a client/server relationship between adjacent layer networks: server layers
provide transport facilities from which traffic in the client layer can make use. Within each
layer characteristic information, specific for the layer, is transported across a network
connection. A network connection transparently transfers characteristic information between
two termination connection points (tcp) of the layer. A trail refers to the transfer of monitored
characteristic information. A trail is transported between access points (ap), which form the
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 25
boundaries of the layer. A trail is formed by associating trail terminations with a network
connection.
network connection
link connection matrix
connection
trail termination
Trail
ap ap
tcp
layer network
matrix
link
cp
cp
tcp
Figure 3: Layer network
The topology of a layer is represented by links and matrices. A matrix represents the
atomic limit to the recursive partitioning of a subnetwork. A link represents the topological
relationship and available transport capacity between a matrix or an access group and another
matrix or access group. An access group represents the access interface to a layer network, as
a group of co-located trail termination functions connected to the same link or subnetwork. A
link can contain multiple link connections, which are responsible for the transfer of the
characteristic information across a link between connection points (cp). The matrix
connections represent flexible transport entities between connection points on the boundary of
the matrix. A network connection is formed by the concatenation of link connections and
subnetwork connections. At the lowest level of detail, subnetwork connections consist of a
concatenation of link connections and matrix connections. These subnetwork connections are
thus transport entities between connection points at the boundaries of a subnetwork.
Now we have defined the architectural components of a layer network, the
client/server relationship between adjacent layer networks can also be better explained. As
depicted in Figure 4, a link connection in the higher client layer is supported by a trail in the
lower server layer network. Client layer information must be adapted for transmission in the
server layer. This process, called adaptation, occurs between the connection point of the link
connection and the access point of the trail. The nature of the adaptation process depends on
the characteristic information within each layer, but typically involves transmission rate
changing, multiplexing, aligning, coding or some combination of these.
CHAPTER 32
6
Figure 4: Multi-layer representation
According to G.805, the layering concept of the transport network allows:
- each layer network to be described using similar functions;
- the independent design and operation of each layer network;
- each layer network to have its own operations, diagnostics and automatic failure
recovery capability;
- the possibility of adding or modifying a layer network without affecting other layer
networks from the architectural viewpoint;
- simple modelling of networks that contain multiple transport technologies.
3.3.1.3 Circuit, path and transmission media layers
The layering and partitioning principle described above are generic, applying to all
networking technologies. The network layers have been classified into three broad categories:
circuit layers, path layers, and transmission media layers.
Circuit layers provide circuits for the telecommunication services between end users.
The switched telephony network and the packet switched network are examples of circuit
layer networks. Circuits, i.e. circuit layer trails, terminate at the user and their connectivity is
controlled by a process invoked by the user or application.
Path layers provide transport services to the circuit layers, or other path layers.
Paths, i.e. path layer trails, terminate at the circuit layer, or other path layer access points and
their connectivity is determined by the transport requirements of the client service layers (or
client path layers) and controlled by the path layer trail management (hence by the network
operator).
The transmission media layers provide transport services to the path layers.
Transmission media layers can in turn be subdivided in section layers and physical media
layers. The section layers define the format of the transport signal, while the physical media
layers take the physical characteristics of the transmission medium into account. Although
sharing the generic properties common to all layer networks, they are specialized according to
the media for which they have been designed, be it coaxial cable, optical fiber or radio
transmission.
3.3.2 Multiplexing
The transport network was initially designed to interconnect different voice switches
with each other. The concept of multiplexing [1] enables several lower rate signals (also called
tributary signals) to be grouped in one higher rate signal (also called aggregate signal),
which can then be transported at once with reduced cost and complexity. The act of
multiplexing is thus part of the adaptation process in going from a higher layer towards a
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 2
7
lower layer in a multi-layer network. Conversely, demultiplexing disassembles the aggregate
signal again into individual tributary signal, when going from a lower towards a higher layer.
An example of such multiplexing is the 2 Mb/s aggregate signal, comprising 30
digitized voice channels at 64 Kb/s (see section 3.4.2.1), which is used to interconnect the
switches in the telephone network. This example illustrates the concept of time division
multiplexing (TDM), in which more time slots are transmitted per unit of time, such that the
different time slots can comprise different lower rate signals. Other multiplexing alternatives
are space division multiplexing (SDM), which does not use higher bitrates, but simply uses
multiple transmission lines in parallel. For example, SDM is used in the last drop section of
the telephone network, where large cables with thousands of copper pairs run up to the
subscribers premises. Frequency division multiplexing (FDM) on the other hand modulates
different lower rate signals on a different frequency, to transmit multiple signals on the same
transmission medium. FDM was used in the early days of analogue telephony to carry large
numbers of telephony circuits on a single coaxial cable. A recently introduced variant of FDM
is wavelength division multiplexing (WDM), transmitting information on different
wavelengths of an optical fiber (see section 3.4.3).
An important requirement in multiplexing is to find the right balance in the
multiplexing scheme. While it might be more efficient to directly map thousands of telephone
channels in a 2.5 Gb/s signal, this is not a very scalable approach. Using several intermediate
multiplexing steps on the other hand adds extra overhead, but allows to grow the network
along with the traffic requirements. In addition, one large multiplexing step requires a
complete demultiplexing of the entire signal to extract one channel, which is not needed with
a more gradual multiplexing scheme. Some recent products brought to the market have a
multiplexing scheme in which the granularities of the multiplexing steps can be adapted,
which adds increased flexibility to the network.
3.3.3 Cross-connecting
The transport network is responsible for interconnecting the switches and routers at
the different local exchanges. The voice switches in the telephone network are often
interconnected with a fully meshed network. This implies there is at least one direct 2 Mb/s
connection between every 2 switches. In practice, this is not realized by a direct physical (e.g.
cable) connection between every two switches, because this would be a too costly
implementation. The principle of cross-connecting [1][3] allows to terminate and interconnect
different tributary signals between the different incoming and outgoing aggregate links of a
node in the transport network. E.g. Figure 5, represents a node terminating three aggregate
links, which cross-connects three tributary signals between each two links. Thereby, semi-
permanent tributary connections can be set up in a flexible way between 2 nodes in the
transport network. The cross-connected tributary signals are aggregate links of their own for
the tributary signal they contain, which can then all be switched at once. As such, a fully
meshed logical network can be created on top of an underlying transmission network that has
a much sparser topology (this is the principle of layering as explained in section 3.3.1.2). To
support the increase in traffic and the varying traffic patters, the cross-connection
functionality can be made flexibly rearrangeable (with or without the control of a central
management system).
CHAPTER 328
Figure 5: Cross-connecting
A cross-connect (XC) typically has a large flexibility, interconnecting multiple
aggregate links and tributary ports (for locally terminated traffic) using a reconfigurable
matrix. A more downsized network element, is the add/drop multiplexer (ADM), which only
terminates two aggregate links and provides pass-through and local add/drop of tributaries.
ADMs are suitable network elements for implementing ring-based networks. Besides the
capacity management capabilities of XCs and ADMs, these network elements are also
suitable for executing network recovery actions (see section 3.7).
3.3.4 Consolidation, segregation and grooming
Consolidation means multiplexing traffic stemming from different locations into one
facility. As such, aggregate signals that are only partially filled with tributary signals, can be
taken together in one combined aggregate signal, thereby increasing the filling ratio of the
links in the network. Consolidation thus has a positive effect on the link cost, but because of
the extra complexity needed in the nodes of the network, it has a negative effect on the node
cost. Segregation refers to the separation of traffic (e.g. into different traffic streams going to
different locations). Grooming involves the consolidation and segregation of traffic for the
purpose of increasing the network efficiency, e.g. by consolidating traffic streams going to the
same location (grooming by destination). Grooming can also provide segregation of traffic
streams based on the type of traffic, e.g. for voice and data traffic or based on survivability
requirements (e.g. protected and unprotected traffic). This is summarized in Figure 6.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 29
1
-
-
-
-
3
-
1
-
3
-
-
-
2
-13132
12313 1
-
-
-
2
-
-
-
3
1
-
-
-
-
3
3
21
-3
12
--
2
1
2
3
1
2
3
1
2
3
1
2
3
1
2
3
12- 31
-3213
Consolidation
Segregation
Grooming
Figure 6: Consolidation, segregation and grooming
3.3.5 Management
The above described transport network functions deal with how the information in
the network is processed. This refers to the so-called data plane of the network. On the other
hand, there is also a control plane, which deals with how the network is managed. This
network management refers, amongst others, to monitoring and control of the network and its
constituting elements. Network monitoring is concerned with observing and analyzing the
status and behavior of the network and its elements. Network control is concerned with
altering the configuration of various elements and causing those elements to perform
predefined actions.
The Telecommunications Management Network (TMN) [5] standard defines the
following functional areas of network management:
- Fault management, which provides a set of functions that provide the ability to detect,
isolate, report and correct abnormal operations in the network.
- Configuration management, which is responsible for offering a technology
independent view of the physical resources and control these resources such to
provision services across the network.
- Accounting management, which enables charges to be established for the use of the
network and its elements.
- Performance management, which allows to monitor and report parameters that reflect
the effectiveness of the network and its elements and the quality of the services.
- Security management, which prevents unauthorized users from obtaining access to
certain resources or information and to protect the network elements.
The above set of functional areas are collectively referred to as FCAPS.
The TMN model is sometimes represented as a layered pyramid (see Figure 7) with
the following components, from top to bottom: business management, service management,
network management, network element management and network elements [6].
CHAPTER 330
Network Element Layer
Element Management Layer
Network
Management Layer
Service
Management Layer
Business
Management Layer
Figure 7: TMN layered model
The Network Element Layer (NEL) is where network equipment (network elements,
or NEs) operates and collects data on how well the equipment is working. The Element
Management Layer (EML) provides management of the network elements (NEs) individually
or in aggregation as a subnetwork. This layer collects data from the NEs and this data is pre-
processed for managing the NEs and to support higher-layer functions. The Network
Management Layer (NML) is where end-to-end control of the network occurs. Information
processing activities focus on statistics and data involving complete connections and the
identification of customers that are affected by the performance of those connections. The
Service Management Layer (SML) is where customers are furnished with billing and usage
reports and also get statistical proof of network quality, also called Quality of Service (QoS),
either by hard copy or electronic means. This is the layer at which customers are afforded
access to certain network management activities to tailor the network's cost and capabilities
more closely to their needs. Sometimes a last layer is added which is called the Business
Management Layer (BML), and represents high-level business planning: product planning,
budgeting, continuous improvement, external relationships and legal arrangements.
3.3.6 Connection monitoring
Monitoring of the integrity, validity and quality of connections in the network is a
task of performance management, which is of paramount importance in order to supervise the
network performance and take appropriate actions in case problems occur. In layered
networks, different options exist to monitor a connection, depending on what kind of
information is available. Connections can either be monitored end-to-end, or only a segment
can be monitored in case a connection is transported across different administrative domains.
A tandem connection represents such a segment of a trail that requires monitoring
independently of the monitoring of the complete trail. The different methods for the
monitoring will be described hereunder [4][7].
3.3.6.1 Inherent monitoring
The performance of a client layer connection can be inherently monitored through
the supporting server layers. If a trail in the server layer network fails then it may provide an
indication (e.g. alarm indication signal) at the output of the client layer link connections that
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 31
are being supported. Based on this client server relationship, we can deduct the integrity and
performance of the client layer connection by collecting and correlating (via a management
network) the information of each link connection and subnetwork connections that forms the
connection of interest. Since adaptation functions and matrix connections are not included in
this monitoring scheme, the overall status of the connection can not be verified.
3.3.6.2 Non-intrusive monitoring
Non-intrusive monitoring of a server layer connection is defined as listen-only
(without modifying it) monitoring of the overhead of the client layer signal. This requires the
accessibility of compatible trail terminations capable of monitoring the client layer signal and
overhead at intermediate connection points. The information derived from this monitoring
reflects the status of the connection from the trail termination source to the intermediate
connection point. The status of an intermediate segment of a connection, can be derived by
correlation, through non-intrusive monitoring of the connection points that delimit the
segment. Non-intrusive monitoring provides a useful complement to inherent monitoring.
Marginal error performance glitches or defects at the adaptation functions that may evade
detection by inherent monitoring can as such be detected.
3.3.6.3 Intrusive monitoring
Intrusive monitoring implies breaking the original trail and inserting a test trail over
the connection of interest. As such more comprehensive tests can be performed, but this can
only be used when the user trail is out of service, which can only be done at the beginning of
trail setup, or else the user trail has to be interrupted.
3.3.6.4 Sublayer monitoring
Sublayer monitoring uses an additional sublayer trail below the original trail for
monitoring purposes. Some portion of the original trail's capacity is overwritten, such that the
part of the connection of interest can be directly monitored by the as such created sublayer
trail. Assuming sufficient bandwidth can be overwritten, this technique allows comprehensive
testing.
3.3.7 Maintaining network integrity
It is not sufficient to monitor the failures that have occurred, but also to take pro-
active and reactive measures such that the impact of these failures can be limited. Pro-active
measures can be taken by implementing reliable software architectures and hardware
components for fulfilling the transport network functions. Reliable software architectures rely
on the adoption of a standardized framework for defining, implementing, maintaining and
testing the software, such as described in ISO 9000-3 (a software specific interpretation of the
ISO 9001 standard for quality assurance) [8]. On the other hand, reliable hardware
components with low failure rates, ensure a long-lasting operation of network elements. In
addition the critical components in the network elements can be duplicated, such that a stand-
by component can be activated in case the active component fails.
Besides pro-active measures, also reactive actions are required in order to respond to
unexpected failures. Examples of such failures are a cable cut (e.g. because of road works) or
a catastrophic node failure (e.g. due to fire). In order to recover all or part of the affected
CHAPTER 332
traffic, some spare resources and rearrangements in the network are required such that the
traffic can be rerouted avoiding the failure. This rearrangement can be manually executed or
can be automated with or without intervention of the network management. Each of the layers
can support its own suite of recovery mechanisms, specific for the technology of that layer. In
section 3.7, some recovery mechanisms for transport network technologies such as SDH and
WDM will be discussed in more detail. A topic of particular interest related to network
recovery in a multi-layer network, is how to provide and co-ordinate recovery mechanisms in
the different layers of the network (see section 3.7.6).
3.4 Transport network technologies
3.4.1 Introduction
In this paragraph we focus on transport network technologies, in which
communications is made possible through semi-permanent dedicated connections in the
network. Such a connection typically has a predefined (fixed) capacity and route. These
connections are not disrupted at any time during the period that information needs to be
exchanged: this is the semi-permanent character of the connection. The dedicated nature
means that once resources have been reserved for a connection, they can not be used by
another connection. In transport networks, blocking can only occur during connection set-up;
however, once connection set-up is successful and sufficient resources are reserved, the
network will not lose data due to congestion.
In the remainder of this paragraph, we describe two transport network technologies
that will play a prominent role in this thesis: Synchronous Digital Hierarchy (SDH) and
Wavelength Division Multiplexing (WDM).
3.4.2 Synchronous Digital Hierarchy (SDH)
3.4.2.1 The early days: PDH
The plesiochronous digital hierarchy (PDH) [9] was the first effort to standardize the
TDM multiplexing scheme. The mapping of 30 voice channels in a 2 Mb/s signal as already
described in section 3.3.2 is performed according to the PDH standard. Each 2 Mb/s signal
(also referred to as E1 signal) comprises 30 digitized voice channels at 64 Kb/s (and 2
channels used for management and signaling purposes). This multiplexing step is
synchronized and uses byte interleaved multiplexing. For higher multiplexing steps, the
different tributary signals can have slightly different bit rates. In PDH networks the bit rate of
each tributary signal is controlled within a specific limit and is not synchronized with the
multiplexing equipment (hence plesiochronous or nearly synchronized). To multiplex such
signals, bit interleaved multiplexing is used together with bit stuffing to correct the bit rate
anomalies.
A disadvantage of PDH is that the multiplexing scheme is different in Europe and in
the US and Japan. There the basic signal consists of 24 channels of 64 Kb/s, which are
multiplexed in a 1.5 Mb/s signal (also referred to as DS1 signal). Another disadvantage of
PDH is that due to bit stuffing, access to the tributaries is impossible without demultiplexing
the whole aggregate signal. Also different frame structures at different bit rates and the
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 33
limited network management, maintenance and fault recovery capabilities have restricted the
adoption of PDH as a future proof technology.
An overview of the different multiplexing levels that are used in Europe and the US
are shown in Table 1. A European PDH channel of multiplexing level x is also referred to as
an Ex signal, while a US PDH channel of multiplexing level y is referred to as a DSy (also Ty)
signal. For example, the European PDH channel of level 3, multiplexing 512 64 Kb/s
channels, is referred to as an E3 signal.
European multiplexing scheme US multiplexing scheme
Multiplexing
level 64 Kb/s channels Bit rate 64 Kb/s channels Bit rate
0 1 64 Kb/s 1 64 Kb/s
1 32 2.048 Mb/s 24 1.544 Mb/s
2 128 8.448 Mb/s 96 6.312 Mb/s
3 512 34.368 Mb/s 672 44.736 Mb/s
4 2048 139.264 Mb/s 4032 274.176 Mb/s
Table 1: PDH multiplexing levels
3.4.2.2 SDH
In the synchronous digital hierarchy (SDH) technology [10][11], the network is
completely synchronized (all clocks are obtained through a synchronization network) and byte
interleaved multiplexing is used for all multiplexing steps. As such, direct synchronous
multiplexing is enabled, which allows individual tributary signals to be directly multiplexed
into higher rate SDH signals, thereby enabling cost-effective and flexible networks. In
addition, SDH has built-in overhead (restricted to about 5% of the total signal capacity) for
advanced network management, maintenance and fault recovery capabilities. This world-wide
standard theoretically allows to create a multi-vendor network. However, some of the
overhead bytes have not been defined in the standards, and are used by different vendors for
proprietary reasons, making interworking between equipment from different vendors
sometimes impossible in practise. An even more limiting factor, is the fact that in the US, a
different synchronous multiplexing hierarchy was defined, namely SONET (synchronous
optical network) [12].
The information structure used for transmission of SDH signals is called the
synchronous transport module (STM). An STM consists of payload and overhead information
fields organized in a block frame structure, which repeats every 125 microseconds. The basic
STM frame is STM-1 and has a signal rate of 155.52 Mb/s. Higher level signals are obtained
through byte-interleaved multilplexing of STM-1 signals and are denoted as STM-N signals,
where N is integer (typically multiples of 4: 4, 16, 64). The basic frame is organized as series
of 2430 bytes, which are typically represented as a matrix of 9 rows and 270 columns (see
Figure 8 for N=1). The first 9 columns contain section overhead (SOH) and administrative
unit (AU) pointers (see further), while the remaining 161 columns are reserved for payload
information.
CHAPTER 334
Figure 8: STM-N frame structure
The payload information field contains one or more higher order virtual containers
of order n (VC-n), with n = 3 or 4. The position of the higher order VC-n within the STM-N
frame is defined by the AU pointer. Together with the higher order VC-n, the AU pointer
defines the administrative unit of order n (AU-n). A homogeneous assembly of AU-3s or an
AU-4, occupying well-defined positions in the STM-N payload is called an administrative
unit group (AUG). A higher order VC-n can comprise a single container (C-n), in which
actual information can be transported, together with a virtual container path overhead (POH)
appropriate to that level. Adaptation functions have been defined for many common network
rates (e.g. PDH) into a limited number of standard containers. The C-4 container of a VC-4
(149.76 Mb/s) can comprise the European E-4 PDH signal (139.264 Mb/s), while the C-3
containers of a VC-3 (48.384 Mb/s) can comprise the US DS-3 PDH signal (44.736 Mb/s).
Besides a single container, the higher order VC-n can also contain a number of lower order
virtual containers of order n (VC-n) (n = 1 or 2) assembled in tributary unit groups (TUG). A
lower order VC-n also consists of a single container (appropriate for transport of low-rate
signals), together with a virtual container path overhead (POH) appropriate to that level. The
position of the lower order VC-n within the higher order VC-n is determined by the tributary
unit (TU) pointer. The lower order VC-n and the TU pointer define the tributary unit of order
n (TU-n). One or more tributary units, occupying well-defined positions in a higher order VC-
n payload, is called a tributary unit group (TUG). TUGs are defined in such a way that mixed
capacity payloads made up of different size tributary units can be easily constructed to
increase flexibility of the transport network. A TUG-2 can consist of a homogeneous
assembly of identical TU-1s or a single TU-2. A TUG-3 can consist of a homogeneous
assembly of TUG-2s or a single TU-3. This multiplexing structure is summarized in Figure 9.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 35
STM-N AUG VC-4AU-4
AU-3 VC-3
TUG-3
TUG-2
TU-3
TU-2
TU-1
VC-3
VC-2
VC-1
C-4
C-3
C-2
C-1
Figure 9: Multiplexing structure
To convey signals with a bitrate that is not in accordance with the bitrates supported
by the different containers, the possibility to concatenate TUs or AUs exists. When
concatenating AUs or TUs, the combined units should be looked upon as one unit with the
aggregate bitrate of all units. As such m TU-2s can be concatenated in a higher order VC-3 or
VC-4 resulting in a TU-2mc. For further details about the implementation of this
concatenation function (for which several implementations exist), we refer to [7]. To support
bitrates higher than those offered by a C-4 (i.e. 149.76 Mb/s), AU-4s can be concatenated as
well. The payload of the STM-N signal in which N AU-4s are concatenated, should then be
looked upon as one large virtual container, a VC-4Nc. The first row of this VC-4N is used for
the POH, while the next N-1 rows are filled with fixed stuff bytes and the remaining rows
form one large C-4N container with the capacity of N C-4 containers.
3.4.2.3 SDH layers
Circuit layer neworks
VC-11 VC-12 VC-2 VC-3
VC-3 VC-4
+ AUG + MSOH
+ RSOH (= STM-N)
Optical fiber, radio link, coax
circuit
layer
lower order
path layer
higher order
path layer
multiplex
section layer
regenerator
section layer
physical
media layer
path layer
section layer
Figure 10: SDH layers
CHAPTER 33
6
The layering concept as described in 3.3.1 can be directly applied to SDH networks
as depicted in Figure 10. The lowest layer of any network is the physical media layer, which
defines the physical medium and the modulation format used for the transport of information.
In case of SDH, his physical medium can be an optical fiber, a coaxial cable or a radio link.
The SDH signal can be directly transported over the physical layer [13][14], or an
intermediate server layer (e.g. WDM) can be used. The regenerator section (RS) layer is
transported over the physical medium (or the intermediate server layer), and is responsible for
correctly conveying the STM-N frame. Due to signal degradations in the physical layer, it is
necessary to 'clean up' the signal at intermediate points to ensure correct transmission, using
so-called regenerators, which terminate the regenerator section layer. This regeneration
involves reamplification, reshaping and retiming of the signal (called 3R regeneration).
Besides, these points where the RS is terminated, also form a monitoring point in the network.
The overhead associated with the regenerator section (RSOH) is also transported in the STM-
N frame, namely in the first 3 rows of the section overhead. The most important bytes in the
RSOH are the framing bytes A1 and A2 (which indicate the start of the frame), the trace byte
J0 (to verify the continuity of the RS) and the parity check byte B1 (for error performance
monitoring).
The next layer in the SDH network is the multiplex section (MS) layer. Each
multiplex section consists of a set of regenerator sections, following the client/server
relationship. A multiplex section is typically defined between two multiplexers, ADMs or
cross-connects. The main function of the multiplex section layer is to multiplex a number of
AUs in the STM-N frame. The overhead associated with the multiplex section (MSOH) is
also transported in the STM-N frame, namely in the last 5 rows of the section overhead. The
parity check bytes B2 in the MSOH allow performance monitoring of the MS. Other
important bytes in the MSOH are the K1 and K2 bytes, which are used for signaling
automatic protection switching (see section 3.7) messages between the network elements.
The multiplex section layer serves the path layers, which are the layers which
assemble traffic in VCs and provide the cross-connecting flexibility. We distinguish two path
layers: the higher order path (HOP) layer and the lower order path (LOP) layer1. Each higher
order path is transported over a set of multiplex sections. The characteristic information of the
HOP layer is a VC-4 (or VC-3). Each HOP connection can transport native HOP traffic (i.e.
150 Mb/s connections) or it can also transport aggregate lower order path link connections.
This LOP layer provides flexibility at a lower granularity. The characteristic information of
this layer is a VC-3, VC-2, VC-11 or VC-12.
3.4.2.4 SONET
The main difference between SONET and SDH is the transmission rate used for the
basic frame. Whereas in SDH, an STM-1 is used with a bit rate of 155.53 Mb/s, SONET
relies on a Synchronous Transport Signal Level 1 (STS-1) of 51.84 Mb/s. Higher level STS
frames are typically multiples of 3 of the STS-1. The optical interface for the STS-N signal is
the Optical Carrier Level N (OC-N). The most popular STS rates and the correspondent STM
rates in SDH are represented in Table 2.
1 Note the confusing nomenclature: the higher order path layer is a lower layer than the lower order path layer. It
is called higher order because of the higher bit rate.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 3
7
OC Level Bit rate (Mb/s) STS Level (SONET) STM Level (SDH)
OC-1 51.48 STS-1 -
OC-3 155.52 STS-3 STM-1
OC-12 622.08 STS-12 STM-4
OC-48 2488.32 STS-48 STM-16
OC-192 9953.28 STS-192 STM-64
Table 2: SONET and SDH levels
The STS-1 frame can contain a number of Virtual Tributaries (VT), similar to virtual
containers. The VT levels have been chosen such that they efficiently transport, the legacy
asynchronous digital signals used in the US, and are therefore different from the VC levels
used in SDH.
3.4.2.5 SDH equipment
SDH is a proven technology, and equipment has been established for a number of
years. The main vendors offer products in three categories: terminal multiplexers (TM), add-
drop multiplexers (ADM) and digital cross-connects (DXC).
A. Terminal Multiplexers
Terminal multiplexers combine several lower rate signals in a single higher rate SDH
signal. TMs exist for the different multiplexing steps as described in G.782. Access
multiplexers typically multiplex E1s and/or E3s in an STM-1 signal, while core multiplexers
typically multiplex STM-1 signals in an STM-N signal (N = 4, 16, 64). Of course, many
intermediate forms exist. As multiplexers are deployed in point-to-point applications, their
functionality is often limited besides the pure multiplexing.
B. Add-Drop Multiplexers
Add/drop multiplexers are an extension of TMs, and can also be used as TM. In
contrast to TMs, ADMs have two aggregate sides and one side for add/drop of tributaries. The
line-rate of the aggregate sides can range from STM-1 all the way up to STM-64, depending
on whether we are dealing with an access ADM or a core ADM. The STM-16 rate is most
popular in today’s transport networks. Two main types of ADMs can be distinguished: LO
ADMs and HO ADMs. LO ADMs can add/drop traffic on the LO and in most cases also the
HO level, while HO ADMs can only access the HO traffic. Thus LO ADMs require an
internal matrix that can treat both LO traffic and HO traffic. Typically both functionalities are
divided over distinct matrices. To limit the size of both matrices (in particular the LO matrix),
the LO ADM is often limited in add/drop capability: typically only 25 or 50% of the total
capacity can be added/dropped, while the remaining part must be passed through. Such ADMs
are also referred to as ‘compact’ ADMs. HO ADMs only require a HO matrix and can thus be
easily made fully non-blocking. LO ADMs can be found with aggregates from STM-1 up to
STM-16. Tributaries rates typically include E1, E3, E4 and (optical or electrical) STM-1. The
ADM can be configured for the various combinations in add/drop rates, by installing the
appropriate tributary cards on the ADM. HO ADMs typically range from STM-16 up to
STM-64. The STM-64 ADM can support tributaries up to STM-16, to interconnect high-
speed internet routers with 2.5 Gb/s output links.
CHAPTER 338
ADMs typically host a suite of protection mechanisms optimized for ring topologies
Both HO and LO ADMs support the most common protocols (linear MSP, SNCP, MS-
SPRing, NUT, RIP, drop & continue, …), although some protection protocols can only be
applied on the HO level or even multiplex section level (for more details on the different
protection mechanisms, we refer to section 3.7).
C. Cross-Connects
SDH cross-connects typically host a large set of input/output ports at the STM-1
level, between which traffic can be cross-connected. Most DXCs are fully non-blocking,
meaning that traffic from any input port can be routed to any output port, by appropriately
configuring the DXC. DXCs come in two basic flavors. The DXC 4/3/1 (also referred to as
wideband DXC) has an internal switching matrix on the LO level. This means LO signals can
be segregated and consolidated in different HO signals inside the DXC. In contrast, the DXC
4/4 (also known as broadband DXC), can only cross-connect HO signals and no internal
consolidation is possible. In fact a broadband cross-connect is nothing more than a flexible,
programmable interconnection between input/output ports.
Besides interconnection functionality, the DXC also supports a large set of protection
and restoration mechanisms, as described in section 3.7. DXCs are well-suited building
blocks for building mesh protected or restorable networks. In contrast with ADMs, DXCs
only host STM-1 aggregate ports. To handle higher line rates, they must be combined with a
core multiplexer. Recently vendors have also developed so-called ‘multi-service nodes’ or
‘bandwidth managers’ integrating the functionalities of ADMs and DXCs. Such equipment,
terminating high-speed line rates up to STM-16 or STM-64 and providing internal
connectivity on the HO and/or LO level, can be both used in mesh and ring configuration, and
are very well suited for interconnecting rings in a flexible way.
3.4.3 Wavelength Division Multiplexing (WDM)
3.4.3.1 Optical fiber
Optical fiber has been heralded as the transport medium of the future, because of its
theoretically almost unlimited intrinsic bandwidth (about 100000 GHz). Other than that,
optical fiber has a relatively low production cost, a long lifetime, and does not suffer from
electro-magnetic interference (EMI), making it an ideally suited transmission medium in
transport (and other) networks.
The most widely deployed fiber type is standard single mode fiber (SMF) [15]. Such
fiber has three frequency regions where signal attenuation reaches a local minimum: around
800 nm, 1310 nm and 1550 nm. These are called the three optical windows of the fiber. In
practice, only the second and third window are used because they have an attenuation below
0.5 dB/km, suitable for long-haul transmission. Besides attenuation, other optical
imperfections come into play as well. A limiting factor is chromatic dispersion, which is the
result of light traveling at a different speed at different wavelengths, which results in a
broadening of the pulses that are sent over the fiber. The zero-dispersion point for SMF (i.e.
the frequency at which light does not suffer from dispersion), is located around 1300 nm,
while at 1550 nm a positive2 dispersion of 17 ps/nm.km is perceived. Chromatic dispersion is
2 Meaning light with a higher wavelength travels faster than light with a lower wavelength.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 39
a linear effect, and can be compensated, e.g. by inserting dispersion compensating fiber
(DCF) with a negative dispersion, at intermediate amplifier stages. Because of the extra fiber
length (typically 1/5th of the SMF length), extra loss is induced and the power has to be
adjusted to maintain acceptable signal to noise ratios (SNR). Therefore Bragg gratings
[16][17] (inducing a wavelength dependent delay) have often been proposed to compensate
for the dispersion, however these devices are still a premature technology. To reduce
dispersion in the 1550 nm window, dispersion shifted fiber (DSF) [18] has been developed
with the zero-dispersion point shifted to 1550 nm. Whereas DSF has proven very suitable for
single wavelength transmission, it is not compatible with multi-wavelength transmission (see
3.4.3.2). This is a result from the Kerr non-linearities, induced by the linear dependence of the
fiber's refractive index on the injected channel power. This effect depends on the area of the
single mode fiber core, also called effective area [19]. The larger the effective area, the
stronger the non-linear effect. One of the most limiting effects of the Kerr non-linearities is
four-wave mixing (FWM), creating additional frequencies (intermodulation products), if the
relative dispersion between channels is low. To overcome this, the channels must be moved
away from the zero dispersion point. This is achieved in so-called non-zero dispersion shifted
fiber (NZDF), shifting the zero-dispersion point towards a slightly lower or higher frequency.
If shifted to a lower frequency, the dispersion is slightly positive in the third window, and we
talk about NZDF+. If shifted to a higher frequency, the dispersion is slightly negative in the
third window, and we talk about NZDF-. NZDF, combined with an enlarged effective area of
the fiber, results in very effective fibers for multi-wavelength transmission, such as Corning
LEAF fiber based on ITU-T recommendation G.655 [20].
3.4.3.2 Multi-wavelength transmission
Traditionally, SMF has been used for transmission of a single TDM channel in the
1310 window (because of the low dispersion characteristics), e.g. using SONET/SDH
equipment with an optical interface transmitting/receiving at this frequency. To achieve
higher bit rates, higher TDM rates of 10 Gb/s or 40 Gb/s were envisaged. This is however not
a very scalable solution. First of all, there is a practical limit on the modulation frequency of
the transmitting lasers, however this can be overcome using optical time division multiplexing
(OTDM) schemes. Secondly, chromatic dispersion and polarization mode dispersion (PMD)
restrict the propagation distance, which decreases as the square of the channel rate. PMD
occurs due to imperfections in the fiber, which result in the polarization components of the
electrical field traveling at different speeds, causing pulse broadening. In contrast to
chromatic dispersion, PMD is much harder to compensate, because it is a much more random
effect.
The difficulties encountered with higher TDM rates on fiber, can be overcome by
using wavelength division multiplexing (WDM). WDM uses multiple channels, each
modulated with a different (TDM) signal, transmitted at distinct wavelengths over a single
optical fiber. Thereby the total capacity of the fiber can be multiplied by the number of
wavelengths used. The very first of such systems were the so-called coarse wavelength
division multiplexing systems using one wavelength at 1310 nm and one wavelength at 1550
nm, thereby doubling the fiber capacity. Next generation systems used multiple wavelengths
in the same window. Initially large spacings between adjacent wavelengths were used (e.g.
200 GHz), but later spacings of 100 GHz and 50 GHz have been achieved, realizing so-called
dense WDM (DWDM) systems. The channel spacing depends on the bit rate of the signal, the
CHAPTER 340
stability of the lasers at the multiplexer, the precision of the filters at the demultiplexer and the
gain flatness of the intermediate optical amplifiers. At present, 50 GHz spacings can be
achieved for 2.5 Gb/s signals, while 100 GHz spacing are required for 10 Gb/s systems. In the
research labs, 25 GHz and 50 GHz spacings have already been demonstrated for 2.5 Gb/s and
10 Gb/s respectively [21].
The 1550 nm window has been chosen for DWDM transmission, because it was
possible to manufacture optical amplifiers, capable of amplifying all wavelengths
simultaneously in large parts of this window (see section 3.4.3.5). These amplifiers are needed
to compensate for the power loss in the fiber (typically after a loss of 25 dB the signal is
amplified). The fact that multiple channels can be amplified at once, results in important cost
savings for point-to-point transmission, in particular in long-haul networks. When multiple
TDM channels would be transmitted on separate fibers (thus using SDM), each of the
channels would have to be amplified separately, while in the WDM case the cost of the
amplifier can be shared over all the channels on the fiber. In addition, the use of WDM offers
modular capacity upgrades, while delaying fiber exhaustion, an important consideration for
operators with a sparse or leased fiber infrastructure.
3.4.3.3 The all-optical network
As the amount of traffic passing through the nodes is ever increasing, routing
functionality at the optical layer - complementing WDM point-to-point transmission - also
provides clear benefits. All-optical networking elements allow to terminate individual
wavelength channels in a node, while transparently passing through the remaining
wavelengths, which alleviates the need for expensive opto-electronic conversions and
electrical processing equipment. Systems and subsystems that allow such all-optical
networking functions are underway from research labs to commercial availability. Meanwhile
standardization bodies are making efforts to render the optical network a unified platform
with full transport network functions, including network management and recovery schemes
[22]. Providing such recovery schemes at the optical layers provides many cost and
complexity savings [23], because it operates at coarser granularities (see section 3.7.6).
As the SDH network, also the WDM network can be described using the generic
layering principles as defined in G.805. In G.872 [24], three layers are proposed for the
optical transport network: the optical channel (OCh), optical multiplex section (OMS) and
optical transmission section (OTS) layer. The OCh layer, which is the optical counterpart of
the SDH path layer(s), provides end-to-end networking of optical channels for transparently
conveying client information of varying formats. This includes OCh connection
rearrangement for flexible network routing, OCh overhead processes for ensuring the integrity
of the optical channel adapted information and OCh optical channel supervisory functions for
enabling network level operations and management functions. The OMS layer is the optical
equivalent of the SDH MS layer and provides functionality for networking of a multi-
wavelength optical signal3. This includes OMS overhead processes for ensuring integrity of
the multi-wavelength optical multiplex section adapted information and OMS supervisory
functions for enabling section level operations and management functions, such as multiplex
section survivability. Finally, the OTS layer, similar to the SDH RS layer, provides
functionality for transmission of optical signals on various types of optical fiber.
3 Note that a "multi-wavelength" signal includes the case of just one optical channel.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 41
Despite all the technological innovations of the past few years, the evolution to an
all-optical network will not happen overnight. The initial vision of all-optical networking was
that once an optical signal has been generated at a certain wavelength, the signal remains in
the optical domain until its final destination is reached. In such a so-called transparent optical
network, the signal may be amplified or translated to a different wavelength underway, but at
no point the signal is converted to the electrical domain and then reconverted to an optical
signal. The latter is called optical-to-electrical-to-optical (OEO) conversion. In contrast,
optical networks in which OEO conversion takes place at every intermediate network element
are called opaque optical networks [25]. An opaque optical cross-connect is depicted in
Figure 11.
.
.
.
OEO
OEO
OEO
OEO
OEO
OEO
OEO
OEO
OEO
OEO
OEO
OEO
OEO
OEO
OEO
OEO
OEO
OEO
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Figure 11: Opaque optical cross-connect
The major advantage of transparent optical networks over opaque optical networks is
the cost savings induced by eliminating the OEO converters (also called transponders), which
can take up to 65% of the total optical network cost over a 10 year time period, according to
some studies [26]. Other advantages of transparent networks are the facts that they are
independent of the signal rate (to a certain extent) and somewhat more reliable because they
contain fewer network elements. On the counter-side, transparent optical networks suffer from
signal impairments, induced by the optical fiber and network elements, making such networks
less scalable. In order to ensure proper signal quality, the network has to be completely
engineered before building it. For long-haul networks, it might not even be feasible to build
such a network, because after a certain distance, the signal has to be regenerated anyway to
guarantee error-free transmission. As all-optical 3R (reamplification, reshaping and retiming)
regeneration is not possible yet, an opaque network, using electronic 3R regeneration, seems
to be a pragmatic near term solution. In such an opaque network, the network can be
engineered link-by-link. Transponders can also be used in a multi-vendor environment to
interconnect WDM gear from different vendors, which use different wavelength sets in their
equipment. In addition the opaque optical network also overcomes the monitoring difficulties
at the optical layer, because at the regeneration points, the client layer signal can be non-
intrusively monitored. E.g., the B1 and J0 bytes in the SDH regenerator section overhead can
be monitored to assess the digital performance and to verify the connectivity respectively.
CHAPTER 342
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Figure 12: Transparent WP optical cross-connect
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Figure 13: Transparent VWP optical cross-connect
Another advantage of opaque optical networks is that if facilitates the wavelength
assignment problem. Through the use of transponders, a signal entering a node at a certain
wavelength can leave the node at any other wavelength. Such networks are much easier to
design than networks in which a signal must maintain a unique wavelength on the path
between its endpoints, because in this case it has to be assured that no wavelength conflicts
between different paths occur. This is the case in transparent networks which do not have
wavelength conversion facilities in the nodes of the network, also denoted as wavelength path
(WP) networks [3][27] (see Figure 12 for a WP OXC). Recently, components have become
available that allow all-optical wavelength conversion of a signal [28]. This enables to
provide wavelength conversion facilities in the nodes of a transparent network, thereby
realizing so-called virtual wavelength path (VWP) networks (see Figure 13 for a VWP OXC
with wavelength converters at the output). In a VWP network, a signal is no longer
characterized by a unique wavelength, which gives additional freedom in the routing of
signals with the potential of increasing the network throughput. However, in transparent
optical networks this freedom comes at the cost of extra wavelength converters, while in
opaque networks it is for free because transponders are used anyhow. In addition, current
techniques for optical wavelength conversion (e.g. based on semiconductor optical amplifiers
[28]) deteriorate the signal too much, such that regeneration is needed anyhow. Besides the
cost of the wavelength conversion facilities in a VWP network, also the cost of the switching
facilities in the node have to be taken into account. The complexity of the node depends on
the use of wavelength converters: while the node of a WP network only requires connectivity
between corresponding wavelengths, the node of a VWP or opaque network requires full
internal connectivity between all possible wavelengths, which results in a more complex
switch architecture and an addition cost. Therefore, also intermediate forms, featuring only a
limited amount of wavelength converters have been engineered [29]. A more elaborated
discussion about node architectures and wavelength conversion in optical networks is held in
[30].
In conclusion, while the initial vision of the all-optical network still remains, it is
hampered by optical impairments, which can currently only be overcome via the electrical
domain, and the lack of suitable management and monitoring capabilities the optical layer.
Therefore the opaque optical network offers a pragmatic near-term solution. As some of the
above issues become resolved, expanding islands of transparency will be witnessed, with
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 43
transponders at the boundaries [31]. Finally, once optical regenerators, wavelength converters
and optical management functionalities (see next section) become widely available, the all-
optical network will become a reality.
3.4.3.4 Management of the optical network
One of the major contributors to the success-story of SDH, was its powerful
operations, administration and maintenance (OAM) framework supported by the overhead
bytes. Currently, the OAM capabilities of WDM are limited because until recently the only
WDM gear that was deployed were simple point-to-point systems. Such systems do not need
advanced OAM functionalities and the basic management tasks for such systems (e.g.
performance management) can be done using some vendor-specific mechanism. However,
with the advent of more sophisticated optical network elements in an all-optical network,
more elaborated OAM tasks need to be performed. In addition, when these OAM tasks need
to be performed on a large scale and across multiple optical subnetwork boundaries,
proprietary techniques have to be replaced with standardized solutions.
The key requirements for optical layer management are configuration, fault and
performance management. Configuration management allows setting up and tearing down
optical channel connections in a dynamic fashion, in order to resolve current manual
configuration headaches. Performance management measures the signal quality and inspects
whether the guaranteed level of performance is maintained. Fault management detects,
isolates and initiates recovery actions in order to restore the affected traffic.
The implementation of an effective optical network management system, comparable
to the one used in SDH, depends on the adoption of the Telecommunications Management
Network (TMN) framework [5] as well as on the implementation of overhead capacity in
order to exchange the necessary management information.
Today’s TMN implementation for WDM networks is functionally restricted to the
Element Management Layer (EML). The point-to-point WDM systems are managed as
individual network elements (NE) from a central Network Operations Center (NOC). To cope
with a catastrophic outage of the NOC, a redundant NMS is usually installed in a remote
standby operations center. With the advent of new optical network elements, such as
dynamically re-configurable OADMs and OXCs, a more sophisticated NMS is required to
operate and manage future WDM networks. As such, it is mandatory that the WDM NMS
supports the Network Management Layer (NML) as well. The NML manages the WDM
network as a whole while the EML manages the WDM network elements individually.
Moreover, it is required to implement trouble ticketing systems, service availability
monitoring and reporting systems as well as other Operational Support Systems (OSS). These
OSS form the Service Management Layer (SML) of the WDM NMS.
As stipulated in [24], end-to-end management of optical channels involves continuity
and connectivity supervision, maintenance indication, signal quality supervision, adaptation
management and protection control in the Optical Transmission Section (OTS), Optical
Multiplex Section (OMS) and Optical Channel (OCh) layers of the Optical Transport
Network (OTN). To support these functions, the necessary amount of overhead needs to be
provisioned in the OTN. Several possibilities exist for implementing this overhead [32],
including [33][34][35][36][37][38]:
• Use of an optical supervisory channel (OSC) outside the wavelength grid used by the
WDM system (out-of-band solution).
CHAPTER 344
• Use of one of the wavelengths of the WDM system as OSC (in-band).
• Subcarrier modulation (SCM) of an optical channel (pilot tone).
• Using part of the unused client layer overhead (e.g. SDH RSOH bytes) in opaque
networks (rate preserving).
• Using a digital wrapper around the OCh client (non rate preserving).
Several of these options can potentially coexist within the same control architecture.
For the OCh overhead, using an OSC (either in-band or out-of-band) does not seem to be an
appropriate solution. As the OCh overhead is used for the OCh connectivity supervision
(using an OCh trace), the OCh overhead must be impossible to divide from the OCh payload,
which is not the case if an OSC is used. The other overhead implementations do not suffer
from this disadvantage. The pilot tone, which is a subcarrier modulated signal added at the
transmitter, however suffers from another potential drawback. The pilot tone introduces
transmission penalties and might cause cross-talk between different OCh signals (especially if
the same frequency is used for all pilot tones). These physical impairments increase as the
amount of overhead information increases (which is a limiting factor for the amount of
overhead) and as number of channels increases (which is not desirable in view of future
upgrades of the network) and thus requires a very careful design of the system. This is not the
case when a TDM frame is used to carry the overhead. Using part of the client layer overhead
might be a temporary solution, but is not preferred because the amount of free overhead bytes
(e.g. in SDH) is too limited and some of these bytes are already used by certain vendors for
proprietary implementations. In addition, these bytes are only accessible at regenerator
locations or at the edges of the network. The latter is not so much of a restriction in current
WDM networks. As already mentioned the initial vision of an all-optical network has been
temporarily abandoned and has been replaced by the expectation of an opaque optical network
carrying a wide range of digital clients. In this light, the adoption of a digital (or TDM)
wrapper around the OCh client to add a new set of overhead bytes seems an attractive solution
[38]. This scheme takes advantage of the need for regeneration points, to add additional
capacity to the OCh client. As such, enough overhead can be provisioned to be future proof,
without distorting the signal. Digitally added overhead also solves the problem of
performance monitoring. In an all-optical network it is difficult to estimate the BER of the
client accurately based on indirect measurements of parameters such as optical power and
signal to noise ratio (SNR). This is no longer the case when a digital wrapper is used around
the OCh client. In addition, it is even possible to add forward error correction (FEC) to the
signal, which leaves an additional margin to increase the bit-rate on the channels or to
increase the distance spanned before regeneration. As such the digital wrapper overhead does
not degrade the signal as the pilot tone does, but it even improves it through the use of FEC.
For the OMS and OTS overhead (e.g. for OMS protection signalling), the OSC can
be used as proposed by ITU. Part of the OSC capacity can then also be used to transport OCh
overhead as proposed in [33]. While we have already stated that some OCh overhead (such as
signal trace) is best transported using OCh associated overhead (e.g. digital wrapper), such
channel associated overhead is not accessible at every possible point in the optical network.
E.g., the digital wrapper is accessible at any point where regeneration is performed, however
such regenerators might not be available at every OMS termination (e.g. at optical amplifiers).
There are however OMS terminations that perform an OCh maintenance function. To carry
OCh overhead information to such locations where the OCh channel associated overhead is
not available, the OSC overhead can be used.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 45
Recently, some OAM mechanisms stemming from the data world have been entering
the WDM world. For example, Ciena proposes an Optical Signalling and Routing Protocol
(OSRP) of a distributed nature that reminds of former ATM implementations. This
mechanism enables automated topology discovery, distributed connection management
(connection establishment, tear-down, modification and re-routing) and traffic engineering
using different QoS parameters [39]. Recently, some proposals have also been made to use
IP-based protocols, such as the open shortest path first (OSPF) routing protocol, resource
reservation protocol (RSVP) and multi-protocol label switching (MPLS) signaling protocol to
handle optical signals, which facilitates scalability and interworking with IP equipment
[40][41][42][43][44]. The principle of applying MPLS for managing data-centric optical
networks has attracted attention in the IETF, where the idea of multi-protocol lamba
switching has been proposed. This solution seems very attractive for IP-over-WDM networks,
because it merges the management systems of both technologies. On the other hand, the
adoption of such an approach is questionable for more traditional WDM networks supporting
a wide range of client signals.
3.4.3.5 WDM equipment
After a more generic description of WDM technology, we now take a closer look at
the products that are currently on the market, or are expected to be released soon. We split the
equipment classification in line systems, optical add-drop multiplexers (OADMs) and optical
cross-connects (OXCs).
A. Line Systems
WDM line systems consist of terminal multiplexers, demultiplexers and intermediate
optical amplifiers. Two types of WDM line systems are currently deployed: opaque (or open)
and transparent (or closed) systems. Open systems use transponders at the input and allow
equipment from any vendor at virtually any wavelength to be connected to the WDM system,
which is not the case for closed systems, where each channel can only interface to equipment
at a well defined wavelength. The major differences between transparent and opaque
networks have already been discussed in section 3.4.3.3.
Transponders can be fixed or flexible with respect to both wavelength and bit rate.
Whereas fixed wavelength transponders are cheaper, wavelength tuneable transponders can
substantially relieve the equipment sparing requirements. Indeed, instead of holding a fixed
wavelength transponder in spare for every working fixed wavelength transponder in the
system, we can use a single tuneable wavelength transponder that can be used as a back-up for
any fixed wavelength transponder. Transponders suited for all bit rates, of which the
particular mode has to set through software, are called broadband transponders. The use of
broadband transponders allows for a more flexible and reconfigureable network in case a wide
variety of digital client signals is used. In addition, such transponders relieve the sparing
requirements. On the other hand, broadband transponders come at a higher cost compared to
bit rate specific transponders.
The multiplexers aggregate the optical signals of a series of lasers transmitting at a
particular wavelength, in a single multi-wavelength signal using a combiner element [16]. The
lasers have to comply to certain requirements regarding frequency stability and narrow laser
linewidth (dependent on the bit rate) to allow precise alignment with standardized
wavelengths. The wavelength grid currently standardized in ITU uses 100 GHz and 50 GHz
CHAPTER 34
6
spacings [45]. The used laser is either coming from the transponder (in case of an open
system) or separate 'colored' lasers (at a compliant wavelength) have to be used at the input
(closed system). These colored lasers can either have a fixed wavelength or can be tuneable
over a wide wavelength range. As for transponders, the use of tuneable lasers can
considerably reduce the cost and inventory requirements for spare lasers.
The demultiplexers are based on filters, extracting the individual wavelengths from
the WDM comb, and photodetectors receiving each wavelength. The filters can either be fixed
or tuneable. Tuneable filters offer more flexibility, but also more challenges (e.g. tuning speed
and wavelength range of the filters). Today, fixed filters based on Fiber Bragg gratings and
arrayed waveguide gratings are mostly used. Other possible implementations for fixed and
tuneable filters can be found in [16]. Demultiplexing can either be done in one single step,
filtering out each individual wavelength at once, or in multiple parallel steps, first filtering out
a set of wavelengths and then filtering out wavelengths from this set. The second approach
allows for less precise (and thus cheaper) filters and a more modular design of the
demultiplexer [46].
Optical amplifiers provide gain to amplify optical signals suffering from propagation
losses in the fiber. Different kinds of amplifiers exist [47]. The first such amplifiers were the
Erbium-doped fiber amplifiers (EDFA), using a section of optical fiber, doped with Erbium
ions and stimulated by a pump laser, in order to amplify optical signals in a broad frequency
range with a flat gain [48]. Both 980 nm and 1410 nm pump lasers can be used [49]. Whereas
1410 nm pump lasers are more mature and offer more efficient pumping and can significantly
boost signal power, the 980 nm counterparts introduce less noise, which is of particular
interest in long-reach, high channel count systems. The usable frequency range of the EDFA
is between 1525 to 1560 nm (the so-called conventional or C-band), which is only about 30%
of the total bandwidth available in the third window. When changing the doping of the fiber,
other frequency ranges can be reached at well, e.g. the short or S-band (below 1525 nm).
Raman-scattering amplifiers have been devised to cover a wide frequency range, including the
S-band [50]. For the so-called long or L-band (between 1570 and 1610 nm), the silica erbium
dual-band fiber amplifiers (DBFA) have been demonstrated successfully [51] (also Raman-
scattering amplifiers can be used). As such the third window, is split up in a number of sub-
bands using different types of amplifiers (eventually integrated in one product). This allows
for a modular system, where sub-bands can be added as bandwidth requirements evolve.
Optical amplifiers can be deployed in three configurations: pre-amplifiers, in-line
amplifiers and boosters. The pre-amplifier is used before the receiver to enhance the signal-to-
noise ratio, and requires a high gain and a low noise figure. The booster is used after the
transmitting laser to boost the power to a level not available from the laser. Boosters are
operated in saturated mode providing high optical power, while the gain is modest. In-line
amplifiers compensate for the signal power degradation due to propagation losses in the fiber
and are deployed between transmitter and receiver. Current in-line amplifiers are capable of
compensating losses up to 30 dB.
Recent advances in both multiplexers, demultiplexers, amplifiers and dispersion
compensation techniques have increased the number of wavelengths on commercial systems
up to 128 using both the C-band and L-band.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 4
7
B. Optical Add-Drop Multiplexers (OADM)
OADMs terminate two fiber pairs and allow to extract and insert one or more
wavelength channels from the WDM comb, while (transparently) passing through the
remaining wavelengths. Current OADMs are in fact optical amplifiers, in which an inter-stage
loss element can be introduced, such that they can be extended with a module able to add/drop
wavelengths. In current OADM implementation this module allows to add-drop only a fixed
sub-set of all available wavelengths. Such fixed add/drop modules can be realized by
circulators and fiber Bragg gratings [52]. Multiple of such elements in series allow to
add/drop multiple wavelengths. However the amount of elements to be placed in serial is
limited due to the optical impairments induced. In addition, it is difficult to make in-service
upgrades to such a configuration. Another possible implementation of fixed OADMs is by
using complete demultiplexing (i.e. parallel implementation). As such each wavelength can be
either added/dropped or passed through. Whereas complete demultiplexing is more expensive,
it allows 100% add/drop of all channels and can easily be upgraded in service. This
implementation also enables flexible add/drop multiplexing, when optical switches (see
section C) are used to configure each wavelength either in the pass-through state or in the
add/drop state [53]. Other possible implementations [54] of flexible add/drop multiplexing,
not using complete demultiplexing, rely on the use of tuneable fiber Bragg gratings, silica
planar lightwave circuits, Mach-Zender chains or acousto-optical switches.
C. Optical Cross-Connects (OXC)
Whereas OADMs terminate two fiber pairs, OXCs enable interconnectivity of
wavelengths between multiple fiber pairs, in addition to local add/drop. As most OXCs are
envisaged to be non-blocking, they rely on complete optical demultiplexing of the WDM
signal. A large matrix of switches then enables to interconnect wavelengths between the
different fibers. Opaque OXCs either have an optical cross-connect matrix with transponders
at the boundaries or an entirely electrical cross-connect matrix [55]. For transparent optical
cross-connects, a multitude of internal architecture exist, depending on whether wavelength
conversion is enabled or not (or to a limited extent). As already mentioned in section 3.4.3.3,
non-blocking opaque OXCs and transparent VWP OXCs require interconnectivity between all
possible wavelengths of all terminating fibers, and thus require a much larger space switch
than transparent WP OXCs.
Also for the optical switch fabric used inside the cross-connect, many different
implementation options can be considered [56]. The size of the OXC that can be constructed
with a good yield is an important metric for choosing the switch technology. In particular,
some OXC technologies, require a large number of switching stages, because they are based
on small building blocks. The more switching stages, the higher the total insertion loss, thus
the more expensive the transmitters/receivers and amplifiers need to be to compensate for the
losses. Most switch fabric architectures require multiple interconnection stages. The switches
used inside the switch fabric can be characterized according to the medium through which the
light travels − free-space or waveguide − and the mechanism used to activate the switch −
opto-mechanical, thermo-optic or electro-optic. The switching speed decreases from electro-
optic (sub-microsecond) to thermo-optic (5 ms) and then opto-mechanical (hundreds of ms)
switches. Similarly, the insertion loss increases from free-space to waveguide technology.
Thus the most interesting switches are free-space electro-optical switches such as liquid
crystal switches [57]. Another promising technique uses micro electro-mechanical switches
CHAPTER 348
(MEMS), which are essentially tiny free-space tilting mirrors, which can be rotated to steer
the optical input channel to the required output port [58]. Using MEMS, single stage optical
switches with very low insertion loss (3 dB) can be manufactured, which are ideally suited for
large port count OXCs.
3.5 Packet-switched network technologies
In this paragraph we briefly describe packet-based network technologies. As
opposed to SDH and WDM transport networks, packets-switched networks do not use fixed
connections, but use small packets for transporting traffic. Each packet carries part of the
information to be exchanged and travels autonomously through the network using the
available resources. Because of the rigid multiplexing structure of fixed dedicated
connections, SDH and WDM technologies on their own are not very well suited for
transporting traffic with a wide range of bitrates, different statistical natures, and varying
quality of service requirements. In contrast, packet-switched networks better fit the needs for
such traffic, because the amount of packets to be transmitted can be easily and rapidly scaled.
In addition, no resources are reserved for a packet, such that different packets can effectively
share the same resources. We give a short overview of two packet-based technologies: ATM
and IP, without going in detail, because the research performed in this thesis does not
consider packet-switched networks. Nevertheless, we feel it is important to mention these
technologies, because in future transport networks the integration of packet-switching is
becoming more and more important (see section 3.6).
3.5.1 Asynchronous Transfer Mode (ATM)
The Asynchronous Transfer Mode (ATM) is a packet based transport protocol, in
which information for multiple service types, such as voice, video, or data, is conveyed in
short (53 bytes) fixed size packets or cells [59]. Each packet contains a 5-byte header
(containing cell identification and destination information) and a 48-byte payload field. In the
intermediate switches, packets are stored and forwarded based on their header information.
ATM acts both as a transport network technology (since it allows guaranteed capacity and
constant transmission delay) and as a packet-switched technology (flexibility and efficiency
for intermittent traffic). It provides scalable bandwidth from a few megabits per second
(Mb/s) to many gigabits per second (Gb/s). As SDH, ATM also uses time division
multiplexing, but in a asynchronous manner: information is not transferred in fixed time slots,
but in packets which can be sent at varying frequencies depending on the current rate at which
data needs to be sent.
The two important layers in the ATM transport network are the ATM layer and the
physical layer. The physical layer can either be a physical transmission medium (cell based
interface) [60][61] or it can be an STM frame (SDH based interface) [10]. The ATM layer
itself is divided in two levels: the virtual channel (VC) and virtual path (VP) level. On the VC
layer, ATM services are offered on virtual channels. The VP layer transports virtual paths,
which are bundles of virtual channels that can be switched at once. The header of each cell
contains a virtual channel identifier (VCI) and a virtual path identifier (VPI), defining the
correct VP and VC. In each node of the ATM network, the cell is forwarded to the correct
outgoing link based on the lookup of the VCI and VPI value in the routing table. All VCIs
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 49
and VPIs, have only local significance across a particular link and are remapped, as
appropriate, at each switch to allow reuse of VPI and VCI values in the network. Depending
on whether the VPI or both the VPI and VCI are translated at an intermediate ATM node, this
node is denoted as a VP switch or a VP/VC switch.
Two types of connection-oriented ATM services exist4: permanent virtual circuits
(PVC) and switched virtual circuits (SVC). A PVC allows direct connectivity between sites
(similar to a leased line), whereas SVC is created and released dynamically and remains in
use only as long as data is being transferred (similar to a phone call). VP connections are
responsible for bulk transport of VC links and are of a more permanent nature (i.e. they are
not constantly established and removed). Strategic ATM transport network planning is thus
more involved with the VP allocation and dimensioning and not so much with the VC
establishment, which is more an operational planning issue.
ATM supports QoS guarantees based on a traffic contract established with the user.
A traffic contract specifies the permitted values for peak bandwidth, average sustained
bandwidth, and burst size, among others. Based on this traffic contract, a connection can be
permitted or refused to enter the network depending on the amount of resources available (this
is the process of connection admission control). Once the connection is admitted to the
network it has to be ensured that the traffic contract is obeyed. This can be accomplished
through traffic shaping and policing. Traffic shaping is the use of queues to constrain data
bursts, limit peak data rate, and smooth jitters so that traffic will fit to the traffic contract.
ATM devices are responsible for adhering to the contract by means of traffic shaping. ATM
switches can use traffic policing to enforce the contract, by marking cells eligible to be
dropped in case network congestion occurs.
One of the key questions, receiving a lot of attention, concerns the future use of
ATM. Initially ATM was engineered as an end-to-end multi-service platform with QoS
capabilities. However, ATM never made it to the desktop because of the high cost of the
network interface cards (NIC) compared to its Ethernet counterparts. While initially ATM
was seen as a replacement of FDDI rings for some applications in the LAN backbone, Fast
and Gigabit Ethernet and layer 3 switches have been developed to provide a cheaper and less
complex solution. The only place where ATM has been extensively adopted to date is in the
transport network, in order to provide VP connectivity with flexible bitrates and adequate
QoS. However, in recent years QoS mechanisms have also entered the IP world [62] with less
overhead than ATM (because they don't rely on small fixed sized cells with 10% overhead).
Also the argument that ATM could provide faster packet forwarding than IP routers, because
of its fixed length cells has been outdated. IP terabit routers that operate at wirespeed on 2.5
or 10 Gb/s links are emerging [63], while ATM-like mechanisms such as MPLS [64] provide
a useful complement to make more efficient routing decisions. As such, ATM is threatened in
its further existence, although some of its principles will still be used in other technologies
(e.g. MPLS).
3.5.2 Internet Protocol (IP) networks
The Internet Protocol (IP) is a network-layer (layer 3 of the OSI protocol stack [65]) protocol
that contains addressing information and some control information that enables packets to be
routed. IP transports data in variable length packets (with a maximum length of 64 Kbytes)
4 ATM also supports a connection-less service, which is less used however.
CHAPTER 350
with a fixed length header (20 bytes). This header contains amongst others the source and
destination address. IP uses a hierarchical addressing scheme, in which an address consists of
a network number and a host number, making it very scalable. The packets are routed through
the network by so-called routers. Routing involves two basic activities: determining optimal
routing paths and transporting packets through the network. The latter of these is referred to
as switching (or forwarding). Although switching is relatively straightforward, path
determination can be very complex. To aid the process of path determination, routing
algorithms initialize and maintain routing tables in the routers, which contain route
information. IP packets can then travel through the network one router at a time, without the
entire route being known at the onset of the journey. Instead, at each stop, the packet is
temporarily stored, and the next destination is calculated by matching the destination address
of the packet with an entry in the current router's routing table. Afterwards the packet is
forwarded to the appropriate outgoing link (or dropped in case the router is saturated). Routers
within the Internet are organized hierarchically. Routers used for information exchange within
autonomous systems are called interior routers, which use a variety of Interior Gateway
Protocols (IGPs) to accomplish this purpose. These IGPs include Open Shortest Path First
(OSPF), Intermediate System to Intermediate System (IS-IS), and Routing Information
Protocol (RIP). Routers that move information between autonomous systems are called
exterior routers. These routers use an Exterior Gateway Protocol (EGP) to exchange
information between autonomous systems. The Border Gateway Protocol (BGP) is an
example of such an exterior gateway protocol.
In complement to IP routing, the Transmission Control Protocol (TCP) provides
reliable connection-oriented transmission of data in an IP environment. TCP corresponds to
the transport layer (layer 4) of the OSI reference model. Among the services TCP provides are
stream data transfer, reliability, efficient flow control, full-duplex operation, and multiplexing.
As such the TCP is suitable for efficiently supporting e-mail, FTP, and other error-critical
traffic. Another layer 4 protocol belonging to the IP family is the User Datagram Protocol
(UDP), which is a connectionless transport-layer protocol. Unlike the TCP, UDP adds no
reliability, flow-control, or error-recovery functions to IP. Because of UDP's simplicity, UDP
headers contain fewer bytes and consume less network overhead than TCP. Therefore UDP is
suitable for streaming applications for which error-free transmission is less critical (e.g. voice,
video, …).
Currently two versions of IP exist. While version 4 [66] is commonly used today, IP
version 6 [67], provides additional address space and a streamlined header to improve routing
efficiency.
3.6 Integration of packet-switching in transport network
technologies
At the time of writing a heated debate is going on about the structure of future
transport networks. All discussants agree that at the top layer of the network, IP-based
applications will dominate for at least the next decade. On the other hand it is also accepted
that optical fiber will remain the preferred transmission medium and WDM will be used as
transport technology at the lowest layer of the network. However, for the technologies to be
used in between IP and WDM, a plethora of possibilities exist [68] as shown in Figure 14.
Efficient IP over WDM transport requires an implementation with low overhead, QoS support
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 51
and service differentiation, high availability (enabled by built-in recovery mechanisms),
scalability, flexible configuration possibilities and low cost. We will now describe some
options for IP over WDM transport and discuss how the options fulfil the requirements
mentioned above.
Cell-based
interface
WDM
IP
ATM
SDH
PPP
HDLC
SDL GigE Optical
packets
Figure 14: IP-over-WDM architectures
3.6.1 IP over ATM over SDH over WDM
In this implementation IP packets are segmented in ATM cells (e.g. classical IP over
ATM [69]), via adaptation layer 5 (AAL) of ATM [59]. These ATM cells can then be packed
in SDH VC-4s or concatenated VC-4 payload [10]. Finally the SDH high bandwidth optical
interface (e.g. at 2.5 Gb/s [14]) can be connected to the WDM system, either using a colored
laser at the SDH equipment (transmitting at a wavelength of the WDM system) or using an
intermediate transponder (converting the optical signal to a WDM compliant wavelength).
Because 2 intermediate layers are used between IP and WDM, a lot of overhead is added to
the useful data (e.g. for a 350 byte long IP packet, 22% of overhead is added [68]). However,
this implementation has the advantage that it relies on proven technologies, which offer added
value to the IP-over-WDM transport and can be implemented today. ATM can offer the QoS
support and service differentiation, which IP currently lacks. SDH offers fast protection
switching, as opposed to slow reconfiguration in the IP layer and immature recovery actions
in the WDM layer. In addition, the performance monitoring capabilities of SDH can be used.
3.6.2 IP over ATM over WDM
As in the previous scenario, IP packets are mapped into ATM cells. Afterwards,
ATM cells are not mapped in SDH frames, but are directly transmitted over the physical
medium, using a cell-based physical layer specific for the ATM protocol. In ITU-T [61], a
155 Mb/s and 622 Mb/s cell-based interface has been standardized, while the ATM forum
also has a specification for 622 Mb/s and 2488 Mb/s [60]. As opposed to SDH, the use of
such a cell-based interface, provides a simpler and cheaper alternative with low physical
overhead (around 16 times less than SDH). The only overhead, besides the ATM headers, is
CHAPTER 352
created by the insertion of OAM cells (1 OAM cell is inserted every 431 contiguous data
cells), which allow performance monitoring by checking the BIP-8 value calculated over the
payload of the data cells. While the overhead of the OAM cells is low, the use of ATM itself
still adds a lot of overhead to the useful date (e.g. for a 350 byte long IP packet, 19% of
overhead is added [68]). When eliminating SDH, fast recovery can still be achieved, either in
the ATM layer, or in the future also in the WDM layer.
3.6.3 IP over SDH over WDM
In this scenario encapsulated IP packets are mapped in the SDH frame. The
encapsulation is performed using PPP5 [70], after which HDLC6 framing [71] is performed.
HDLC framing adds delimited flag sequences at the start and end of the frame, and also
includes a cyclic redundancy check (CRC) checksum field for error control. The HDLC
frames can then be transported in SDH VC-4s or concatenated VC-4 payload. It is proposed
[72], to first scramble the data to minimize the risk that a malicious user sends data that
causes SDH to loose synchronization. This implementation of IP over SDH over WDM is
also referred to as packet over SONET/SDH (POS) [73]. In this scenario, the fast protection
switching and performance monitoring of SDH are retained. By eliminating ATM, the
amount of overhead is reduced (e.g. for a 350 byte long IP packet, only 6% of overhead is
retained [68]), but on the counter-side, also the QoS that ATM provides is lost. The latter can
be overcome by two solutions. A first possibility is by ‘over-provisioning’, which provides
abundant bandwidth and IP routing capacity such that there is no contention for resources
between different traffic streams. A second possibility is to use QoS mechanisms in the IP
layer, which are now starting to emerge [62]. The integrated service model (IntServ) [74]
defines a set of services with different QoS characteristics for which the required resources
can be reserved using a resource reservation protocol (RSVP) [75]. As such the required QoS
can be offered per flow. However the large set of services and flows and the temporary
reservations (i.e. soft-state) make this a poorly scalable solution. Therefore, the differentiated
services (DiffServ) model has been devised, which allows differentiation of a small set of
services [76]. As such DiffServ offers QoS to traffic aggregates (and not per flow) by defining
a set of per-hop forwarding behaviours (including queuing, shaping and policing) in the
routers, which define the classes of QoS offered. In addition, the multi-protocol label
switching (MPLS) protocol [64], allows to forward packets at layer 2, based solely on an
additional label they carry (also called shim header). In this way, packets that belong to
different applications can be switched in bulk to satisfy particular QoS requirements. The
MPLS layer also offers the potential to implement fast protection switching mechanisms [77].
3.6.4 IP over SDL over WDM
The Simple Data Link (SDL) protocol [78], which has been proposed by Lucent,
replaces HDLC framing for PPP encapsulation by an SDL frame. The SDL frame uses no
delimiting flag sequences but is started with a packet length field instead, which allows for
easier synchronization (especially at high bit rates). The SDL frame can be directly encoded
5 The Point-to-Point Protocol (PPP), provides a standard method for transporting multi-protocol datagrams (e.g.
IP) over point-to-point links.
6 High-level Data Link Control
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 53
on an optical carrier. As such this is a protocol with very low overhead (e.g. for a 350 byte
long IP packet, only 3% of overhead is retained [68]). The use of optional error checking
codes (CRC) can allow bit error rate (BER) monitoring. The SDL frame does not include any
bytes dedicated for protection switching protocols. Thus this scheme relies on recovery
actions either in the higher layer (either through routing table updates in the IP layer, or
protection switching in MPLS) or in the WDM layer. As for IP over SDH over WDM, QoS
mechanisms are needed in the IP layer.
3.6.5 IP over Gigabit Ethernet over WDM
Ethernet accounts for 85% of the world-wide local area network (LAN) traffic. Using
Gigabit Ethernet, high-capacity LAN switches can be interconnected with each other [79][80].
When Gigabit Ethernet operates in full-duplex mode, it is simply an encapsulation and
framing method for IP packets. In this case, the CSMA-CD protocol (to avoid collisions) is
not used, such that the distance limitations imposed by this protocol do no longer come into
play, and interconnection over large distances is possible. The cost of Gigabit Ethernet cards
is sufficiently lower than SDH cards. On the other hand, Gigabit Ethernet adds a fair amount
of overhead, because 8B/10B encoding is used for transmission on an optical carrier. This
means that every octet is encoded using 10 bits to ensure sufficient transitions in the signal for
clock recovery (e.g. during idle periods). Gigabit Ethernet also provides support for different
classes of service (CoS), by providing a means of ‘tagging’ packets with an indication of the
priority or class of service desired for the packet. Again, recovery actions can either be
implemented at the IP or WDM layer.
3.6.6 Optical packet switching
In the near term, all switching actions will still take place in the electrical domain.
Meanwhile, the WDM layer can be used to interconnect electronic switches with
wavelengths, thereby establishing a logical topology, which differs from the underlying fiber
topology. This logical topology can be adapted in response to changes in the traffic
distribution and can be used to bypass intermediate routers for large aggregate flows. This can
be achieved in a dynamic fashion with an optical flow switching protocol similar to MPLS.
However this implementation still uses electrical switching and the optical processing is
limited to bypassing an entire wavelength at a node. More futuristic architectures promise to
deliver optical packet switching. In such an architecture the switches optically route packets
based on optical packet header information. Such networks pose considerable challenges,
including the generation of ultra-short pulses, synchronization, high-speed optical switching,
optical buffering (i.e. optical memory) and packet header processing [81]. An intermediate
solution, that overcomes some of these issues, can be realized through WDM burst switching
[82]. WDM burst switching combines the complementary strengths of optical and electronical
technologies. In such a network, each link contains a separate control channel (or more than
one), independent from the WDM data channels, that carries burst header cells (BHC)
describing the bursts carried on the WDM data channels. As such, data and control
information is separated, which allows the control information to be processed digitally, while
the data can remain in optical format. Based on the control information (including a
destination address), the control system of the burst switch determines a free channel on the
CHAPTER 354
outgoing link, and converts the burst to the appropriate wavelength. As such, no optical
buffering is required.
3.6.7 Proprietary mechanisms
Recently, some vendors have come up with their own solutions for efficiently
transporting IP over WDM networks. These proprietary mechanisms focus on an optimized
architecture for IP-centric networks, by reducing the total amount of networking layers (i.e.
eliminating intermediate ATM and/or SDH layers) while still providing QoS and recovery
mechanisms, using some other intermediate protocol in between IP and WDM, at a much
lower cost.
A first example of such proprietary solutions is the Dynamic Packet Transport (DPT)
protocol proposed by Cisco [83]. DPT works in a ring topology, in which IP routers hosting
DPT line cards can be interconnected. These DPT cards support the Spatial Reuse Protocol
(SRP) [84], which is a medium access control (MAC) protocol, ensuring an optimized use of
the shared ring bandwidth. The ring bandwidth can be shared because SRP uses destination
stripping: the packet is removed from the ring at the destination node, such that this
bandwidth can be reused by another packet on the remaining segments of the ring (which is
particularly interesting in case most of the traffic is between adjacent nodes on the ring).
Furthermore, the MAC protocol enables that only packets destined at an IP router terminate in
that router, while other packets are passed through at the MAC layer. This can be done
without setting up dedicated communications channels (as ATM VPs) between two nodes on
the ring, such that the entire ring bandwidth is available to all packets between any two nodes,
and statistical multiplexing can be fully exploited. The SRP protocol includes a fairness
algorithm, which makes sure that at each node a fair share of the ring bandwidth is devoted to
both introduced and forwarded packets. This fairness algorithm also takes into account the
priority of the packets. Currently, two priorities are supported: high and low, which make it
possible to achieve some sort of service differentiation, but not real QoS. The ring also
supports a protection switching protocol called Intelligent Protection Switching (IPS)
optimized for IP traffic. In contrast to SDH based protection mechanisms (see section 3.7),
this mechanism does not devote any protection bandwidth in advance. At the moment a
failure occurs, only half of the ring bandwidth is available, such that less traffic can be
transported (e.g. some of the low priority traffic might be dropped).
A second example of a proprietary mechanism is Dynamic Transfer Mode (DTM),
pioneered at the Royal Institute of Technology (Sweden), and later commercialized by Dynarc
and NetInsight [85][86][87]. DTM is essentially a Time Division Multiplexing (TDM)
scheme, like SDH but of a more dynamic nature and without the rigid multiplexing structure
of SDH, suitable for packet based transport. DTM uses a 125 microseconds TDM frame
(compatible with SDH), divided in 64 bits timeslots. Thus each timeslot represents a 512 Kb/s
channel. As such a dedicated circuit with the capacity of a multiple of 512 Kb/s can be set up
between two (point-to-point) or more (point-to-multipoint) nodes, configured in a ring or bus
topology. Moreover, the bandwidth of such a circuit can be dynamically adapted according to
the traffic requirements between two nodes. As such the static, fixed capacity SDH
connections are replaced by dynamic variable bit rate circuits suited for packet-based
transport. The attractive features of SDH are retained, such as separation of data and control
plane, service separation (QoS), and fast protection switching. Separation of data and control
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 55
plane is enabled by reserving a small set of slots in the DTM frame as control slots, used for
signalling between the different nodes. These control slots are used by the slot reservation
process, which is based on a distributed algorithm that maintains the status of free time slots
and hands out time slots upon request. DTM supports different reservation schemes,
according to the QoS that is required: constant delay, guaranteed bandwidth, best effort.
Recovery of a failure in a DTM ring can be based on two mechanisms: the ring can either
reconfigure itself through detection by the link state protocol, which identifies changes in the
topology (the recovery time depends on the ring size). In this way no protection capacity
needs to be set up in advance and in failure free conditions, the entire ring bandwidth can be
used. Alternatively a circuit can be protected by a dedicated protection circuit along the other
side of the ring, such that fast protection switching can be achieved (see section 3.7). This at
the expense of nailing up protection capacity in advance, which is wasted in case no failures
occur. Hence the latter is only done for mission critical traffic.
3.7 Reliable network architectures
In this paragraph we give an overview of different network architectures, which can
be applied to transport networks. The described architectures mainly target the design of
reliable SDH en WDM network architectures. As mentioned in section 3.3.7, these network
architectures should allow reactive actions to be taken in the event of network failures, in
order to recover the affected traffic. We distinguish network architectures based on their
topological structure, as well as based on the recovery mechanism that is used.
The three network topologies considered are linear, ring and mesh topologies. Linear
topologies refer to point-to-point transmission between two network elements. Ring
topologies collect a number of nodes in a closed loop, such that 2 transmission paths exist
between each node pair on the ring. In each node, traffic can either be added/dropped locally
or passed through to the next node. The equipment used in the nodes of the ring to achieve
this are add-drop multiplexers (ADM). Unlike ring networks, meshed networks typically have
a much higher and less regular connectivity and rely on cross-connects capable of routing
traffic between multiple incoming and outgoing transmission links as well as adding and
dropping local traffic.
The two recovery mechanisms considered in transport networks are protection and
restoration [88]. Both mechanisms reroute affected traffic along a different path in the
network on which spare resources (also denoted as reserve capacity) are available. However,
the way the recovery mechanism is implemented and executed is different for both recovery
schemes. Protection reroutes the traffic on pre-assigned spare resources that have been
provisioned for a pre-determined failure or set of failures, which makes the rerouting 100%
predictable. Restoration, on the other hand, makes use of a pool of spare resources in the
network and for each failure, a restoration algorithm computes (real-time or pre-planned) a
restoration path within these spare resources. As such, restoration can make better use of the
shared spare resources. The restoration path and reliability of the network depend on the
amount of spare resources and the restoration algorithm.
Protection can be implemented in 3 ways, depending on the way the protection
resources are used. The simplest architecture has one dedicated protection entity for each
working entity (1+1). Traffic is sent ("bridged") on both entities at the transmitter side and the
receiver end selects one of both traffic streams, based on monitoring information (single
CHAPTER 35
6
ended switching). Alternatively, the protection entity can also be used to send low priority
traffic in case the working entity is not affected (1:1), using dual ended switching. In this case,
the low-priority traffic has to be pre-empted from the protection entity in case of a failure.
Thus, some form of communication is required between transmitter and receiver to co-
ordinate the switching actions, using a so-called automatic protection signalling (APS)
protocol. Protection entities can also be shared amongst multiple working entities. The most
complex architecture has M protection entities shared amongst N working entities (M:N),
with M<N. This architecture also requires an APS protocol to co-ordinate the switching.
In the following, an in-depth overview of the different recovery schemes will be
given. This overview will mostly be based on the standardized protocols in SDH/SONET
[89][90][91]. Because of the functional resemblance of SDH and WDM technology and
equipment, most of the SDH recovery schemes are also expected to become available for
WDM networks. Therefore, we will always try to discuss the correspondent WDM
architecture as well.
3.7.1 Trail protection
Trail protection [4] is a dedicated end-to-end protection mechanism, that can be
applied to both linear, ring and mesh architectures. A working trail is replaced by a protection
trail if the working trail fails or if the performance falls below the required level. The
performance is monitored by the trail termination function (e.g. using the POH in SDH). Trail
protection can operate in a unidirectional or bi-directional manner, and it may be 1+1, where
the dedicated protection trail is used only for protection purposes, or 1:1 where extra traffic
may be supported on the protection trail. Revertive operation, as well as non-revertive
operation is possible. Revertive operation means that after a failure on the working trail has
been repaired, the traffic on the protection trail is switched back to the working trail.
Trail protection is a simple recovery mechanism that provides end-to-end
survivability for any single failure in the network (except for failures of the terminating
nodes), provided working and protection trails are physically diversely routed in the network.
However, a failure or maintenance action on the working trail leaves the protection trail
unprotected. In case of long trails in a large network, this might seriously deteriorate the
network availability.
In SDH, trail protection can be implemented both on the LOP, HOP as MS layer. In
WDM it can be implemented on the OCh and OMS layer. Advantages and disadvantages of
protecting in higher or lower layers are discussed in section 3.7.6. Two examples of trail
protection are given below.
3.7.1.1 Linear protection
Linear architectures refer to point-to-point transmission between two network
elements. In SDH, linear protection switching is traditionally executed on the multiplex
section layer. This multiplex section protection (MSP) uses 2 parallel multiplex sections
between two network elements (in this case, multiplexers), and can be based on 1+1 or 1:N
protection. This is demonstrated in Figure 15 (in which bridging of the traffic at the
transmitter is represented by the symbol, while the selector at the receiver is indicated by
the symbol). Switching is executed based on the monitoring overhead in the MSOH.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 5
7
Diverse routing of working and protection multiplex sections reduces the chance of both MSs
failing at the same time.
M
U
X
M
U
X
Working MS
Protection MS
Figure 15: Linear multiplex section protection
In WDM, the same principle of linear protection can be applied. In fact, for first
generation WDM systems, this was the only protection mechanism available, mainly based on
proprietary techniques from the vendors. Switching at OMS level was executed based on
optical performance parameters, such as power and signal-to-noise (SNR) ratio. Some
vendors also support linear protection switching at the OTS level, to protect weak spans
between amplifiers.
3.7.1.2 Path protection
In a meshed network, or in an interconnected ring network, the path between 2
network elements can be end-to-end protected by using trail protection. Two diversely routed
trails between both network elements act as working and protection trail. The working and
protection trail can be link disjoint if the protection mechanism only has to protect link
failures or it can also be node disjoint if the protection mechanism also has to protect node
failures. The latter does not include the terminating nodes at the network elements, which can
not be protected. An example is shown in Figure 16.
MUX MUX
Working path
M
U
X
M
U
X
M
U
X
M
U
X
Protection
path
Figure 16: Path protection
CHAPTER 358
3.7.2 Subnetwork Connection Protection (SNCP)
Subnetwork connection protection (SNCP) is a dedicated protection mechanism that
can be applied to both linear, ring and mesh architectures. It may be used to protect part or all
of a network connection, in the same way as done for trail protection. However, since the
subnetwork connection has no access to the trail, performance monitoring has to be done
different [92]. Subnetwork connection protection using inherent monitoring (SNCP/I) protects
against failures in the server layer. The switching process and defect detection process are
performed by two adjacent layers: the server layer, providing the defect detection process and
the client layer, receiving a server signal fail (SSF) indication generated by the server layer
upon which switching is executed. Subnetwork connection protection using non-intrusive
monitoring (SNCP/N) uses client layer information (e.g. POH in SDH) to protect against
failures in the server layer and failures and degradations in the client layer.
SNCP can be used to protect connections end-to-end or within the different
administrative domains or tiers of a network. As such, the protection in each subnetwork can
be performed independent from the other subnetworks. This principle of protection
independence increases the overall reliability of the network. Indeed, by splitting up a
connection in multiple subnetwork connections that are all independently protected, multiple
failures occuring in different subnetworks can be protected, thus leading to an increased
network availability. The interconnection point between two subnetworks however remains a
vulnerable point (single point of failure); a problem which can be overcome by using dual
node interconnection schemes such as drop & continue (see section 3.7.3.3).
In SDH, SNCP typically works on the LOP or HOP level, e.g. as in SDH rings (see
section 3.7.3). Such rings are well suited for SNCP, since they intrinsically provide two node
and link disjoint paths between every node pair on the ring. The nodes on the ring typically
host ADMs that can either provide SNCP on the HOP layer or simultaneously on the LOP and
HOP layer. In the first case, all traffic needs to be protected on the HOP layer. In the latter
case, LO traffic can be protected on the LOP layer as well. In WDM, SNCP typically works
on the OCh layer, either in rings or in meshed (sub)networks.
3.7.3 Ring based architectures
Ring based architectures are of particular interest for constructing reliable transport
networks because rings provide a cheap and simple 2-connnected topology. We can
categorize rings according to three criteria:
• The direction in which working traffic flows on the ring: uni-directional or bi-
directional. In a uni-directional ring, working traffic can only be transmitted in either
the clockwise or counter-clockwise direction of the ring (see Figure 17a). Conversely,
protection traffic can only be transmitted in the counter-rotating direction. As such,
both directions of bi-directional working traffic have to be routed on a different side of
the uni-directional ring. In a bi-directional ring, working (and protection) traffic can be
transmitted in both directions (see Figure 17b). As such, both directions of bi-
directional working traffic can be routed on the same side of the bi-directional ring.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 59
working
protected
Figure 17a: Uni-directional ring Figure 17b: Bi-directional ring
• The way in which the protection is executed. This can be either linear protection at the
multiplex section layer, allowing to recover all failed connections of a multiplex section
at once. Alternatively, path protection can be executed, either at the LOP or HOP layer
for SDH rings or at the OCh layer for WDM rings. As such, each connection can be
individually protected (or left unprotected) by the terminating ADMs on the ring.
• The fact whether a capacity slot (i.e. time-slot in SDH, wavelength in WDM) around the
entire ring can be shared by multiple non-overlapping connections. The latter also
implies that protection capacity around the ring can be shared, resulting in reduced
capacity needs for these ring types. Hence these ring types are also denoted as shared
protection rings, as opposed to dedicated protection rings where such sharing is not
possible.
In fact, not all combinations of the above classification criteria are possible. For
instance, as we will see further on, shared protection rings require to be bi-directional. The
most important classification with respect to network planning is the distinction between
dedicated and shared protection rings. Hence we will first describe the dedicated protection
ring types, followed by their shared protection counterparts.
3.7.3.1 Dedicated protection ring
The unidirectional7 dedicated protection ring (DPRing) consists of a working ring
and a corresponding counter-rotating protection ring. As such, each connection on the
working ring can be protected along the reverse side on the protection ring. Because a
protected bi-directional connection is thus routed on both sides of the ring, it uses up an entire
capacity slot along the ring, such that this capacity can not be shared by any other connection.
As such, the capacity requirements of the ring equal the total amount of traffic to be protected
on the DPRing. Two implementations of the DPRing can be distinguished: path protection
rings and multiplex section protected (MSP) rings.
A. Dedicated path protection rings
Dedicated path protection rings use the path protection concept on a per ring level.
The protection switching is executed in the ADMs that the nodes of the ring host. In case 1+1
protection is used, the traffic is duplicated at the transmitting ADM side, and single-ended
switching occurs at the receiving ADM (using either trail protection or SNCP). Alternatively,
1:1 protection can be used, which does not permanently bridge the incoming traffic at the
transmitter side and thus requires dual-ended switching in event of failures (using trail
protection). The advantage of 1:1 protection is that the spare capacity can also be used for
7 Also a bi-directional implementation of the dedicated protection ring is possible
CHAPTER 360
transporting pre-emptible low priority traffic (also known as extra traffic). In the event of a
failure, the extra traffic will be pre-empted to free capacity for recovery of higher-priority
failing connections. Since the protection occurs per individual connection, it is also
straightforward to disable the protection mechanism for certain connections that are better left
unprotected (e.g. because protection or restoration occurs in other layers).
In Figure 18a, a uni-directional ring is used and thus both working paths of the bi-
directional demand are routed along different sides of the ring. As such, one fiber on the ring
only contains working traffic and the fiber in the reverse direction only contains protection
traffic. This means that any failure affects only one direction of every connection on the ring,
except for a failure of one of the terminating nodes of a connection. The latter can be useful
for in-service reparations of one of the fibers. Alternatively, a bi-directional ring can be used,
such that both working paths can also be routed on the same side of the ring.
Figure 18a: Dedicated path protection ring Figure 18b: Dedicated path protection ring after failure
The conceptual simplicity of the path protection ring is also appealing to emerging
technologies such as WDM. Many vendors are announcing so-called optical channel
dedicated protection rings (OCh-DPRing), capable of protecting individual wavelength
demands in a dedicated fashion using the path protection concept.
B. Dedicated MSP rings
Another possible implementation of the DPRing is to protect traffic on a multiplex
section basis. Using this protection scheme, a failure in the working ring is first detected by
the two ADMs adjacent to the failure, based on the monitoring information in the MS
overhead. Subsequently, both ADMs loop back all the affected traffic of the failed multiplex
section on the protection ring in the opposite direction (see Figure 19). This loop-back
procedure effectively replaces the failed multiplex section by a new multiplex section on the
protection ring. This implementation is however more complex than 1+1 path protection,
since it requires an APS protocol to coordinate the switching. One advantage over path
protection rings is that by performing the switching at the multiplex section layer, less
switching actions are required than when switching each subnetwork connection individually.
On the other hand the MSP ring can only deal with MSP failures, hence it does not allow to
disable the protection mechanism for certain connections in a straightforward way.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 61
Figure 19a: Dedicated MSP ring Figure 19b: Dedicated MSP ring after failure
3.7.3.2 Shared protection ring
The shared protection ring (SPRing) requires a bi-directional implementation, in
order to allow sharing of the ring capacity by multiple connections. As such, both directions
of a bi-directional connection can be routed on the same side of the ring, such that capacity on
the other side of the ring can indeed be reused by one or more non-overlapping connections.
Because a single failure will only affect (both directions of) one of the connections sharing the
same ring capacity, also protection bandwidth can be shared among these connections. Some
node failures might affect 2 connections sharing the same ring capacity, but since these
connections terminate in that node, they can not recover form the node failure anyway. The
implementation of the shared protection ring can consist of a 2- or 4-fiber implementation and
can use shared path protection or multiplex section protection.
A. Shared MSP rings
Shared protection rings have originally been implemented in SDH on the multiplex
section level. These rings are referred to as multiplex section shared protection rings (MS-
SPRing) [89] and optical MS-SPRings (OMS-SPRing) [93] in WDM. The 2-fiber
implementation uses two counter-rotating fiber rings. In case of SDH-over-WDM MS-
SPRings, the fibers can also be wavelengths of the WDM system, on which an SDH multiplex
section is carried. Half of the capacity (i.e. timeslots or wavelengths) on each fiber is reserved
for protection traffic. As such, working connections in one fiber can be protected using the
protection capacity in the other fiber, traveling in the other direction around the ring. Both
directions of a bi-directional connection are routed on the same side of the ring in different
fibers (see Figure 20a). The same ring capacity can be reused for another connection between
different nodes. There is no dedicated protection connection but instead, a pool of spare
capacity that can be accessed by different connections. In the event of a failure condition,
detected at the multiplex section level, the ADMs adjacent to the failure will loop back all the
affected connections of the failed multiplex section at once on the protection capacity of the
ring (see Figure 20b). An APS protocol is required to co-ordinate the switching and the shared
use of protection capacity (e.g. to prevent that loop-back is initiated after a second failure).
CHAPTER 362
Figure 20a: Shared MSP ring Figure 20b: Shared MSP ring after failure
If the same capacity slots (i.e. time slots in SDH and wavelengths in WDM) are used
for both directions of bi-directional working traffic in both fibers of the 2-fiber
implementation, the protection capacity has to use different capacity slots than the working
capacity on the same fiber [22]. In case of a protection switch, the slots thus have to be
interchanged, using either time slot interchange (TSI) in case of SDH, or wavelength
conversion in case of WDM. Because of the additional complexity and cost of these
conversions, it might be better to use a different assignment for working and protection
capacity. When the same capacity slots are used for working traffic on the one fiber and
protection capacity on the other fiber (and vice-versa, such that both directions of bi-
directional working traffic have to use complementary capacity slots), no such conversions
are needed.
Another possibility for avoiding conversions of time slots or wavelengths is to use a
4-fiber implementation of the MS-SPRing. In this case working and protection traffic
traveling in the same direction are carried over different fibers, which enables to use the same
capacity slot for both directions of bi-directional working traffic. The 4-fiber arrangement
combines both ring protection and span protection on the same architecture. The latter implies
that if only the multiplex section in the working fiber of the ring is affected, the parallel
protection fiber can be addressed by a simple span switch and no loop-back occurs. Certain
multiple failures can as such be fully protected.
MS-SPRings can recover from both nodal (i.e. ADM) and link failures. However,
when an ADM failure occurs, appropriate measures have to be taken to avoid misconnection.
Misconnection occurs when an ADM fails that terminates two connections sharing the same
capacity slot on the ring. The loop-back procedure will try to tie both connections to each
other thereby setting up an unsolicited communications channel between the two other
terminating nodes of both connections. To avoid misconnection, a principle named squelching
[89] has been devised. Based on so-called squelching tables distributed in all ADMs of the
ring, the ring can undertake suitable actions to prevent such misconnection.
B. Shared path protection rings
The shared path protection ring uses the same configuration as the MS-SPRing, and
the same advantages in terms of capacity sharing, only the protection switching is different
(using 1:N path protection). When a failure occurs, the affected connections are individually
switched at the terminating ADMs to the other side of the ring, using the protection capacity
on the ring (see Figure 21). Again an APS protocol is required, in order to co-ordinate the
switching and shared use of protection capacity.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 63
Figure 21a: Shared path protection ring Figure 21b: Shared path protection ring after failure
Shared path protection rings have gained particular attention in the WDM field,
where they are being referred to as optical channel shared protection rings (OCh-SPRing)
[94][95]. OCh-SPRings are seen as a logical successor of OCh-DPRings, because they also
protect traffic on a per-channel basis, but in a shared manner. As such, WDM rings can
gracefully evolve from OCh-DPRings to OCh-SPRings or even combine both protection
mechanisms [96]. Although OCh-SPRings have the same capacity requirements as OMS-
SPRings, they have different characteristics, which are particularly advantageous for WDM
application:
• OMS-SPRings rely on loop-back protection switching, such that the protection path of a
connection can potential span almost twice the entire ring circumference, whereas the
length of the protection path in a OCh-SPRing is in the worst case only once the ring
circumference. As optical signals suffer from power degradation and signal distortion,
they have to be amplified and regenerated at regular places. Longer protection paths, as
in the OMS-SPRing case, thus put more stringent requirements on the physical design
of the optical ring.
• A second advantage of the OCh-SPRing over the OMS-SPRing is that is
straightforward to provide protection selectivity. As protection occurs on a per-channel
basis it is easy to mix and match protection mechanisms using either shared or
dedicated protection or leaving certain channels unprotected around the ring. Shared
protection has to be implemented per 2 channels (working + protection), whereas
dedicated protection or no protection is implemented per single channel around the ring.
• A third advantage relates to the performance monitoring, which is usually easier to do
on an optical channel than on an optical multiplex section basis. One can either monitor
the client layer signal transported by the optical channel non-intrusively, or use
monitoring based on the optical channel overhead. While a consensus has already been
reached on the overhead of an optical channel (using a digital wrapper) making it
possible to assess the performance of the channel, methods to appraise the quality of an
optical multiplex section are still under investigation.
• A fourth important advantage of OCh-SPRings is that they do not suffer from possible
misconnection, as is the case for OMS-SPRings. Hence, no complicated squelching
mechanisms have to be developed as is the case for SDH MS-SPRings.
3.7.3.3 Ring interconnection
Rings are typically restricted in terms of their length and number of nodes allowed
on the ring, to avoid transmission impairments and to minimize the protection switching
CHAPTER 364
delay, or to guarantee a high ring availability. Therefore - or because of economical reasons -
large-scale networks are typically covered with multiple rings. While initially overlapping
rings can be created such that no interconnection between rings is required [97], ring
interconnection is a long-term requirement for an economical and manageable large-scale ring
network. The rings can be interconnected logically in two ways, either by defining a hierarchy
[98], or in a flat way without defining a hierarchy [99]. In the former case, nodes of the lower
level rings are selected to be combined in a hierarchically higher ring. The difference between
both approaches is in the routing of the inter-ring traffic, which is in a hierarchical set-up
straightforward, but more complex in a flat interconnected set-up, as will be explained in
chapter 6. The hierarchical interconnected ring networks have therefore a simpler routing and
management model.
On the physical level, there are three options for ring interconnection [100]. The
simplest option is to directly interconnect the traffic dropped at the tributary side of the ADM
in one ring to the tributary side of the ADM in the other ring. Typically, both ADMs are co-
located in the same building in a so-called ‘back-to-back’ configuration. Such a hard-wired
interconnection (e.g. using a fiber distribution frame) does not offer any flexibility and is
therefore only appropriate for static traffic. A second option, adding more flexibility, can be
achieved by inserting a cross-connecting device between both ADMs. In that case, all the
traffic added/dropped in the ADM is fed into the cross-connect, which can be configured
(using software) in such a way to either terminate traffic locally, or pass it through to the
ADM of an adjacent ring. Finally, a third and most flexible option integrates both ADMs and
cross-connect in a single piece of equipment, providing both ring pass-through, add/drop and
interconnection functionality. This reduces the amount of wiring required between multiple
network elements. Such equipment, typically a large cross-connect, should thus also support
the protection switching protocols of the ring.
Ring interconnection offers protection independence (i.e. a failure occurring in one
ring does not trigger protection switching in the other ring), which allows surviving from
multiple failures taking place in distinct rings. Failures of the ring interconnection gateways
can be survived from by using dual node interconnection strategies, such as drop & continue
or other matched node mechanisms. In the remainder of this paragraph we will discuss drop &
continue for SNCP rings and MS-SPRings. For more information on drop & continue
implementations for different ring combinations, we refer to [90].
3.7.3.4 Drop & continue for interconnected SNCP rings
The drop & continue concept applied on SNCP rings is shown in Figure 22 for bi-
directional traffic between node A on ring 1 and node B on ring 2. For simplicity reasons, we
will only explain the uni-directional case from node A to B, while the explanation for the
reverse direction is analogous. At node A, traffic is bridged in a working and protection path
along both sides of the ring, following the SCNP scheme. At the first interconnection node
that the working path encounters, traffic is bridged again. One part is dropped locally at the
interconnection node, while the other part continues on the ring towards the second
interconnection node. This action explains the term drop & continue. Also the protection path
in the opposite direction is dropped & continued as shown on the figure. As such, both
interconnection nodes receive the working and protection signal from different directions,
from which they can choose. The selected signals at both interconnection nodes are then
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 65
routed on the second ring in different directions. In the terminating node B, one of both
signals is again selected.
As can be seen on the figure, the capacity usage for drop & continue is different on
both rings. For uni-directional traffic from ring 1 to ring 2, capacity is used on all sections of
ring 1, while on ring 2 the section(s) between both interconnection nodes are not utilized.
However, for the return path from ring 2 to ring 1, the situation is asymmetric, such that for
bi-directional traffic between both rings, all sections on both rings are fully utilized, as is the
case when no drop & continue is used.
Ring 1
Ring 2
A
B
Primary
interconnection
node pair
Secondary
interconnection
node pair
Figure 22: Drop & continue on interconnected SNCP rings
The use of drop & continue for interconnecting SNCP rings allows to protect each
ring individually against failures, such that double failures occurring in different rings can
also be protected. In addition a failure at one of the interconnection nodes, can also be
survived from (see Figure 23a). However, certain double failures, including an
interconnection node breakdown and a failure in the second ring can not be survived form. In
the example of Figure 23b, such a double failure affects the traffic flowing from node A to B,
while traffic in the reverse direction can still be recovered from.
CHAPTER 36
6
Ring 1
Ring 2
A
B
Ring 1
Ring 2
A
B
Figure 23a: Interconnection failure Figure 23b: Double failure
3.7.3.5 Drop & continue for interconnected MS-SPRing
The drop & continue concept for interconnecting MS-SPRings is illustrated in Figure
24. As MS-SPRings are bi-directional rings, both directions of the signal are routed along the
same route and the operation of the drop & continue protocol is symmetric for both directions.
For simplicity reasons, again only one direction (from node A on ring 1 to node B on ring 2)
will be described here. At node A, traffic is routed to one of both interconnection nodes, along
the side of the ring not crossing the second interconnection node. Traffic is dropped at this
interconnection node and continued towards the second interconnection node, where it is also
dropped. The dropped signal at the second interconnection node is then again routed to the
first interconnection node on ring 2 (along the side of the ring not crossing node B). At the
first interconnection node on ring 2, a selection can thus be made between the latter signal and
the signal dropped at the first interconnection node of ring 1. At ring 2, the selected signal is
then routed towards node B (along the side of the ring not crossing the second interconnection
node).
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 6
7
Ring 1
Ring 2
A
B
Ring 1
Ring 2
A
B
Figure 24a: Drop & continue on MS-SPRings Figure 24b: Drop & continue after failure
This implementation of drop & continue routes signals on both rings along the same
side and is thus referred to as same side routing (SSR). Alternatively, the signal can be routed
on different sides on both rings. This principle of opposite side routing (OSR), works along
the same way as SSR in ring 1. However the signal dropped at the first interconnection node
is now routed on ring 2 towards the second interconnection node (along the side of the ring
not crossing node B). At this second interconnection node in ring 2, a selection can now be
made between the latter signal and the continued signal of ring 1 dropped at the second
interconnection node. At ring 2, the selected signal is then routed towards node B along the
other side of the ring (thus not crossing the first interconnection node).
Both SSR and OSR implementations of drop & continue, use up capacity between
both interconnection nodes on both rings. This leads to an increased capacity usage on both
rings, compared to the case when no drop & continue is used (as opposed to SNCP rings,
where the use of drop & continue does not require any additional capacity to be nailed up). A
possible solution to relieve both rings, is to transport the drop & continue part of the traffic
between both interconnection nodes on the protection bandwidth of the ring as extra traffic.
This principle is also referred to as ring interworking on protection (RIP) [91]. In this case,
the working bandwidth between both interconnection nodes remains available for other
traffic. A failure within one of the rings pre-empts the drop & continue traffic on the
protection bandwidth of this ring, and addresses the protection bandwidth for executing loop-
back. On the other hand, a failure of the ring interconnection does not require any loop-back
in the ring, and thus the protection bandwidth can indeed be used for transporting the drop &
continue traffic. Another possibility for lowering the amount of traffic needed between the
interconnection nodes for the use of drop & continue, is to use a third interconnection node
located in between both other interconnection nodes. At the expense of such an extra
interconnection node, traffic can as such be better balanced between the interconnection nodes
[101].
Both SSR and OSR implementations of drop & continue, either using RIP or not, can
recover from any single node or link failure. However, with respect to double failure
CHAPTER 368
scenarios, the implementations have a different availability [102]. Using drop & continue on
working capacity is clearly the most reliable architecture. For this architecture there is no
difference between the SSR and OSR implementation. The use of RIP, results in a lower
availability, in particular, the OSR implementation which is very vulnerable to double failures
including one of the interconnection gateways.
3.7.3.6 Ring switched matched nodes for shared protection ring interconnection
Recently, a new concept for interconnecting shared protection rings has been
introduced by Nortel, called ring switched matched nodes (RSMN) [103]. This concept relies
on two main features. Firstly, ring interconnection occurs in two matched nodes, instead of
two matched node pairs, as in traditional ring interconnection strategies. As such, two nodes
belonging to two different rings are collapsed in one node, which provides add-drop
functionality on both rings, cross-connecting between both rings, and protection switching
functionality. This considerably reduces the amount of equipment and wiring needed. This
also reduces the amount of failure scenarios. Possible failures of the ring interconnection link
or one of both interconnection nodes are now reduced to a single − being it more catastrophic
− node failure. Secondly, when such a node failure should occur, protection switching can be
executed in an efficient way, without nailing up protection capacity in advance on any of the
rings, as is done in traditional drop & continue implementations. However, for each
connection that is interchanged between both rings at the primary node, the secondary node
must be 'aware' of this connection. This is done by establishing a so-called 'shadow
connection' in this secondary node, as shown in Figure 25a. In case of a failure, this shadow
connection is used to bridge traffic from one intra-ring protection channel in the one ring to
the adjoining intra-ring protection channel in the other ring (see Figure 25b). If the primary
node fails, the terminating nodes on both rings of the failed connections will detect the
failures. These nodes will switch to the protection path, independent of the fact whether the
failed connection is an intra-ring connection or an inter-ring connection. If the failure affects
an inter-ring connection, the secondary node will receive alarm signals from both rings, such
that it knows the shadow connection has to be activated. If the failure affects an intra-ring
connection, this failure can be restored within the ring.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 69
Ring 1
Ring 2
A
B
Ring 1
Ring 2
A
B
Figure 25a: RSMN before failure Figure 25b: RSMN after failure
While the above described architecture provides considerable advantages in terms of
bandwidth usage, there are also some questions to be asked about this implementation, of
which few details have been released so far. A disadvantage mentioned in [103] is the
somewhat lower availability of RSMN compared to traditional drop & continue. Secondly,
questions arise related to the scalability and stability of the switching process, especially when
multiple rings are interconnected. Thirdly, it seems difficult to provide independent
management of network elements within different subnetwork domains. As the matched
nodes belong to both rings, they can not be independently managed. Hence such an
architecture might not be suited for interconnecting rings belonging to different operator
domains.
3.7.4 Restoration mechanisms
Restoration relies on a more autonomous determination of backup routes and is
therefore more flexible and capacity efficient. Restoration strategies can be classified
according to three main criteria [100]:
• Link-based or path-based restoration
• Pre-computed or real-time restoration
• Centralized or distributed restoration
Figure 26 shows the principle of link restoration. In this case, the traffic is rerouted
between the end nodes of the failed link (provided that there is spare capacity). The capacity
is either kept together or rerouted in separate parts. This strategy typically consumes a lot of
spare capacity in the vicinity of the failure. Yet, different links can be protected sharing the
same spare capacity. In some cases backhauling may occur, i.e. the final route runs up and
down in the same link as can be seen in the example of Figure 26. This approach requires the
ability to identify the failed link. In the event of a node failure, the more complex scheme of
adjacent node restoration can be applied [104].
CHAPTER 370
MUX MUX
M
U
X
M
U
X
M
U
X
MUX MUX
M
U
X
MUX
M
U
X
M
U
X
MUX
M
U
X
M
U
X
Working paths
Link restoration
Figure 26: Link restoration
Figure 27 shows end-to-end path restoration. The restoration route can coincide
partly with the original route or a completely disjoint restoration route can be preferred. In
case of a link or node failure, all affected paths are individually restored. Thus much more
switching actions are required compared to link restoration, and the restoration time could be
slower. On the other hand, path restoration works at a smaller granularity and can use capacity
more globally, such that is makes better use of spare capacity in the network.
MUX MUX
M
U
X
M
U
X
M
U
X
MUX MUX
M
U
X
MUX
M
U
X
M
U
X
MUX
M
U
X
M
U
X
Working paths
Path restoration
Figure 27: Path restoration
In case of WDM networks, one also has to take into account the ability of the nodes
to perform wavelength conversion. For WP networks, continuity of the wavelength along the
restoration route is necessary. This restoration wavelength can however be different from the
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 71
wavelength of the working route if the transmitter and receiver are tuneable to another
wavelength. These wavelength requirements add additional complexity to the restoration
route calculation.
The restoration routes can be pre-calculated (before the failure happens) or
determined in real-time (after the failure happens). The speed and optimality requirements for
the restoration algorithms to compute the alternative routes, have a different priority in both
cases. Pre-calculation requires to calculate and store the restoration routes each time the
network status changes or at regular time intervals. The algorithm then has knowledge of the
entire network status, and the restoration routes for a set of pre-determined failure scenarios
can be stored in a database. Upon a failure, a capacity reservation protocol then verifies the
proposed restoration routes. Pre-calculation gives room for optimization, multiple failures are
not considered to occur in a short timeframe and thus the optimization can be done off-line.
Real-time restoration on the other hand requires, upon a failure, to figure out the current
network status and to propose adequate restoration routes. While real-time restoration can
typically survive from more unexpected failures than pre-calculated restoration, real-time
restoration is slower. The restoration process can be speeded up however, by sacrificing on
the optimality of the restoration algorithm.
In centralized restoration, computation of restoration routes is done in a centralized
network controller, where all necessary and up-to-date network information is available. After
computation, the routes are downloaded into the databases of the nodes. Real-time centralized
restoration is based on alarm messages to identify the failure and obtain topology information,
which is typically a slow process. Pre-planned centralized restoration is much faster, but
requires frequent communications between the centralized controller and network elements to
acquire up-to-date topology information, which might not be scalable for large size networks.
In addition, the dependency on a single centralized controller, makes this architecture less
reliable.
Distributed restoration can be real-time, based on flooding messages, sent out by the
terminating nodes of a failed link, that search for alternative routes. Although simple, this
mechanism is fairly slow and more importantly it has not yet shown to scale beyond single
link failures. Therefore, pre-computed distributed restoration seems like a more viable
alternative [105]. To restrict the amount of restoration routes to be stored in memory, it is best
to use failure independent restoration paths (i.e. completely disjoint from the working path).
One example is the ATM backup VP concept [106], which is path based and uses a node and
link disjoint restoration path (VP) per connection, such that each node and link failure can be
restored. Although every connection is assigned a restoration route, no dedicated capacity
needs to be provisioned. Other examples of pre-computed distributed restoration can be found
in [107].
3.7.5 Comparison between different protection and restoration mechanisms
In the previous sections, the principles and main characteristics of the different
protection and restoration mechanisms have been explained. Some advantages or
disadvantages of a particular scheme have already been briefly mentioned, but we have not
gone into the discussion about how the different mechanisms compare to each other. This
section will be entirely devoted to this comparison. Comparing the different network
architectures can be done according to a variety of criteria [108]. As some criteria are
CHAPTER 372
antagonistic in nature, it is sometimes difficult to find a single best architecture. Cost is often
an important benchmark for such a comparison. This cost consist on the one hand of
installation cost or capital expenditure cost (CAPEX), which in turn consists of link cost (i.e.
cost of fibers and transmission equipment) and nodal cost (i.e. cost of processing equipment
such as cross-connects). On the other hand, there is also a cost incurred with managing the
network. This does not only include the fixed cost of a management system, but more
importantly a recurrent cost for operating, administrating and maintaining the network on a
daily basis. The later cost is also referred to as operational expenditure cost (OPEX). Besides
cost, other considerations come into play as well. Flexibility involves the ability to cope with
unpredicted traffic patterns and failures (e.g. double link failures) and the scalability of the
network to reflect future growth. Also the availability of the network, subject to a particular
recovery scheme, should be considered. The availability of a path/network is the proportion of
time for which this path/network is expected to be available relative to the total time [88].
Finally, the different recovery schemes also have different recovery times (i.e. the time
between the occurrence of the failure and the restoration of the failure). As some applications
are critical with respect to recovery time, this is also an important basis for comparing the
different schemes.
In Table 3 we rank different recovery schemes according to the different criteria
mentioned above. We consider 6 types of recovery schemes. Two of these schemes rely on
the use interconnected rings, either using dedicated or shared protection. Two other schemes
use dedicated protection in a meshed network. This protection can occur either on the path
level (e.g. using trail protection or SNCP between the end-points of each connection) or on
the link level (e.g. linear protection of each link using a diversely routed protection route).
The last two considered schemes are restoration schemes in a meshed network, again either on
the path or link level.
Link cost Node cost Management
cost
Flexibility Availability Recovery
time
Ded. prot. ring Higher Lowest Low Mid/Low High Fast
Shared prot. ring Low Lowest Mid Lower High Fast
Mesh path prot. High High Low/Mid Mid Mid Fast
Mesh link prot Highest High/Mid Low/Mid Mid Mid Fast
Mesh path res. Lowest Mid/Low Higher High Mid/High Slowest
Mesh link res. Low Mid Higher High Mid/High Slower
Table 3: Comparison of network architectures
The link cost is highest for the dedicated protected schemes, because no protection
capacity can be shared. The meshed link protection scheme has the highest link cost because
this scheme requires a diverse protection route per protected link, and each protection route
typically spans a multitude of links on which capacity needs to be nailed up. In the meshed
path protection scheme only one protection path is required between the end points of each
connection. As such the protection works at a lower granularity. In addition the protection is
more global, such that the amount of capacity required (i.e. the amount of links spanned) by
working and protection path is more balanced and redundant capacity can be better spread out
over the network. In literature it has indeed been shown that path protection schemes use
protection capacity more efficiently than link protection schemes [109]. In the dedicated
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 73
protection ring a diverse protection path is provided within each ring that the working path
traverses. As such, the link cost of the dedicated ring protection scheme is typically situated
between the link costs of dedicated link and path protection.
With respect to the capacity requirement, protection requires substantially more
redundant capacity than restoration. On top of the working capacity, 100 to 200% redundant
capacity is required for protection, while only 50-100% for restoration [110]. Therefore the
lowest link cost can be expected for the shared protection and restoration schemes: shared
protection rings, but even more for shared restoration in a mesh because spare capacity can be
used more globally [111]. Also for mesh restoration, path based schemes have shown to use
spare capacity more efficiently than link based schemes [112].
Concerning the node cost, ADMs are typically much cheaper than comparative
cross-connect devices. Consequently, the ring based protection schemes should have the
lowest node cost if the amount of required ADMs does not considerably exceed the equivalent
amount of cross-connects. When interconnecting a large amount of rings, all ADMs of the
different rings have to be interconnected, which becomes highly complex. In this case, the use
of a more expensive cross-connect for ring interconnection can be considered to improve the
scalability. In addition, the use of drop & continue to interconnect rings, can increase the node
cost for ring based schemes. Although mesh based schemes relying on cross-connects are
expected to be more expensive, link based schemes can also rely on fiber switches to reduce
the switching matrix dimension and thus the node cost. As restoration schemes require less
spare capacity compared to the path protection scheme, the amount of traffic to be cross-
connected for restoration schemes is also lower, which in turn also contributes to a lower node
cost.
The management cost is lowest for the protection schemes, since they require no
signaling protocol (as long as 1+1 protection is used). For the shared protection ring, an APS
protocol is required, which adds up to the management cost. Also for restoration schemes in
meshed networks, signaling is required (either centralized or distributed). This signaling is
even more complicated, since multiple possible routes for the restoration path exist and may
need to be evaluated (in a dynamic restoration scheme). In addition, the restoration path has to
be set up by reconfiguring all the cross-connects on its route, while the shared protection ring
only needs to switch and bridge at two ADMs to perform the switching. Therefore mesh
based restoration comprises a higher management cost than shared protection rings.
The ring based schemes offer the lowest flexibility because they do not fully exploit
the connectivity of the network and the route of the protection paths is fixed. Rings are also
more difficult to upgrade compared to meshed networks, because the capacity should be
upgraded along the entire ring or a whole ring system must be added at once, while in meshed
networks upgrades can be made per link or per node. The flexibility with regard to
unpredicted demand dispersions is higher for the dedicated protection ring, because its sizing
only depends on the amount of traffic and not on the pattern. Meshed based schemes, relying
on OXCs, are most flexible as they can be configured to use any potentially free route in the
network. In addition, real-time restoration is also very flexible with regard to unpredicted
network failures.
Interconnected ring networks offer the highest availability, as multiple failures
occurring in different rings can be recovered from (protection independence). Ring
interconnection gateways can be protected using drop & continue, resulting in a very secure
architecture, consisting of survivable subnetworks. Long paths running over several rings can
CHAPTER 374
therefore recover from several failures along its track. Meshed link protection can also survive
from certain double failures, but can not cover node failures. Mesh path protection schemes
protect the paths end-to-end and if both working and protection path have failed the
connection can not be restored. Restoration schemes are more flexible and can react
adequately for several failure scenarios. However, it depends on the amount of spare
resources in the network, whether failures can be survived from or not. In this case, the
availability depends both on the over-dimensioning for spare resources and on the behaviour
of the restoration scheme. The link based schemes cannot recover from node failures, but they
can potentially recover better from multiple link failures. Overall availability thus depends on
the ratio of availability of nodes and links and the amount of spare resources.
Ring and path based 1+1 protection schemes allow to invoke the protection
switching in less than 50 ms as already demonstrated in SDH, because no complex signaling
is required. Also for shared protection rings fast recovery times have been demonstrated even
for WDM rings [113]. Restoration schemes, either real-time restoration (requiring extensive
messaging) or pre-planned restoration are expected to achieve slower recovery times than
protection. However, recent advances might allow more competitive recovery times [114].
So far, the network architectures and their availability were described based on one
type of recovery scheme. Normally the required availability is imposed by the traffic demand.
When considering different traffic types and some types require a higher availability than
others, protection selectivity is required. For example, traffic that has been protected in the
client layer should not be protected again in the lower layers. E.g. for the support of an IP
network, protection in the server layer may not be required, since IP routers can effectively
reroute traffic from broken links through other links. However restoration in the server layer
may still be preferred as a second-line of protection or in order to avoid congestion in the IP
network. Protection schemes in the client layer can also coexist with server layer recovery
schemes. This allows to protect the network against outages in the client layers, which the
server layer is unable to protect. On the other hand this also adds an additional cost and
interworking complexities to the network [115]. The existing myriad of services has very
different availability requirements from a server network point of view. In a meshed network
the different kinds of service can be easily combined. In rings, this could be achieved through
the support of non pre-emptible unprotected traffic and pre-emptible extra traffic in the ring.
To combine advantages of multiple architectures, one can think about alternative
architectures based on both ring and mesh influences. In [116] a hybrid ring mesh network is
proposed. Traffic is first routed in a mesh architecture and afterwards rings are inserted where
traffic well matches the traffic conditions to a ring architecture. Also in [88] and [117] hybrid
ring-mesh schemes have been investigated.
3.7.6 Multi-layer recovery
As already mentioned, current transport networks typically exist of several layers,
each with their own recovery mechanisms. In such networks, a major challenge is where and
how to provide and co-ordinate recovery mechanisms in the different layers of the network
[115][118], in order to optimize cost, availability, flexibility and maintainability of the
network.
Recovery at the lowest layer recovers the affected traffic in the lowest possible layer
(i.e. the layer closest to the failure). Because of the coarser granularities at which the lower
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 75
layers operate, less switching actions are required (compared to recovery in higher layers) to
recover large amounts of traffic, potentially stemming from multiple client layers (some of
which do not support recovery mechanisms of their own). However, not all failures can be
recovered at the lowest network layer: e.g. a node failure in one of the higher layers. To
recover from such failures, it is also necessary to implement recovery actions at the higher
layers. This leads to multiple complexities. First of all, the reserve capacity needed to recover
from failures in the higher layers also has to be supported by the lower layers. However,
protecting this client-layer protected traffic (and its reserve capacity) again in the server layer
leads to over-protected traffic and a very inefficient use of the network capacity. A solution is
to leave client-layer protected traffic unprotected in the server layer. An even more elegant
approach is to use a common pool of spare capacity at the server layer, which can be accessed
by either the server layer or client layer (as extra traffic), depending on the nature of the
failure [119]. Secondly, the co-existence of multiple recovery mechanisms in different layers
needs to be treated with extreme caution. Certain failures at the lower layers might trigger
recovery mechanisms in multiple layers at the same time, leading to uncontrolled and
unwanted behaviour of the network [120]. In such cases suitable recovery interworking
strategies (also called escalation strategies [121]) have to be engineered that consist of a set of
guidelines describing when to start and stop recovery in which layer, and how to co-ordinate
these activities in the different layers. Recovery can either be started parallel in multiple
layers (resulting in potentially fast recovery times) or in a sequential way by starting recovery
in one layer after the recovery activities in another layer have been halted (which is easier to
implement and control, e.g. using hold-off timers or explicit messaging between layers). The
sequential scenario, can either bottom-up (staring at the lowest layer) of top-down (starting at
the highest layer).
Recovery at the highest layer recovers traffic at the layer closest to the origin of the
traffic. This mechanism can deal with more failure scenarios at once because it can handle
both failures in the client layer, as well as failures in the server layers. However, one failure in
the server layer typically results in multiple concurrent failures in the client layer (in network
elements far away from the root cause of the failure), which makes recovery of these failures
in the client layer more complex and slow and possibly also more expensive. In addition,
protection or restoration paths in the higher layer need to be physically diverse from the
working paths. However diverse paths in the higher layer do not necessarily run over different
physical resources in the client layers, which adds an additional complexity to recovery at the
highest layer. On the positive side, recovery at the highest layer operates the granularity of the
traffic, such that different recovery mechanisms can be used for traffic with different
reliability requirements (e.g. leaving best effort services unprotected). The latter is also
referred to as protection selectivity. In addition, recovery at the highest layer also has the
advantage that it protects traffic against failures in any layer, such that no recovery
mechanisms at lower layers are required. This avoids the burden of implementing recovery
interworking strategies.
When making an economic comparison of recovery at the highest or lowest layer,
most studies [122][123] reveal that recovery at the lowest layer is cheaper than recovery at the
highest layer. Indeed, with recovery schemes operating in the highest network layer, the
reserve capacity allocated in the highest layer also has to be supported by the lower layers. It
is thus important to understand that deploying highest layer recovery schemes implies
additional investments in lower layers as well. Shifting the recovery towards the lowest layers
CHAPTER 37
6
thus allows to lower the cost of the higher layers (potentially increasing the cost of the lower
layers). The overall investment to obtain a survivable network is therefore usually lower using
lowest layer recovery instead of highest layer recovery. More details about multi-layer
recovery and planning can be found in [124].
3.8 Conclusion
In this chapter we have introduced the transport network technologies and
architectures, which form the building blocks for the subsequent chapters. First, the desired
functionalities were described, followed by an overview of the enabling technologies. This
technological overview was mainly focussed on SDH and WDM technologies, because these
are the prominent transport network technologies. Most planning problems in this thesis
indeed focus on SDH and/or WDM networks. For completeness, packet-switched
technologies have also been described, with particular attention paid to the integration of
packet-switched networks in transport networks. The last part of this chapter was devoted to
reliable network architectures. An overview was given of the main recovery mechanisms that
will be used in this thesis for SDH and WDM networks. This includes both ring-based
schemes (which will dominate in this thesis) and meshed architectures.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 7
7
3.9 References
[1] P. Demeester, "Telecommunicatienetwerken B", course at the Ghent University, 1999-2000.
[2] Alex Gillespie, "Access networks: Technology and V5 interfacing", Artech House, 1997.
[3] K.-I. Sato, "Advances in transport network technologies: Photonic networks, ATM and SDH", Artech
House, 1996.
[4] ITU-T Recommendation G.805, "Generic functional architecture of transport networks", 1995.
[5] ITU-T Recommendation M.3010, "Principles for a Telecommunications management network", 1996.
[6] S. Aidarous, T. Pleyvak (eds.), "Telecommunications management in the 21st century", IEEE Press,
1993.
[7] M. Sexton, A. Reid, "Broadband networking: ATM, SDH and SONET", Artech House, 1997.
[8] ISO 9000-3 Guideline, "Guidelines for applying the ISO 9001 standard to the development, supply,
installation and maintenance of computer software", 1997.
[9] ITU-T Recommendation G.702, "Digital hierarchy bit rates", 1988.
[10] ITU-T Recommendation G.707, "Synchronous digital hierarchy bit rates", 1993.
[11] ITU-T Recommendation G.708, "Network node interface for the synchronous digital hierarchy", 1993.
[12] ANSI T1.105-1995, "Synchronous Optical Network (SONET) - Basic description including multiplex
structure, rates and format", 1995.
[13] ITU-T Recommendation G.703, "Physical/electrical characteristics of hierarchical digital interfaces",
1998.
[14] ITU-T Recommendation G.957, "Optical interfaces for equipments and systems relating to the
synchronous digital hierarchy ", 1999.
[15] ITU-T Recommendation G.652, "Characteristics of a single-mode optical fibre cable", 1997.
[16] B. Mukherjee, "Optical Communication Networks", McGraw-Hill, 1997.
[17] R. Ramaswami, K.N. Sivarajan, "Optical networks: A practical perspective", Morgan Kaufmann, 1998.
[18] ITU-T Recommendation G.653, "Characteristics of a dispersion-shifted single-mode optical fibre
cable", 1997.
[19] R. Baets, "Optische Vezelcommunicatie", course at the Ghent University, 1999-2000.
[20] ITU-T Recommendation G.655, "Characteristics of a non-zero dispersion shifted single-mode optical
fibre cable", 1996.
[21] A.K. Srivastava et al., "Ultra-dense terabit capacity WDM transmission in L-band", Proceedings of
OFC 2000, Post-deadline paper PD27, Baltimore (MD), March 2000.
[22] P. Bonenfant, "Protection and restoration in optical networks", Proceedings of OFC'99, Tutorial
Sessions, pp. 199-214, San Diego (CA), February 1999.
[23] O. Gerstel, "Opportunities for optical protection and restoration", Proceedings of OFC'98, Session ThD,
pp. 269-270, San Jose (CA), February 1998.
[24] ITU-T Recommendation G.872, "Architecture of optical transport networks", 1999.
[25] E. Goldstein, J. Nagel, J. Strand and R. Tkach, "National-scale networks likely to be opaque",
Lightwave, February 1998.
[26] E.A. Dowdell, P.C. Dejneka, C. Weinstein, "Economic modeling for long-haul systems", Proceedings
of AMTC'98, San Diego (CA), July 1998.
[27] N. Wauters, "Modeling and evaluation of optical WDM transport networks", Ph. D. thesis (in Dutch),
Ghent University, 1996-1997.
[28] J.M.H. Elmirghani, H.T. Mouftah, "All-optical wavelength conversion: Technologies and applications
in DWDM networks", IEEE Communications Magazine, Vol. 38, No. 3, pp. 86-92, March 2000.
CHAPTER 378
[29] W. Van Parys, B. Van Caenegem, B. Vandenberghe, P. Demeester, "Meshed wavelength division
multiplexed networks partially equipped with wavelength converters", Proceedings of OFC’98, Session
ThU, pp 359-360, San Jose (CA), February 1998.
[30] J.M. Yates, M.P. Rumsewicz, and J.P.R. Lacey, "Wavelength converters in dynamically-reconfigurable
WDM networks", IEEE Communications Surveys, Vol. 2, No. 2, Second quarter 1999.
[31] M.J. Soulliere, "Interconnection of Optical Network Administrative Domains", T1X1.5 Contribution
98-118R1, September 1998.
[32] G. Harning, R. Ramaswami, "Optical transport network management", Proceedings of 11th Tyrrhenian
International Workshop on Digital Communications, Portofino (Italy), September 1999.
[33] G. Newsome, P. Bonenfant, "Optical channel overhead carried on the optical supervisory channel",
ANSI T1X1.5 Contribution 99-004, January 1999.
[34] X. Liu, "Subcarrier modulation of client signals: An effective means for carrying OCh overhead", ANSI
T1X1.5 Contribution 98-121, September 1998.
[35] M. Wright, "A proposal for a rate-preserving OCh overhead approach", ITU-T Contribution D.561 (WP
3,4/15), Geneva, June 1999.
[36] J. McDonough, "Optical channel header implementation in the optical transport network", ITU-T
Contribution D.372 (WP 3,4/15), Geneva, October 1998.
[37] G. Newsome, P. Bonenfant, "A proposal for providing channel-associated optical channel overhead in
the OTN", ANSI T1X1.5 Contribution 99-002, January 1999.
[38] J. Ballintine, "A proposed implementation for a digital 'wrapper' for OCh overhead", ANSI T1X1.5
Contribution 99-003, January 1999.
[39] C. Liou, "Operational and management solutions for the intelligent optical transport network",
Proceedings of NFOEC'99, Chicago (IL), September 1999.
[40] D. Awduche, Y. Rekhter, J. Drake, R. Coltun, "Multi-Protocol Lambda Switching: Combining MPLS
Traffic Engineering Control with Optical Crossconnects", IETF Internet Draft, November 1999.
[41] S. Chaudhuri, G. Hjalmtysson, J. Yates, "Control of ligthpaths in an optical network", IETF Internet
Draft, February 2000.
[42] K. Kompella et al., "Extensions to IS-IS/OSPF and RSVP in support of MPL(amba)S", IETF Internet
Draft, February 2000.
[43] B. Rajagopalan, D. Saha, B. Tang, K. Bala, "Signaling framework for automated provisioning and
restoration of paths in optical mesh networks", IETF Internet Draft, April 2000.
[44] J. Luciani, B. Rajagopalan, D. Awduche, B. Cain, B. Jamoussi, "IP over optical networks - A
framework", IETF Internet Draft, April 2000.
[45] ITU-T Recommendation G.692, " Optical interfaces for multichannel systems with optical amplifiers",
1998.
[46] O. Gerstel, R. Ramaswami and W.-K. Wang, "Making use of a two stage multiplexing scheme in a
WDM network", Proceedings of OFC 2000, Paper ThD1, Baltimore (MD), March 2000.
[47] J. Jungjohann, R. Schafer, "Emerging optical amplifier technology and its role in next-generation
networks", Lightwave, April 2000.
[48] Y. Sun, A.K. Srivastava, J. Zhou, J.W. Sulhoff, "Optical fiber amplifiers for WDM optical networks",
Bell Labs Technical Journal, pp. 187-206, January-March 1999.
[49] T. Strite, G. van de Hoven, "Trends in pump laser diode markets and technology", Ligthwave, February
1999.
[50] H. Masuda, "Review of wideband hybrid amplifiers", Proceedings of OFC 2000, Paper TuA1,
Baltimore (MD), March 2000.
[51] D. Yang, "Amplifiers at 1590 nm double the WDM bandwidth", Ligthwave, September 1998.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 79
[52] "Corning fixed WADM drop module and fixed WADM add module", Corning product information,
1999.
[53] F. Bentivoglio, B. Hein, "Assessing the developments in OADMs: How far is the technology from
commercial viability", Proceedings of Euroforum Conference on Optical Networks, London, November
1998.
[54] C.R. Giles, M. Spector, "The wavelength add/drop multiplexer for lightwave communication
networks", Bell Labs Technical Journal, pp. 207-229, January-March 1999.
[55] K. Bala, "Electronic switch fabrics for opaque optical crossconnects", Proceedings of OFC 2000, Paper
TuM3, Baltimore (MD), March 2000.
[56] R. Chua, Y. Hu, "Optical switches are key components in high-capacity, data-centric networks",
Lightwave, November 1999.
[57] S. Hardy, "Liquid-crystal technology vies for switching applications", Ligthwave, December 1999.
[58] L.Y. Lin, E. Goldstein, "Optical-layer networking: Opportunities for and progress in lightwave
micromachines", Proceedings of OFC 2000, Tutorial Session ThN, Baltimore (MD), March 2000.
[59] M. De Prycker, "Asynchronous Transfer Mode", Prentice Hall, 1995.
[60] ATM Forum AF-PHY-0128.000, "622 and 2488 Mb/s cell-based physical layer", July 1999.
[61] ITU-T Recommendation I.432, "B-ISDN user-network interface - Physical layer specification",
February 1999.
[62] X. Xiao, L. M. Ni, "Internet QoS: A big picture", IEEE Network, Vol. 13, No. 2, pp. 8-18, March-April
1999.
[63] D. Greenfield, "Terabit routers: A lesson in carrier-class confusion", Network Magazine, Vol. 15, No. 3,
pp. 78-84, March 2000.
[64] G.-S. Kuo (ed.), "Multiprotocol label switching", Special Issue of IEEE Communications Magazine,
Vol. 37, No. 12, December 1999.
[65] A.S. Tanenbaum, "Computer networks", Prentice-Hall, 1996
[66] "Internet protocol", IETF RFC 791, September 1981.
[67] S. Deering, R. Hinden, "Internet Protocol, version 6 (IPv6) specification" IETF RFC 1883, December
1995.
[68] Carl Wickman et al., "IP over WDM, transport and routing", Deliverable 1 of Eurescom project P.918,
October 1999.
[69] Eric Mannie et al., "IP and ATM integration", ACTS Guideline NIG-G3, June 1999.
[70] W. Simpson, "The point-to-point prototcol (PPP)", IETF RFC 1661, July 1994.
[71] W. Simpson, "PPP in HDLC-like framing", IETF RFC 1662, July 1994.
[72] J. Manchester, J. Anderson, B. Doshi, S. Dravida, "IP over SONET", IEEE Communications Magazine,
pp. 136-142, Vol. 36, No. 5, May 1998.
[73] W. Simpson, "PPP over SONET/SDH", IETF RFC 1619, May 1994.
[74] R. Braden, D. Clark, S. Shenker, "Integrated services in the Internet architecture: An overview", IETF
RFC 1633, June 1994.
[75] R. Braden et al., "Resource reservation protocol (RSVP)", IETF RFC 2205, September 1997.
[76] S. Blake et al., "An architecture for differentiated services", IETF RFC 2475, December 1998.
[77] T.M. Chen, T.H. Oh, "Reliable services in MPLS", IEEE Communications Magazine, pp. 58-68, Vol.
37, No. 12, December 1999.
[78] B.T. Doshi et al., "A simple data link protocol for high-speed packet networks", Bell Labs Technical
Journal, pp. 85-104, January-March 1999.
[79] IEEE Standard 802.3z, June 1998.
[80] http://www.gigabit-ethernet.org
CHAPTER 380
[81] V.W.S. Chan, K.L. Hall, E. Modiano, K.A. Rauschenbach, "Architectures and Technologies for high-
speed optical data networks", Journal of Lightwave Technology, Vol. 16, No. 12, pp. 2146-2168,
December 1998.
[82] Y. Chen, J. Turner, "WDM burst switching for petabit capacity routers", Proceedings of Milcom'99,
Atlantic City (NJ), October 1999.
[83] "Dynamic packet transport technology and applications overview", Cisco Systems White Paper,
February 1999.
[84] D. Tsiang, G. Suwala, "The SRP MAC layer protocol", IETF Internet Draft, June 1999.
[85] C. Bohm, M. Hidell, P. Lindgren, L. Ramfelt, P. Sjödin, "Fast circuit switching for the next generation
of high performance networks", IEEE Journal on Selected Areas in Communications, Vol. 14, No. 2,
pp. 298-305, February 1996.
[86] L.H. Ramfelt, "DTM offers value-added IP services to metropolitan area networks", Proceedings of
NFOEC'99, Chicago (IL), September 1999.
[87] M. Holmlund, "IP traffic in a DTM network", Net Insight White Paper, November 1999.
[88] D. Vercauteren, "Modeling and analysis of the integrity of telecommunications networks", Ph. D. thesis
(in Dutch), Ghent University, 1997-1998.
[89] ITU-T Recommendation G.841, "Types and characteristics of SDH network protection architectures",
1998.
[90] ITU-T Recommendation G.842, " Interworking of SDH network protection architectures", 1997.
[91] ANSI T1.105.01-1997 Standard, "Synchronous Optical Network (SONET) - Automatic Protection
Switching", 1997
[92] C. Coltro, "Evolution of transport network architectures", Alcatel Telecommunications Review, pp. 10-
18, 1st quarter 1997.
[93] X. Fang, R. Iraschko, R. Sharma, "All-optical four-fiber bi-directional line-switched ring", Journal of
Lightwave Technology, Vol. 17, No. 8, pp. 1302-1308, August 1999.
[94] M. Cvijetic, T. Shiragaki, and A.J. Weissberger, "OCh shared protection ring", ANSI T1X1.5
Contribution 99-178, July 1999.
[95] P. Arijs et al., "Architecture and design of optical channel protected ring networks", submitted to
Journal of Lightwave Technology, April 2000.
[96] P. Arijs, P. Demeester, "The merit of shared and dedicated protection WDM rings in a hybrid network
design", Proceedings of OFC 2000, Paper FE6, Baltimore (MD), March 2000.
[97] L. Wuttisittikulkij, M.J. O'Mahony, "Design of a WDM network using a multiple ring approach", Proc.
of Globecom'97, pp. 551-555, Phoenix (AZ), November 1997.
[98] B. Van Caenegem, P. Demeester, "Design of interconnected WDM ring networks", Proceedings of
ONDM'99, pp. 61-68, Paris (France), February 1999.
[99] P. Arijs, P. Demeester, "Dimensioning of non-hierarchical interconnected WDM ring networks",
Proceedings of ONDM 2000, pp 147-159, Athens (Greece), February 2000.
[100] T.-H. Wu, “Fiber network service survivability", Artech House, 1992.
[101] R. Nagarajan, P. Harshavardhana, "Dual SONET/SDH ring interworking: Capacity and cost penalty",
Proceedings of the IEEE Symposium on the Planning and Design of Broadband Networks", Mont
Tremblant (Canada), October 1998.
[102] J. Nikolopoulos, J. Leung, M. El-Torky, "Inter-ring bandwidth management planning considerations",
Proceedings of AMTC'99, San Diego (CA), July 1999.
[103] J. Nikolopoulos, J. Leung, M. El-Torky, "Optimization of inter-ring bandwidth management",
Proceedings of NFOEC'99, Chicago (IL), September 1999.
TRANSPORT NETWORK TECHNOLOGIES AND ARCHITECTURES 81
[104] K. Struyve, P. Demeester, L. Nederlof, L. Van Hauwermeiren, "Design and evaluation of distributed
link and path restoration algorithms for ATM meshed networks", Proceedings of Broadband
Communications, pp. 13-25, Zurich (Switzerland), February 1996.
[105] B.T. Doshi, S. Dravida, P. Harshavardhana, O. Hauser, Y. Wang, "Optical network design and
restoration", Bell Labs Technical Journal, pp. 58-84, January-March 1999.
[106] R. Kawamura, K. Sato, and I. Tokizawa, "Self-healing ATM networks based on virtual path concept"
IEEE Journal on Selected Areas in Communications, Vol. 12, No. 1, pp. 120–127, January 1994.
[107] I. Lievens, "Use of distributed rerouting in meshed ATM Networks", Ph.D. thesis, Ghent University,
1999-2000.
[108] J. Späth, H. Weißschuh, "Investigation of protection strategies: Problem complexity and specific
aspects for WDM networks", Proceedings of NOC'99, Vol. 2, pp. 68-75, Delft (Netherlands), June
1999.
[109] T. Shiragaki, H. Saito, "Optimum protection architecture for reliable dense-WDM lightwave networks",
Proceedings of ICC'98, pp. 181-186, Atlanta (GA), June 1998.
[110] P. Arijs, W. Van Parys, B. Van Caenegem, P. Achten, P. Demeester, P. Lagasse, "Design of ring and
mesh based WDM transport networks", forthcoming in Optical Networks Magazine, 2000.
[111] M. Garnot et al., "Dimensioning and optimization of the wavelength-division-multiplexed optical layer
of future transport networks", Proceedings of ICC'98, pp. 202-206, Atlanta (GA), June 1998.
[112] B. Van Caenegem, W. Van Parys, F. De Turck, and P. Demeester, "Dimensioning of survivable WDM
networks", IEEE Journal on Selected Areas in Communications, Vol. 16, No. 7, pp. 1146-1157,
September 1998.
[113] R. Iraschko et al., "An optical 4-fiber bi-directional line-switched ring", Proceedings of OFC'99,
Session TuK, pp. 154-156, San Diego (CA), February 1999.
[114] R. Doverspike, G. Sahin, J.L. Strand, R.W. Tkach, "Fast restoration in a mesh network of optical cross-
connects", Proceedings of OFC'99, Session TuL, pp. 170-172, San Diego (CA), February 1999.
[115] P. Demeester et al. "Resilience in multilayer networks", IEEE Communications Magazine, Vol.37, No.
8, pp. 70-76, August 1999.
[116] W.D. Grover, R. Martens, "Optimized design of ring-mesh hybrid networks", Proceedings of DRCN
2000, pp. 291-297, Munich (Germany), April 2000.
[117] P. Arijs et al., "SDH protection in long distance networks: A practical case study", Proceedings of
DRCN98, paper O3, Brugge (Belgium), May 1998.
[118] J. Meijen, E. Varma, R. Wu, Y. Wang, "Multi-layer survivability", Lucent (Optical networking group)
White Paper, 1999.
[119] M. Gryseels, P. Demeester, R. Clemente, M. Ravera, "Common pool survivability for ATM on SDH
networks", Proceedings of Networks'98, pp. 405-410, Sorrento (Italy), October 1998.
[120] J. Manchester, P. Bonenfant, "Fiber optic network survivability: SONET/Optical protection layer
interworking", Proceedings of NFOEC'96, Vol.3, pp. 907-918, Denver (CO), November 1996.
[121] K. Struyve, "Design and evaluation of distributed recovery in SDH-based ATM transport networks",
Ph.D. thesis (in Dutch), Ghent University, 1997-1998.
[122] M. Gryseels, S. Ohta, R. Clemente, P. Demeester, "A cost evaluation of service protection strategies in
ATM on SDH networks", Proceedings of DRCN'98, Paper O38, Brugge (Belgium), May 1998.
[123] S. Baroni, M.A. Qureshi, A. Rodriguez-Moral and D. Sugerman, "Backbone network architectures for
IP optical networking", Proceedings of OFC 2000, Paper TuK2, Baltimore (MD), March 2000.
[124] M. Gryseels, "Planning of multi-technology telecommunication networks", Ph.D. thesis, Ghent
University, 2000-2001.
CHAPTER 382
CHAPTER 4
Comparison of SDH network
architectures
4.1 Introduction
In this chapter we describe the design of a long-distance (e.g. pan-European)
network, based on SDH and WDM technology. Such networks are currently being deployed
and operated by several incumbent and start-up international operators, taking advantage of
the liberalization of the global telecommunications market. The early adoption of WDM and
SDH technology, allows to rapidly deploy a broadband network in order to tap new revenue
streams from cross-border services [1]. These services include both wholesale bandwidth
offered to other emerging carriers and providers, as well as data-centric services offered to
corporate customers.
Whereas WDM has already been widely deployed in the core network of operators
around the globe, its current application is limited to point-to-point systems between adjacent
nodes. This enables cost reductions in the amount of required regenerative equipment on the
links, delays fiber exhaustion and allows rapid capacity upgrades. However, all networking
functionality, such as routing and execution of recovery actions, still resides in the WDM
client layers. These client layers, in particular SDH, are not expected to disappear overnight,
because services at E1 (=2 Mb/s), E3 (= 34 Mb/s) or E4 (=140 Mb/s) rates continue to
contribute to a substantial amount of the current network traffic [2]. Therefore, in the near
future, SDH remains a key technology for provisioning, managing and protecting these
services. The flexibility that SDH offers related to circuit configuration and its protection
capabilities have currently not reached a sustainable level of maturity in the optical layer. In
addition, SDH cross-connecting allows to improve the utilization of the optical channels by
consolidating SDH traffic streams stemming from multiple origins [3].
In the remainder of this chapter we will describe, design, evaluate and compare
network architectures for near-term SDH-based WDM networks. For such networks, a myriad
of architectures, involving different recovery strategies and manifold equipment
configurations can be used. Therefore an accurate modeling and design of these different
architectures is of paramount importance, in order to compare the architectures and make
justified decisions about the best suited network architecture for the specific case under study.
Not only the architecture itself should be modeled accurately, but also the used equipment.
Indeed SDH equipment, often suffers from several limitations, that impact the network
design, and thus can not be neglected.
Design and comparison of SDH network architectures has been the subject of
numerous studies in literature. Early work by Wu [4], considers network designs for meshed
path protection and interconnected dedicated protection rings, using a DS3 (45 Mb/s) demand
CHAPTER 484
matrix, however without any detailed equipment model. The different architectures are
compared based on cost and availability, and based on this, the ring-based architecture was
found to be most economical. A similar study, targeting specifically at DXC based
architectures was done in [5]. In this study, hybrid models, combining both dedicated
protection rings and meshed path protection or restoration are brought forward, which turn out
to be particularly advantageous. In [6] meshed path protection is compared to both dedicated
and shared protection ring schemes. This is one of the few works that models the equipment
in detail, including certain add/drop limitations. The results, considering a network design
based on a DS1 (1.5 Mb/s) traffic matrix, show that path protection has the lowest installation
cost. In [7], mesh-based architectures (both path protection and restoration) are compared
with interconnected rings (both dedicated and shared protection), not mentioning equipment
limitations. The design was done for a VC-12 traffic matrix, and considers DXC 4/3/1
equipment (for grooming and path protection), DXC 4/4 (for cross-connecting, restoration and
ring interconnection) and ADMs for the rings. In this study, the restoration scheme turned out
to be most economical, followed by the shared protection ring scheme, while the path
protection scheme was most expensive. In [8], a very similar study was done using a number
of commercial planning tools. While the different tools sometimes produce incompatible
results, the general trend is that interconnected rings are the cheapest solution, followed
closely by restoration architectures. Surprisingly the difference between dedicated and shared
protection ring is very small. More comparisons can be found in [9], for path protection, mesh
restoration and interconnected shared protection rings. This time the design was done for an
STM-1 (155 Mb/s) traffic matrix, and the shared protection ring scheme came out on top. As
in most other work, no equipment limitations were taken into account. Finally, we would like
to finish this literature overview by mentioning the network planning tool produced by Lucent
[10], which is the most elaborated tool to the best of our knowledge. It does not only include
all possible network architectures, it also allows to consider different equipment
configurations and to consider the specific limitations of such equipment. In addition,
different traffic matrices with varying bit rates can be specified. However, this is a vendor
dependent tool, and might be hard to use with equipment from another supplier.
Most research described above considers a number of network architectures designed
for a traffic matrix with different bit rates. In most cases, a lot of effort is spent on developing
optimal planning algorithms to design the different architectures, but often the used
equipment model is kept very simple (not taking into account certain limitations of the
equipment). While this allows to make a rough comparison on the required network capacity
for the different architectures, it is not always suitable for detailed cost studies or equipment
procurement, because it might lead to a serious sub-estimation of the amount of required
equipment. Because of the different traffic matrices, different cost models, and different
simplified models of equipment that are made in the different studies, results differ
considerably and it is hard to draw general conclusions. While it is true that each case study is
specific, we try to make our study somewhat more general. We do this in the following ways:
• A more elaborated node and link model is used, revealing the full details of current
SDH equipment and its limitations. This requires an extensive modeling effort. In
addition, it seriously complicates the planning process, because it adds a lot of extra
constraints, which are often difficult to model. On the positive side, such a model gives
a much more realistic estimation of the equipment needed and the actual network cost
for the different network architectures.
COMPARISON OF SDH NETWORK ARCHITECTURES 85
• Most research regarding network architecture studies is based on one type of equipment
configuration (specifying which equipment is used in each node of the network and how
it is configured). Such an equipment configuration is often fixed in the model, and no
alternative options can be considered. Our work uses a model, which considers 5
different equipment configuration options that can be applied to any network
architecture (both ring and mesh based), and could potentially be extended to even more
options.
• Another extension of our work, compared to most other work, is that we consider
multiple traffic matrixes with varying bit rates. Because the equipment has different
limitations for different traffic types, the results of the network design will thus depend
on the traffic mix that is given as an input.
• In contrast to some other studies, we chose not to consider all possible network
architectures, but the ones that seem most interesting. From the above literature
overview, the three most appealing architectures seem to be restoration, path protection
and interconnected shared protection rings. However, we chose not to consider
restoration because it is hardly used in practice for SDH networks, due to the lack of
relevant standards, slower restoration times, and unpredictable network configuration
effects. Besides path protection and interconnected shared protection rings, we also
consider both these architectures combined in a hybrid architecture.
In the following sections, our approach will be outlined. The next section describes the input
parameters to the planning problem we consider. These different input parameters are network
topology, network traffic, adopted recovery strategy, link and node configuration and a cost
model. Afterwards the network design method is explained, which renders a feasible network
design for each combination of the input parameters. Finally, we discuss results for the
different equipment configurations, recovery strategies and traffic scenarios and draw relevant
conclusions. The main results of this chapter can also be found in [2][11].
4.2 Input to the network design problem
In this section the network design problem will be fully specified. The aim of the
design problem is to develop methods that allow the different network architectures to be
modeled, dimensioned and finally compared to each other based on a number of metrics (of
which cost will be one of the most important ones). The input parameters to this model are:
the network topology, the traffic matrix, the considered network architectures, the used
equipment and equipment configurations (including certain limitations), a cost model and
some external restrictions to be taken into account. Each of these input parameters will now
be discussed in more detail.
4.2.1 Network topology
We consider the network topology to be given as input, i.e. the location of nodes and
links between the nodes is given. While the network topology is fixed, we still have to
dimension the nodes and links accordingly for the different network architectures. The links
represent cables containing a bundle of fiber pairs, on which WDM can be deployed. On each
wavelength of the WDM system, the appropriate SDH equipment can then be connected.
CHAPTER 486
Thus, dimensioning of these links implies determining how many fibers to lit (i.e. how many
WDM systems to use) and how many wavelengths to use. The nodes of the network can be
divided in so-called points of presence (POP) and flexibility points. A POP is a node in which
traffic is injected. Such POPs can be found in the large cities, where global interconnectivity
services are offered. Each such city typically hosts two POPs, where customer networks can
interconnect, which results in a very secure architecture that can even survive POP failures
without any traffic impact (e.g. using drop & continue). On the other hand, a flexibility point
does not terminate any local traffic, but it only processes through traffic. Its main aim is cross-
connecting and consolidation of traffic. Dimensioning of the nodes (both POPs and flexibility
points) involves determining the amount of required equipment in the nodes for each of the
equipment configurations (see section 4.2.5).
4.2.2 Network traffic
To study the influence of the nature of the traffic on the equipment requirements,
traffic matrixes with varying bit rates are considered. We consider three traffic matrixes: at E1
(=2 Mb/s), E3 (=34 Mb/s) and E4 (=140 Mb/s) rates. We will consider 4 different
combinations of these traffic matrices, to which we will refer as traffic scenarios:
• A realistic mix of the three matrices (E1, E3 and E4)
• Only traffic at the lowest bit rate (E1 only)
• Only traffic at the highest bit rate (E4 only)
• No traffic at the lowest bit rate (E3 and E4 only)
When comparing the impact of each of these 4 traffic scenarios on the equipment
requirements, we will make sure that the total amount of traffic for each of these different
traffic scenarios is the same. E.g. translation of the 'E4 only' scenario to the 'E1 only' scenario
would involve multiplying all elements in the E4 matrix by 63.
4.2.3 Recovery strategies
In this section, we will discuss the three recovery strategies we have considered in
our study and we stress the relevant design constraints they impose. We refer to chapter 3 for
a more detailed description of SDH protection.
4.2.3.1 Path protection using end-to-end SNCP
In our design approach, we consider path protection based on end-to-end SNCP. The
protection switching is realized at the lowest possible layer, dependent on the equipment used
(e.g. at the VC-12 layer for E1 traffic). Each demand pair is supported by a working and a
protection path. In order to ensure survivability of any node or link failure, we stipulate that
this protection path should be node and link disjoint with the working path.
4.2.3.2 Multiplex Section Shared Protection Rings (MS-SPRing)
We considered STM-16 two-fiber MS-SPRings to be used throughout the network.
Protection thus occurs within the rings, on the multiplex section layer. Since the MS-SPRing
protocol inherently relies on a ring-structured network, a set of interconnected rings must be
determined in the underlying meshed network, covering all the POPs, such that all
connections in the network can be set up. When determining rings in the network, one has to
COMPARISON OF SDH NETWORK ARCHITECTURES 87
take into account the restriction that the maximum number of ADMs on the ring is limited to
16 (due to the implementation of the APS protocol). Each connection is set up along one path,
which travels over one or more rings. In case of a network failure, the ring in which the
failure occurs can independently resolve the failure. In order to protect ring interconnection
gateway failures, we consider that interconnected rings must have at least two nodes in
common, and the drop & continue mechanism is used.
4.2.3.3 Combination of SNCP path protection and MS-SPRing
SNCP path protection and MS-SPRings can be combined within the same network,
leading to a hybrid protection strategy. In our approach we consider connections to be
protected either using one or more interconnected MS-SPRings, or no MS-SPRings at all but
using SNCP path protection between the end points of the connection. The MS-SPRings will
thus carry both intra-ring traffic and inter-ring traffic (in case of interconnected rings), while
traffic that can not be handled by the rings is protected with SNCP. As such, the rings do not
have to cover all the POPs. The SDH resources for SNCP and MS-SPRing are completely
separated, i.e. one connection is always protected using one of both protection schemes. Other
approaches, where a connection is split up in several subnetwork connections of which some
are protected by SNCP and others by MS-SPRing, are not considered here, because of the
difficulties they impose on the network design method and the network management. The
main idea behind this hybrid protection strategy is to use the benefits of both protection
strategies by installing rings only where they are cost-efficient (because of the traffic pattern
in this region) and using SNCP path protection for the remainder of the traffic.
4.2.4 Link model
All links in the network are assumed to use a WDM system that can scale up to 40
wavelengths. The used WDM system is an open system using transponders. On each
wavelength an STM-16 SDH signal is conveyed. The sites for boosters, pre-amplifiers, in-line
amplifiers and regenerators are fixed. The modularity of the system is reflected by the amount
of transponders and regenerators, which scales with the amount of wavelengths deployed.
4.2.5 Node scenarios
For each of the three protection architectures mentioned in section 4.2.3, a wide
variety of equipment configurations can be applied in the nodes of the network. We consider
5 so-called node scenarios (NS), representing different ways of deploying equipment in the
nodes, such to handle traffic in a different way. These 5 node scenarios are all based on some
combination of the following types of equipment:
• LO ADMs for accessing higher and lower order traffic on an STM-16 link, supporting
SNCP protection switching (both on the HO and LO level) and MS-SPRing operation
(only on the HO level). As shown in Figure 1, the HO matrix has 8 STM-1 add/drop
ports: as such, up to 8 protected (or unprotected) VC-4s can be add/dropped. When
traffic is protected in the LO matrix, 2 (unprotected) add/drop ports are required on the
HO matrix per equivalent VC-4 containing protected LO traffic. Thus, only 4 equivalent
VC-4s can be protected on the LO level. If no protection is required on the LO level, 8
equivalent VC-4s can be added/dropped.
CHAPTER 488
.
.
..
.
.
HO matrix
LO matrix
STM-16STM-16
16 STM-1 ports
LO protected
traffic
HO protected
traffic
unprotected
traffic
8 STM-1
ports
8 trib.
ports
Figure 1: Lower order add-drop multiplexer (LO ADM)
• HO ADMs for accessing higher order traffic on an STM-16 link, supporting SCNP
protection switching and MS-SPRing operation. No LO traffic can be treated by the HO
ADM and therefore an additional LO STM-1 MUX is required for multiplexing the LO
traffic. The HO ADM can add/drop up to 16 protected or unprotected VC-4s. Thus for
protected traffic the ADM is 100% non-blocking. For unprotected traffic, only 50% of
the bandwidth (on east and west side of the ADM) can be accessed.
.
.
..
.
.
HO matrix STM-16STM-16
16 STM-1 ports
HO protected
traffic
unprotected
traffic
16 trib.
ports
Figure 2: Higher order add-drop multiplexer (HO ADM)
• LO STM-1 MUX to multiplex lower order signals in an STM-1 signal.
• DXC 4/3/1 for cross-connecting, consolidation and protection switching LO and HO
traffic between STM-1 ports. The DXC 4/3/1 comes in different modularities (we
consider 56, 112 and 224 ports) and is 100% non-blocking on both HO and LO level.
• DXC 4/4 for cross-connecting and protection switching HO traffic between STM-1
ports. Also the DXC 4/4 comes in different modularities (we consider 64 and 256 ports)
and is 100% non-blocking on the HO level.
• Regenerators, not performing any networking functions, but merely 'cleaning up' (i.e.
re-timing, re-shaping, re-powering) the distorted SDH signal.
The basic model of a node (with degree 2) is depicted in Figure 3. A node terminates
at least two fiber pairs. First a WDM terminal (de)multiplexer, splits up the individual
wavelengths, with each wavelength carrying an STM-16 signal. The STM-16 signals are then
terminated by the appropriate SDH equipment combination. Traffic must be terminated in the
SDH equipment for two reasons. First to add/drop (and protect) locally terminated traffic in
COMPARISON OF SDH NETWORK ARCHITECTURES 89
that node. Second, to route traffic, i.e. cross-connect and consolidate HO and LO traffic
between the different STM-16 signals (or wavelengths) of different fibers. In practice we
assume that the nodes always host at least the required equipment for cross-connecting all HO
traffic between the different STM-16s passing through the node, in order to make the network
100% non-blocking (in terms of node capacity) for routing HO traffic. This makes the
network flexible with respect to traffic changes. Additional equipment for add-drop of local
traffic can then be added as needed. The 5 node scenarios we considered for equipping the
node are discussed next. Each node scenario can be used both for SNCP as well as for MS-
SPRing operation, although the required amount of equipment will differ in both cases. The
node scenarios rely heavily on the use of ADMs, and try to restrict to use of DXC as much as
possible because they are much more expensive (see section 4.2.6). ADMs are used both to
access traffic from an STM-16 with the purpose of cross-connecting traffic to another STM-
16 (this is the flexibility part of the node scenario) and also to add/drop locally terminated
traffic from an STM-16 (this is called the access part of the node scenario). The amount of
ADMs required for the flexibility part is not dependent on the traffic matrix but only on the
link capacity, because they should allow a 100% non-blocking network. This means that the
required number of ADMs must be present, in order to cross-connect traffic from any
incoming STM-16 link to any outgoing STM-16 link. The amount of ADMs required in the
access part has to be calculated dependent on the traffic pattern (see section 4.3.3.2). If no
traffic is terminated in a node (e.g. in a flexibility point), no extra ADMs are needed for local
access.
fiber SDH traffic
cross-
connecting
Local Add-Drop traffic
W
D
M
W
D
M
Each wavelength
= 1 STM-16
Figure 3: Basic node model
In the next paragraphs we described the 5 considered node scenarios, based on the
aforementioned equipment.
4.2.5.1 NS-1: Based on LO ADMs
In this node scenario, only LO ADMs are used to terminate or gain access to an
STM-16 (or wavelength) exchanged between two fibers. As shown in Figure 4, the LO ADM
provides access on both the HO and LO level.
CHAPTER 490
fiber W
D
M
W
D
M
REG
LO ADM
LO ADM
HO add-drop traffic
VC-4
SNCP
140 Mb/s
2 and 34 Mb/s
SNCP
VC-4
LO add-drop traffic
VC-4 passing through
VC-4
Figure 4: Add-drop in node scenario 1
For LO protected SNCP traffic, only 4 equivalent protected VC-4s can be
added/dropped using the LO ADM. Thus in the worst case 4 LO ADMs in series are required
if all LO traffic needs to be add/dropped on an STM-16. For HO protected SNCP traffic, 8
protected VC-4s can be add/dropped per LO ADM, thus up to 2 LO ADMs in series are
required in the worst case. For cross-connecting traffic between different STM-16s, all VC-4s
must be accessible on an STM-16. Hence in the worst case, if all traffic (i.e. 16 VC-4s) on an
STM-16 must be dropped (for cross-connecting to other STM-16s) and 16 new VC-4s must
be added instead, also 4 LO ADMs in series are needed.
For MS-SPRing protected traffic, only HO protection is possible. LO traffic can still
be consolidated in the LO matrix of the LO ADM, but no LO protection is possible. Both for
add-drop of local traffic and for cross-connecting, up to 16 VC-4s must be accessible in the
worst case (i.e. 8 working VC-4s on the east and west side). This requires up to 2 LO ADMs
in series.
In case no HO or LO traffic needs to be accessed on a STM-16, it can be
transparently passed through, and the LO ADMs can be replaced by a single regenerator.
Grooming of traffic is only done at the edges of the network, where LO add/drop
traffic to the same destination node is groomed in VC-4s, using the LO matrix of the LO
ADM. In intermediate points of the network all such traffic is treated as HO traffic and no
consolidation is possible to improve the efficiency of poorly filled VC-4s.
Cross-connecting of (both native and groomed) HO traffic is done hard-wired, i.e.
manually interconnecting the VC-4 dropped at the tributary of one ADM to the ADM where it
needs to be added (see Figure 4). This eliminates the use of expensive DXCs but also impairs
the configuration flexibility that such equipment provides.
COMPARISON OF SDH NETWORK ARCHITECTURES 91
fiber W
D
M
W
D
M
REG
LO ADM
LO ADM
Hard-wired
VC-4 cross-
connecting
VC-4
VC-4 passing through
Figure 5: Cross-connecting in node scenario 1
In Figure 6, an example is given of how a node with a higher degree can be
constructed in an equivalent way.
LO ADM
LO ADM
LO ADM
W
D
M
W
D
M
LO ADM
LO ADM
LO ADM
WDM
140 Mb/s
Hard-wired
VC-4 cross-
connecting
HO add-
drop traffic
2 and 34 Mb/s
LO add-
drop traffic
VC-4 passing
through
Figure 6: NS-1 node with degree 3
4.2.5.2 NS-2: Based on LO ADMs and HO ADMs
This node scenario uses HO ADMs to terminate and protect HO add/drop traffic
(either using SNCP or in MS-SPRing operation) and to access HO traffic for cross-connecting
between STM-16 channels. In addition LO ADMs are used in tandem on the same STM-16 to
add/drop local LO traffic as shown in Figure 7.
CHAPTER 492
fiber W
D
M
W
D
M
REG
LO ADM
LO ADM
HO ADM
HO ADM
2 and 34 Mb/s
HO add-
drop traffic
140 Mb/s
LO add-
drop traffic
VC-4
VC-4 passing through
140 Mb/s
Figure 7: Add-drop in node scenario 2
For LO protected SNCP traffic, again up to 4 LO ADMs might be required in series.
For HO protected SNCP traffic, the HO ADM is used, and one such ADM can terminate all
HO protected traffic. For cross-connecting the HO traffic, the worst case requires all 16 VC-
4s on the east and west side to be terminated, thus 2 HO ADMs in series are required
(enabling 32 VC-4s to be terminated).
For MS-SPRing protected HO traffic, only one HO ADM is required to terminate all
8 working VC-4s on east and west side, for local add/drop. Also one HO ADM suffices for
terminating the (native or groomed) HO traffic for cross-connecting purposes. For LO
add/drop traffic, the LO ADM is used to groom traffic on an end-to-end basis in equivalent
VC-4s. Protection then happens on the HO level in the LO ADM (using MS-SPRing), thus in
the worst case 2 such ADMs might be required.
As in NS-1, grooming is only done at the edges of the network and HO traffic cross-
connecting is done hard-wired through the HO ADMs in the flexibility part of the node (see
Figure 8).
fiber W
D
M
W
D
M
REG
LO ADM
LO ADM
HO ADM
HO ADM
Hard-wired
VC-4 cross-
connecting
VC-4 passing through
VC-4
Figure 8: Cross-connecting in node scenario 2
4.2.5.3 NS-3: Based on HO ADMs, LO MUXs, and DXC 4/4
This node scenario only uses HO ADMs for accessing traffic on an STM-16. As such
only HO traffic can be terminated directly using the HO ADM. Therefore LO traffic must first
be mapped in HO traffic by a LO MUX (see Figure 9). As in NS-1 and NS-2, LO traffic is
groomed in VC-4s at the edges of the network, based on the destination.
COMPARISON OF SDH NETWORK ARCHITECTURES 93
W
D
M
W
D
M
REG
HO ADM
HO ADM
4/4 DXC
VC-4
140 Mb/s
MUX
VC-4
2 Mb/s
34 Mb/s
SCNP for LO traffic
performed in DXC
SNCP
VC-4 passing through
Figure 9: Add-drop in node scenario 3
Since LO traffic is first mapped in HO containers in this node scenario, protection
also takes place at the HO level. To increase the flexibility, this groomed LO traffic is
added/dropped via the DXC 4/4 in the HO ADM instead of directly via the ADM. As such,
the SNCP protection can occur in the DXC, and working and protection path can even be put
on a different wavelength (see Figure 9). Add/drop of LO traffic is thus looked upon as cross-
connected traffic from the HO ADM point of view (and thus accessed via the HO ADMs in
the flexibility part of the node). Add/drop of native HO traffic is still done via the HO ADM
in the access part of the node, such that only one HO ADM is required to add/drop this local
traffic per STM-16. This is true both in the SNCP as MS-SPRing case. For cross-connecting
in the MS-SPRing case, also one HO ADM suffices, while in the SNCP case two such ADMs
might be required to access all of the 16 VC-4s on the east and west side.
Cross-connecting of traffic between different ADMs is now no longer done hard-
wired but in a more flexible manner using a DXC 4/4, as shown in Figure 10. The size of the
DXC depends on the amount of traffic to be cross-connected. In this study, we assume the
DXC is only used for cross-connecting HO traffic which contains groomed LO traffic. Native
HO traffic is still cross-connected hard-wired, because this traffic is less likely to change over
time, and flexibility is less important.
W
D
M
W
D
M
REG
HO ADM
HO ADM
4/4 DXC
VC-4 cross-
connecting
through DXC
VC-4
VC-4 passing through
Figure 10: Cross-connecting in node scenario 3
CHAPTER 494
4.2.5.4 NS-4: Based on HO ADMs and DXC 4/3/1
This node scenario also uses only HO ADMs for accessing traffic on an STM-16.
This allows HO traffic to be terminated directly in the ADM, while LO traffic must first be
mapped in HO containers, which can now be done via a DXC 4/3/1 (see Figure 11). Since this
DXC 4/3/1 is present in every node, LO traffic no longer has to be groomed by destination at
the edges of the network. Instead, LO traffic bound towards different destinations is
consolidated in the same VC-4, and segregated again at the next node. In contrast to the
previously considered node scenarios, LO traffic can thus be terminated and rearranged in
each intermediate node, allowing to considerably improve the efficient use of capacity.
W
D
M
W
D
M
REG
HO ADM
HO ADM
4/3/1
DXC
Local add-drop traffic
VC-4
140 Mb/s
VC-4
2 Mb/s
34 Mb/s Grooming of LO traffic
in 4/3/1 DXC
VC-4 passing through
Figure 11: Add-drop in node scenario 4
Native HO traffic is again added/dropped directly via the HO ADM and cross-
connected hard-wired. This requires one such ADM per STM-16 in the MS-SPRing case, and
up to two in the SNCP case. LO traffic is added via tributary ports on the DXC 4/3/1 (thus no
LO MUXs are required anymore) and fed into the HO ADM as cross-connected traffic. The
SNCP protection occurs in the DXC 4/3/1. In the MS-SPRing case, LO traffic is just
consolidated in the DXC, while protection still occurs using MS-SPRing protection (at the
HO level). The disadvantage of this scheme is that a failure of the DXC can not be protected
using MS-SPRing, because no LO failures can be protected. Another implication is that no
drop & continue can be used for such LO traffic, because this also works at the HO level.
COMPARISON OF SDH NETWORK ARCHITECTURES 95
W
D
M
W
D
M
REG
HO ADM
HO ADM VC-12
consolidation
through DXC
VC-4
Local add-drop
4/3/1 DXC
VC-4 passing through
Figure 12: Cross-connecting (and consolidation) in node scenario 4
4.2.5.5 NS-5: Based in HO ADMs and LO MUXs
This last node scenario only uses HO ADMs for accessing STM-16 traffic streams,
and is thus much alike NS-3. Also LO traffic is groomed at the edges using a LO MUX (see
Figure 13). In contrast to NS-3, all cross-connecting is done hard-wired again, eliminating the
use of an expensive DXC, but decreasing the configuration flexibility (see Figure 14). The
same statements as in NS-3 hold regarding the required amount of ADMs in the flexibility
part of the node. In addition some extra HO ADMs might be required for add/dropping locally
terminated traffic, because now also LO traffic needs to be accessed via these ADMs, and not
via the ADMs in the flexibility part.
W
D
M
W
D
M
REG
HO ADM
HO ADM
VC-4
140 Mb/s
M
U
X
VC-4
2 Mb/s
34 Mb/s
VC-4 passing through
Figure 13: Add-drop in node scenario 5
CHAPTER 496
W
D
M
W
D
M
REG
HO ADM
HO ADM
Hard-wired
VC-4 cross-
connecting
VC-4 passing through
Figure 14: Cross-connecting in node scenario 5
4.2.5.6 Summary of node scenarios
In Table 1, we give an overview of the above described node scenarios in terms of
the required equipment.
SNCP ADM requirements MS-SPRing ADM requirementsFlexibility
equipment Flexibility Access Flexibility Access
NS-1 None 4 LO ADM 0-4 LO ADM 2 LO ADM 0-2 LO ADM
NS-2 None 2 HO ADM 0-4 LO ADM
0-1 HO ADM
1 HO ADM 0-2 LO ADM
0-1 HO ADM
NS-3 DXC 4/4 2 HO ADM 0-1 HO ADM 1 HO ADM 0-1 HO ADM
NS-4 DXC 4/3/1 2 HO ADM 0-1 HO ADM 1 HO ADM 0-1 HO ADM
NS-5 None 2 HO ADM 0-1 HO ADM 1 HO ADM 0-1 HO ADM
Table 1: Summary of node scenarios
4.2.6 Cost model
The cost of the SDH equipment is shown in Table 2. All costs are relative to the cost
of the LO ADM. As can be seen the LO ADM is more expensive than the HO ADM, because
it contains an additional LO matrix to deal with LO traffic. The DXCs are most expensive, in
particular the 4/3/1 DXCs.
SDH equipment type Relative cost
LO STM-16 ADM 100 %
HO STM-16 ADM 85 %
LO STM-1 MUX 25 %
STM-16 Regenerator 70 %
4/4 DXC with 64 STM-1 ports 210 %
4/4 DXC with 224 STM-1 ports 630 %
4/3/1 DXC with 56 STM-1 ports 800 %
4/3/1 DXC with 112 STM-1 ports 1400 %
4/3/1 DXC with 224 STM-1 ports 2400 %
Table 2: SDH equipment costs
COMPARISON OF SDH NETWORK ARCHITECTURES 97
The cost of the WDM equipment is shown in Table 3, relative to the cost of the
WDM multiplexer.
WDM equipment type Relative cost
Terminal Multiplexer 100 %
Booster 20 %
Pre-amplifier 25 %
In-line amplifier 70 %
Transponder 15 %
Table 3: WDM equipment costs
4.3 Network design method
The aim of our study is to compare the different recovery strategies and node
scenarios based on a number of relevant metrics: the so-called key performance indicators
(KPI). Our main focus will be to take into account the restrictions imposed by the different
node scenarios during the design process and study the impact of these node scenarios.
Ideally, the network design method should be different and optimized for each specific
recovery strategy and each node scenario. However, taking into account all possible
combinations, this would lead to 15 (=3x5) different design methods. To limit the number of
combinations, we will use a less optimal design method that initially only depends on the
recovery strategy. The impact of the node scenarios will only be taken into account in detail
during a later phase of the design process. This reduces the initial problem to 3 design
methods. Since such a design method does not take into account the node scenarios from the
start, it is a less optimal method, compared to using a more complex design method,
optimized for the specific node scenario. However, such a method still allows us to make a
representative estimation of the amount of equipment required for each node scenario and
recovery strategy, in order to make meaningful comparisons.
Our network design method consists of 4 phases, as shown in Figure 15, which will
be explained in more detail in the remainder of this paragraph.
CHAPTER 498
Ring definition
+ Ring routing
Routing
Traffic
evaluation
Topology information
E1, E3, E4 traffic matrix
Equipment
dimensioning
Calculation of
KPI's
OPTION 1
Mesh/SNCP
OPTION 2
Ring/MS-SPRING
OPTION 3
Hybrid Ring/Mesh
Traffic
evaluation
Equipment
dimensioning
Calculation of
KPI's
Ring definition
+ Ring routing
+ Mesh routing
Traffic
evaluation
Equipment
dimensioning
Calculation of
KPI's
Node scenarios
Cost info
equipment
Figure 15: Network dimensioning
In the routing phase, the routes of the working and (in case of SNCP) protection
paths are constructed. In case the topology includes rings, these rings have to be identified at
first.
In the traffic evaluation phase, the amount of traffic carried by each link is evaluated,
as well as the amount of add/drop and through traffic in each node. This is done based on the
routes obtained in the routing phase.
In the equipment dimensioning phase, we determine the amount of fiber transmission
systems required on each link and the amount of equipment needed in each node.
Finally in the last phase, we calculate the key performance indicators (KPIs).
The routing phase is independent of the node scenarios, but dependent on the
adopted recovery strategy. The traffic evaluation differs slightly depending whether end-to-
end grooming (NS-1, 2, 3 and 5) or intermediate consolidation (NS-4) is used. It is only in the
equipment dimensioning phase that the full details of the various node scenarios come into
play, and that each of these scenarios needs to be treated differently.
4.3.1 Routing phase
The three different routing strategies for the different recovery methods will be
shortly described hereunder.
4.3.1.1 End-to-end SNCP path protection
For routing the traffic using end-to-end SNCP, two node and link disjoint paths have
to be identified between the end points (POPs) of each connection. This can be done using the
algorithm of Suurballe [12], which identifies the shortest cycle between two nodes in a
network. The weights used on the links are the distances spanned between the end points of
the link.
4.3.1.2 MS-SPRing
For MS-SPRing, it is not sufficient to route the traffic, but in addition rings have to
be identified in the network. In a first phase we will route all traffic along the shortest path in
COMPARISON OF SDH NETWORK ARCHITECTURES 99
the meshed network, again using the link distances as weights. In a second phase we will
search for the best combination of rings to cover the entire network and carry the routed
traffic. The process for identifying the best ring combination and assigning the routed traffic
to the best-suited ring is presented in [13] and [14]. Special care should be taken, such that all
ring interconnections are realized through at least two nodes to allow the use of drop and
continue. The result of this process is a given set of rings, and the amount of traffic routed on
each ring.
4.3.1.3 Combination of SNCP path protection and MS-SPRing
When SNCP path protection and MS-SPRing protection are combined, the rings do
no longer have to cover the entire network, because the parts not covered by the rings can be
protected using SNCP. However, this means that all possible ring combinations would have
to be identified, which would be intractable in practice. Therefore the combination of SNCP
path protection and MS-SPRings is based on a pre-determined set of rings. These rings are
chosen manually (using a trial and error process) in these places in the network where the
traffic pattern and volume is best suited for such rings.
For routing in a network with a pre-determined set of rings not covering the entire
network, we use the following algorithm. Firstly, all inter-ring traffic is routed in the network
along the shortest path on the rings. For routing this inter-ring traffic, no load balancing on
each individual ring is performed. Secondly, all intra-ring traffic is routed (if multiple rings
are possible, the shortest ring is chosen). The intra-ring traffic is routed over the ring in such a
way to balance the loads evenly over the ring (also taking into account the already routed
inter-ring traffic). This is done by routing the intra-ring demands according to descending
capacity, in such a way that each extra demand results in the lowest increase in ring capacity
[15]. Finally, all traffic that can not be protected on the rings is protected end-to-end in the
same manner as done in 4.3.1.1.
4.3.2 Traffic evaluation phase
After the traffic has been routed and (in case of MS-SPRing) assigned to rings, we
will evaluate the amount of traffic that each link and node in the network processes based on
these fixed routings. This information will then be fed in the equipment dimensioning
process.
4.3.2.1 Link evaluation
The evaluation of the traffic processed by the links L is relatively easy. Two cases
have to be distinguished: whether traffic is groomed at the edges of the network (i.e. NS 1, 2,
3 and 5) or whether intermediate consolidation is possible (NS-4).
In the first case all traffic is mapped in VC-4s on an end-to-end basis, and all traffic
on the links will be VC-4 traffic. We distinguish the following types of VC-4 traffic on each
link i ∈ L:
• i
e4 : native VC-4 traffic (from the original E4 matrix), protected using SNCP (this
includes both working and protection paths).
• r
i
e4: native VC-4 traffic (from the original E4 matrix), protected on ring r (only the
working paths).
CHAPTER 4100
• i
e4' : VC-4 traffic containing groomed LO traffic (from the E1 and E3 matrix),
protected using SNCP (this includes both working and protection paths).
• r
i
e4' : VC-4 traffic containing groomed LO traffic (from the E1 and E3 matrix),
protected on ring r (only the working paths).
In the second case (NS-4), traffic can be optimally consolidated at the edges of each
individual link, resulting in a better filling of the VC-4s on that link. Thus in this second case,
we consider the following types of traffic on each link i ∈ L:
• i
e1 : SNCP protected traffic from the original E1 matrix (working + protection).
• r
i
e1: traffic protected on ring r from the original E1 matrix (only working).
• i
e3 : SNCP protected traffic from the original E3 matrix (working + protection).
• r
i
e3: traffic protected on ring r from the original E3 matrix (only working).
• i
e4 : SNCP protected traffic from the original E4 matrix (working + protection).
• r
i
e4: traffic protected on ring r from the original E4 matrix (only working).
4.3.2.2 Node evaluation
After the links have been dimensioned, this information can be used to dimension the
nodes. The evaluation of the traffic processed in the nodes N is somewhat more complicated.
The following information is required in order to dimension each node n ∈ N properly:
• Ln : the aggregate links terminated in the node n.
• mij : the amount of STM-16 channels (= wavelengths) between link i and j (i, j ∈ Ln)
belonging to MS-SPRings. This amount of channels is symmetrical because all links of
a ring need to be dimensioned equally.
• sij : the amount of STM-16 channels (= wavelengths) between link i and j (i, j ∈ Ln)
reserved for SNCP operation. This amount of channels is symmetrical, because traffic is
either passed through between two links in a node, or is added/dropped in that node
(using two links for working and protection path).
• AD
ij
e4, AD
ij
e4' : amount of native ( AD
ij
e4 ) and groomed ( AD
ij
e4' ) E4 traffic for SNCP, that is
add/dropped in node n using link i and j as working and protection path (i, j ∈ Ln).
• AD
ij
e3, AD
ij
e1 : amount of E3 and E1 add/drop traffic for SNCP, that is add/dropped in
node n using link i and j as working and protection path (i, j ∈ Ln). These values are
only applicable in case of NS-4.
• AD
ij
e4, AD
ij
e4' : amount of native and groomed E4 traffic on MS-SPRing, that is
add/dropped in node n between link i and j (i, j ∈ Ln).
• AD
ij
e3, AD
ij
e1 : amount of E3 and E1 add/drop traffic on MS-SPRing, that is add/dropped
in node n between link i and j (i, j ∈ Ln). These values are only applicable in case of
NS-4.
• PT
ij
e4, PT
ij
e4' : amount of native and groomed E4 pass-trough traffic between link i and j
(i, j ∈ Ln) for SNCP. This includes working and protection paths.
• PT
ij
e3, PT
ij
e1 : amount of E3 and E1 pass-trough traffic between link i and j (i, j ∈ Ln) for
SNCP. This includes working and protection paths. These values are only applicable in
case of NS-4.
COMPARISON OF SDH NETWORK ARCHITECTURES 101
• PT
ij
e4, PT
ij
e4' : amount of native and groomed E4 pass-trough traffic on links i and j (i, j ∈
Ln) of MS-SPRing. This only includes working paths.
• PT
ij
e3, PT
ij
e1 : amount of E3 and E1 pass-trough traffic on links i and j (i, j ∈ Ln) of MS-
SPRing. This only includes working paths. These values are only applicable in case of
NS-4.
4.3.3 Equipment dimensioning phase
Based on the description of how LO and HO traffic is to be treated in the different
node scenarios and the information gathered in the foregoing traffic evaluation phase, it is
now possible to calculate the amount of equipment on each link and in each node of the
network.
4.3.3.1 Link dimensioning
A. SNCP
The link dimensioning is straightforward. In case of end-to-end grooming (i.e. NS-1,
2, 3 and 5), the required amount of STM-16 channels ci required for SNCP on a link i is:
ù
ê
ê
é+
=16
4'4 ii
i
ee
c
Or, in case intermediate consolidation is possible (NS-4):
ù
ê
ê
é++
=16
43/)321/1( iii
i
eee
c
B. MS-SPRing
In case of MS-SPRing, similar formulas can be devised, not per link, but per ring,
because all links in the ring should have the same capacity (according to the most loaded link
on the ring). In addition, we have to take into account that only half of the capacity of the MS-
SPRing can be used for working traffic (the other half is reserved as protection capacity). In
case of end-to-end grooming, the required amount of STM-16 channels i
c required for the
MS-SPRings on link i is thus:
ù
ê
ê
ê
é+
=
∈∈8
4'4
max
r
j
r
j
ir rj
i
ee
c
Or, in case intermediate consolidation is possible:
ù
ê
ê
ê
é++
=
∈∈8
43/)321/1(
max
r
j
r
j
r
j
ir rj
i
eee
c
Hence, the total required link capacity equals:
ii
tot
iccc +=
In case only SNCP or MS-SPRing is used, one of both terms in the above sum is
zero. In case of the hybrid architecture, both terms can be non-zero.
CHAPTER 4102
4.3.3.2 Node dimensioning
The node dimensioning is much more complex, because each node scenario uses
equipment with different characteristics and limitations in different configurations. In
addition, also the recovery strategies adopted in the node impact the dimensioning of that
node. Thus for each node scenario, formulas have to be worked out that calculate the required
amount of equipment for each of the recovery strategies. Unlike the link dimensioning
formulas, it is impossible to devise formulas that calculate the exact amount of equipment
required in the nodes, because we do not know in detail which traffic is transported on which
wavelength. Therefore, our node dimensioning formulas will always reflect the worst case
scenario.
In the flexibility part of the node (i.e. ADMs installed for accessing cross-connected
traffic between STM-16s), it should be possible to access all VC-4s on each STM-16 link
through the node. Therefore, a large amount of ADMs might be required (see Table 1).
However, in most cases such a large amount of ADMs for flexibility is an overkill, since a lot
of traffic is either passed through on the STM-16 level, or terminated locally using the ADMs
for add/drop. Therefore, in our case study we will not always install the maximum amount of
HO ADMs for cross-connecting in each node scenario. The amount of ADMs to be installed
in the flexibility part of the node is shown in Table 4.
NS-1 (LO ADMs) NS-2, 3, 4, 5 (HO ADMs)
SNCP 2 1
MS-SPRing 2 1
Table 4: Amount of ADMs installed for cross-connecting
Calculating the required amount of ADMs for the access part is more difficult, and is
explained in more detail for each of the node scenarios hereunder. We will always calculate
the worst case amount of ADMs required, in order to ensure that all traffic can be carried in
the network.
A. Node scenario 1
A.1. SNCP
From the amount of native HO traffic add/dropped in a node, and the amount of
groomed LO traffic, we can devise the worst case amount of LO ADMs required. Two cases
can be distinguished. In case ij
AD
ij
AD
ij see ≤+ 4'4 , there are less or equal add/drop traffic units
(in terms of equivalent VC-4s) than STM-16 channels. In the worst case, all add/drop occurs
on different channels. In this case the worst case amount of LO ADMs required for SNCP
between i and j equals:
AD
ij
AD
ijij eeadmlo 4'4_ +=
In case ij
AD
ij
AD
ij see >+ 4'4 , the worst case is obtained when one unit of traffic is
add/dropped on each STM-16 channel and the remaining traffic is grouped such that it
accesses as much as possible the same channels. Two cases can be distinguished. In case
ij
AD
ij se ≥4 , the worst case is the situation in which each channel add/drops one unit of HO
traffic, and all LO traffic (which is twice as demanding on the amount of LO ADMs) is
COMPARISON OF SDH NETWORK ARCHITECTURES 103
add/dropped as much as possible on the same channels together with the remaining HO
traffic. If ij
AD
ij se <4 , the worst case is obtained by add/dropping each unit of HO traffic on a
separate STM-16 channel, and add/dropping one unit of groomed LO traffic on the remaining
AD
ijij es 4− channels. The remaining LO traffic is then add/dropped as much as possible on the
same channels. Combining these both cases, the worst case amount of LO ADMs required
between i and j is:
ú
ê
ê
ë
ê−−+
+= 8
)4.2,max()4'.24(
_
AD
ijijij
AD
ij
AD
ij
ijij
essee
sadmlo
The factor 2 for lower order traffic represents the fact the LO traffic is protected in
the LO matrix of the LO ADM, and thus requires 2 ports (working and protection) towards
the HO matrix of the ADM.
A.2. MS-SPRing
Similar calculations can be done for MS-SPRing. In case ij
AD
ij
AD
ij mee ≤+ 4'4, the
worst case amount of LO ADMs required between i and j is:
AD
ij
AD
ij
ij eeadmlo 4'4_ +=
Else, this worst case is:
ú
ú
ú
ê
ê
ë
ê−+
+= 8
4'4
_ij
AD
ij
AD
ij
ij
ij
mee
madmlo
In this case, LO traffic is not protected in the LO matrix of the LO ADM, but on the
HO level (using MS-SPRing), and no factor 2 (as in the SNCP case) is required.
A.3. Total
The total amount of LO ADMs to access traffic in a node can be obtained by adding
the amount of ADMs for SNCP and MS-SPRing for all i, j values applicable to that node n
(denoted as i(n) and j(n)).
)__(_
)(),(
ij
njni
ij
tot
nadmloadmloadmlo +=
B. Node scenario 2
B.1. SNCP
In this node scenario, the LO ADMs are only used to add/drop lower order traffic. In
case of SNCP and ij
AD
ij se ≤4' , in the worst case all add/drop occurs on different channels and
the amount LO ADMs needed for access between i and j is:
AD
ijij eadmlo 4'_ =
Else, the worst case refers to one unit of groomed LO traffic add/dropped on each
channel and all remaining LO add/drop traffic sharing as much as possible the same channels,
thus:
CHAPTER 4104
ú
ê
ê
ë
ê−
+= 8
)4'.(2
_ij
AD
ij
ijij
se
sadmlo
For HO traffic similar formulas can be devised for the amount of HO ADMs
required. In case ij
AD
ij se ≤4 , the worst case amount of HO ADMs for higher order traffic
between link i and j equals:
AD
ijij eadmho 4_ =
Else, since HO ADMs are non-blocking for HO protected traffic, one HO ADM per
wavelength suffices:
ijij sadmho =_
B.2. MS-SPRing
In case of MS-SPRing, and ij
AD
ij me ≤4' , in the worst case all add/drop occurs on
different channels and the amount LO ADMs needed for access is:
AD
ij
ij eadmlo 4'_ =
Else, the worst case refers to 1 unit of groomed LO traffic add/dropped on each
channel and all remaining LO add/drop traffic sharing as much as possible the same channels,
thus:
ú
ú
ú
ê
ê
ë
ê−
+= 8
4'
_ij
AD
ij
ijij
me
madmlo
For HO traffic, in case ij
AD
ij me ≤4 , the worst case amount of HO ADMs required to
obtain access to higher order traffic on the ring between link i and j equals:
AD
ij
ij eadmho 4_ =
Else, since HO ADMs are non-blocking:
ij
ij madmho =_
B.3. Total
Again, the total amount of ADMs to access traffic in a node n can be obtained by
adding the amount of ADMs for SNCP and MS-SPRing for all i, j values applicable to that
node.
)__(_
)(),(
ij
njni
ij
tot
nadmloadmloadmlo +=
)__(_
)(),(
ij
njni
ij
tot
nadmhoadmhoadmho +=
COMPARISON OF SDH NETWORK ARCHITECTURES 105
C. Node scenario 3
C.1. SNCP
In this case only HO ADMs are used and LO traffic is groomed using external
multiplexers. The amount of external multiplexers required for grooming lower order traffic
on HO ADMs equals (between link i and j):
AD
ijij emux 4'=
The LO traffic is protected in the DXC 4/4 and not in the access ADMs. Therefore
this protected LO traffic is accessed via the HO ADMs installed for flexibility (see Table 4)
and does not contribute to the amount of HO ADMs in the access part of the node. Thus, only
HO traffic contributes to the amount of HO ADMs required for access between link i and j.
The worst case amount of such HO ADMs required is thus:
)4,min(_ AD
ijijij esadmho =
Both LO add/drop and LO pass-through traffic is treated by the DXC 4/4. LO
add/drop traffic enters the DXC in one port, is split up in a working and protection part, and
leaves the DXC using 2 ports. LO pass-through traffic enters and leaves the DXC through one
port. Thus the amount of STM-1 ports on the DXC for SNCP operation between link i and j
is:
PT
ij
AD
ij
DXC
ij eep 4'.24'.3 +=
C.2. MS-SPRing
For MS-SPRing operation, similar formulas can be devised. For the amount of
external multiplexers between link i and j:
AD
ijij emux 4'=
And for the worst case amount of HO ADMs required:
)4,min(_ AD
ij
ij
ij emadmho =
Concerning the sizing of the DXC 4/4, LO add/drop traffic enters the DXC in one
port, and also leaves the DXC using 1 port since traffic is protected on the MS-SPRing and
not in the DXC as for SNCP traffic. LO pass-through traffic enters and leaves the DXC
through one port. Thus the amount of STM-1 ports on the DXC for MS-SPRing operation
between link i and j is:
PT
ij
AD
ij
DXC
ij eep 4'.24'.2 +=
C.3. Total
The total amount of HO ADMs required to access traffic in a node n is then:
)__(_
)(),(
ij
njni
ij
tot
nadmhoadmhoadmho +=
And the total amount of external multiplexers:
CHAPTER 4106
)(
)(),(
+=
njni
ij
ijn muxmuxmux
The DXC cost can be shared between SNCP and MS-SPRing operation. The amount
of ports required on the DXC is:
+=
)(),(
)(
njni
DXC
ij
DXC
ij
DXC
nppp
Depending on the amount of ports required, the appropriate size of the DXC can be
determined.
D. Node scenario 4
D.1. SNCP
As in NS-3, only HO ADMs are used, but now in conjunction with a DXC 4/3/1.
Again, LO traffic is protected in the DXC and added/dropped via the HO ADMs installed for
flexibility (see Table 4). HO traffic is added/dropped at the HO ADMs installed for access,
and the amount of such HO ADMs required between link i and j is thus:
)4,min(_ AD
ijijij esadmho =
D.2. MS-SPRing
Similar as in the SNCP case, the worst case amount of HO ADMs required for access
between link i and j is:
)4,min(_ AD
ij
ij
ij emadmho =
D.3. Total
The amount of HO ADMs required to access traffic in a node n is then:
)__(_
)(),(
ij
njni
ij
tot
nadmhoadmhoadmho +=
Because traffic can be segregated and consolidated in the DXC 4/3/1, all LO traffic
added/dropped via this DXC (to whichever destination, using either SNCP or MS-SPRing)
can share the same ports on the DXC for access. For the DXC ports towards the ADMs,
SNCP and MS-SPRing traffic must be segregated. LO SNCP add/drop and pass-through
traffic can be consolidated as well and use the same ports on the DXC. MS-SPRing add/drop
and pass-through traffic can also be consolidated, however we have to take into account
direction (east or west) of the MS-SPRing add/drop traffic. The amount of DXC ports
required for node n is thus:
ù
ê
ê
ê
é+++
+
ù
ê
ê
ê
é+++
=3
3321/)11(
.2
3
3321/)11(
)(),(
PT
ij
AD
ij
PT
ij
AD
ij
njni
AD
ij
AD
ij
AD
ij
AD
ij
DXC
n
eeeeeeee
p
ù
ê
ê
ê
é+++
+
ù
ê
ê
ê
é+++
+3
3321/)11(
3
3321/)11( )()()()( PT
ij
westAD
ij
PT
ij
westAD
ij
PT
ij
eastAD
ij
PT
ij
eastAD
ij eeeeeeee
COMPARISON OF SDH NETWORK ARCHITECTURES 107
The first term accounts for groomed LO add/drop traffic. The second term accounts
for working and protection LO SNCP add/drop and pass-through traffic. The third term
contains the LO MS-SPRing traffic that is add/dropped on the east side, or passed through.
The fourth term contains the LO MS-SPRing traffic that is add/dropped on the west side, or
passed through. Depending on the amount of ports required, the appropriate size of the DXC
can then be determined.
E. Node scenario 5
E.1. SNCP
This node scenario is very similar to node scenario 3, but instead of using a DXC
4/4, manual wiring is used. This implies that LO traffic is no longer protected in the DXC and
requires additional ports for add/drop on the access ADMs in which it is protected.
The amount of external multiplexers is equal to the amount of groomed lower order
traffic on HO ADMs, and equals (between link i and j):
AD
ijij emux 4'=
For HO traffic, the worst case amount of HO ADMs required between link i and j is:
)4'4,min(_ AD
ij
AD
ijijij eesadmho +=
E.2. MS-SPRing
Identical formulas as the above can be devised for MS-SPRing operation:
AD
ijij emux 4'=
)4'4,min(_ AD
ij
AD
ij
ij
ij eemadmho +=
E.3. Total
The total amount of HO ADMs required to access traffic in a node n is thus:
)__(_
)(),(
ij
njni
ij
tot
nadmhoadmhoadmho +=
And the total amount of external multiplexers:
)(
)(),(
ij
njni
ijn muxmuxmux +=
4.3.4 Key performance indicators
In this paragraph we explain the metrics that will be used for evaluating and
comparing the different node scenarios and recovery strategies. The following so-called key
performance indicators (KPI) will be used for this purpose.
The total installation cost of the network is a very important metric. This cost will be
split up in a link cost and a node cost to better asses the influence of the different node
scenarios on the total cost. The link cost comprises the cost of WDM equipment
(multiplexers, amplifiers, boosters, …), transponders and SDH regenerators in intermediate
landing points (without traffic add/drop) on a link. The node cost includes the cost of SDH
CHAPTER 4108
multiplexers, ADMs, DXCs and regenerators deployed in nodes with add/drop and flexibility
points.
The filling ratio gives an indication on how efficient the network uses the installed
capacity. The (total) link filling is the ratio of used capacity on a link versus the installed
capacity on a link. E.g. an STM-16 link conveying one working VC-12 and a protection VC-
12 has a link filling of 2/(16×63) = 0.2%. The primary link filling on the other hand, is the
ratio of used capacity for working traffic versus the installed capacity on a link. Thus for the
example above, the primary link filling is 1/(16×63) = 0.1%.
4.4 Results
We will present results for two sample networks. A first network, denoted as core
network, represents the core of a pan-European network, containing 25 nodes (20 POPs and 5
flexibility points) and 29 links. A second network, denoted as extended network, represents
the core network in addition to some extensions at the edges, and contains 66 nodes (49 POPs
and 17 flexibility points) and 81 links in total. Both considered networks have a node
connectivity of at least 2, to ensure path diversity.
For the core network we have also studied different traffic scenarios as described in
section 4.2.2. The values for the considered traffic matrices have been chosen such that the
total amount of traffic in each scenario is the same, and also the traffic pattern is very similar,
in order to make meaningful comparisons. The total amount of traffic considered was 9072 E1
equivalents (i.e. about 18 Gb/s) for the core network. This represents an average demand of
47.75 E1s per POP pair.
We will start by describing the results for a network design based solely on SNCP.
Afterwards the corresponding results for MS-SPRing will be presented. Finally, a comparison
between both architectures and a third hybrid architecture will be made.
4.4.1 Results for SNCP
In this paragraph we discuss the results for traffic protected using only SNCP. First
we present results for the different node scenarios applied to the core network using the most
realistic traffic matrix containing both E1, E3 and E4 traffic. Afterwards we will study the
impact of a larger network topology (the extended network) and traffic matrixes with different
granularities.
4.4.1.1 Comparison between the different node scenarios
In Figure 16, we represent the relative link cost for the 5 considered node scenarios.
The total link costs for NS-1, 2, 3 and 5 are all equal, because the routing of traffic is fixed
and these scenarios groom LO traffic at the edges of the network, such that each link
transports an equal amount of traffic. By consolidating LO traffic link-per-link, NS-4 provides
opportunities for link cost savings of 12% as can be seen from the figure. These link cost
savings are on the conservative side, because a large amount of the link cost is in fact a fixed
cost, independent of the amount of channels transported on a link. This fixed cost includes the
cost of the WDM multiplexer and optical amplifiers. In the end-to-end grooming case, this
fixed cost takes up 42% of the total link cost, while in the case of link-per-link consolidation,
it is 48% of the total link cost. The benefit of link-per-link consolidation becomes more clear
COMPARISON OF SDH NETWORK ARCHITECTURES 109
when looking only to the variable (i.e. traffic dependent) part of the link cost, where 21% can
be saved compared to end-to-end grooming. This is also witnessed when looking at the
average amount of wavelengths needed per link: NS-4 requires an average of 4.84
wavelengths per link, while the other scenarios require 7.89 wavelengths on average.
0%
20%
40%
60%
80%
100%
NS-1 NS-2 NS-3 NS-4 NS-5
Relative cost
Var. link cost
Fixed link cost
Figure 16: SNCP link cost
The benefit of intermediate consolidation is also shown in Figure 17, where the
average link filling (over all links in the network) is compared to the scenarios with end-to-
end grooming. When using end-to-end grooming only 54% of the installed capacity can be
effectively used, while the remaining part is lost, mainly due to bad filling of VC-4s with LO
traffic. In this case of end-to-end grooming, only 20% of the installed capacity is used for
working traffic (i.e. primary link filling), and 34% is used for protection traffic, which means
that long protection paths are used. If intermediate consolidation is used, almost 90% of the
installed capacity can be effectively used. Again a large portion of this capacity is used for
protection traffic.
CHAPTER 4110
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
End-to-end grooming Intermediate consolidation
Average filling ratio
Primary link filling
Total link filling
Figure 17: SNCP link filling
In Figure 18, the relative node cost of the different node scenarios is depicted. NS-1
is the most expensive scenario, because it only uses LO ADMs, which are limited in port
capacity. Therefore a large amount of such ADMs needs to be installed, inducing a large cost.
NS-2 also uses LO ADMs but only for LO traffic (thereby eliminating the need for LO
multiplexers). Add/drop of HO traffic and cross-connecting is done via the HO ADMs, which
are non-blocking. As such, NS-2 requires only 30% of LO ADMs and 45% of cheaper HO
ADMs compared to the total amount of LO ADMs needed in NS-1. This renders NS-2 32%
cheaper than NS-1. NS-3 requires the same amount of HO ADMs as NS-2 because LO traffic
is accessed via the DXC 4/4 and the HO ADMs in the flexibility part. Thus NS-2 and NS-3
use the HO ADMs in the access part only for native HO traffic. Besides the DXC, NS-3 also
requires external multiplexers. As such NS-3 is 22% more expensive than NS-2, but still 17%
cheaper than NS-1. In addition NS-3, offers flexibility to LO traffic via the DXC. NS-4
requires 15% less HO ADMs than NS-3, because LO traffic can be packed together more
efficiently. In addition no LO multiplexers are required, since LO traffic can be accessed via
the DXC 4/3/1. Nevertheless, the total node cost is 10% more expensive than for NS-3
because of the very expensive DXC 4/3/1. Finally, NS-5 is the cheapest option in terms of
node cost, because only HO ADMs and LO multiplexers are required and no DXC equipment
(which also means that the flexibility for LO traffic is lower). The amount of HO ADMs
required for NS-5 is 18% higher than for NS-2 and NS-3 because extra HO ADMs are needed
for access of LO traffic (whereas in NS-2 and NS-3 LO traffic is added via the LO ADMs or
via the cross-connect and the HO ADMs for flexibility).
COMPARISON OF SDH NETWORK ARCHITECTURES 111
0%
20%
40%
60%
80%
100%
NS-1 NS-2 NS-3 NS-4 NS-5
Relative node cost
Figure 18: SNCP node cost
Putting the link and node cost together, we obtain the total cost as shown in Figure
19. The link cost contributes to about 40 to 50% of the total cost. NS-1 is clearly the most
expensive solution, while NS-5 is the cheapest scenario. Also NS-2 is cost competitive as it
does not require LO multiplexers, but a limited amount of LO ADMs instead. Both options
using DXCs are about equally expensive (NS-3 is slightly cheaper): the lower link cost
obtained through efficient consolidation in NS-4 is thus neutralized by the higher node cost
(because of the more expensive DXC 4/3/1).
0%
20%
40%
60%
80%
100%
NS-1 NS-2 NS-3 NS-4 NS-5
Relative cost
Node cost
Var. link cost
Fixed link cost
Figure 19: SNCP total cost
CHAPTER 4112
4.4.1.2 Results for the extended network
For the extended network we only focus on NS-2, 4 and 5. NS-1 is not considered
anymore because it has shown to be too expensive due to the extensive use of LO ADMs. NS-
2 and NS-5 are the cheapest options for end-to-end grooming, while NS-4 is the only option
with intermediate consolidation.
The total cost of the extended network for the three node scenarios is show in Figure
20, together with a comparison with the results for the core network. NS-2 and NS-5 are
almost equally expensive for the extended network. The extra amount of LO ADMs needed in
NS-2 is offset against the LO multiplexers and additional HO ADMs in NS-5, making both
options about equally expensive. NS-4 has a 33% lower link cost than NS-2 and NS-5 and a
42% lower node cost. The lower link cost is due to the better link filling, as can be seen in
Figure 21. Without intermediate consolidation, the total link filling drops down to 19% for the
extended network (of which more than half is used for protection traffic). If intermediate
consolidation is used, the link filling is 77%, resulting in larger savings than for the core
network as also shown in Figure 21. With end-to-end grooming a lot of installed capacity is
thus very poorly used. Indeed, in a larger network, the amount of connections scales
quadratically with the amount of nodes. In addition the traffic at the edges of the network is
typically smaller than in the core of the network. Smaller traffic divided over more links thus
irrevocably leads to poor link filling in case end-to-end grooming is used. On the other hand,
this effect could be less strong in case the amount of traffic was higher. The fact that less link
capacity is needed in case of intermediate consolidation (an average of 3.43 wavelengths are
used per link compared to 14.24 in case of end-to-end grooming) not only impacts the link
cost but also the node cost. Indeed, since less channels are required, less channels need to be
terminated in the nodes. In addition the equipment is very efficiently used, compared to the
case of end-to-end grooming, where a lot of equipment is required to handle a large portion of
unused capacity.
0%
20%
40%
60%
80%
100%
NS-2 NS-4 NS-5 NS-2 NS-4 NS-5
Relative cost
Node cost
Link cost
Core Network Extended Network
Figure 20: Total cost of core network versus extended network
COMPARISON OF SDH NETWORK ARCHITECTURES 113
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Core network Extended network
Average total filling
End-to-end grooming
Intermediate consolidation
Figure 21: Link filling of core network versus extended network
4.4.1.3 Results for different traffic scenarios
The four different traffic scenarios described in section 4.2.2, have been applied to
the core network for all 5 node scenarios. The results on the link costs are presented in Figure
22, which shows results for end-to-end grooming (NS-1, 2, 3 and 5) and intermediate
consolidation (NS-4). The relative link cost is largely independent on the traffic scenario in
case of intermediate consolidation. Indeed, as the total amount of traffic is the same for all
scenarios and only the granularity is different, the effect of intermediate consolidation is that
all links are equally filled (see also Figure 23). For the traffic scenario with only E4 traffic,
there is no difference between end-to-end grooming and intermediate consolidation. Because
all VC-4s are optimally filled in this case, this is the optimal scenario for the end-to-end
grooming case. For the other traffic scenarios, end-to-end grooming requires a higher link
cost. Intermediate consolidation has the largest benefit over end-to-end grooming in case a
heterogeneous traffic mix with different granularities has to be transported (E1, E3 and E4
traffic), which can also be witnessed from the link fillings shown in Figure 23.
CHAPTER 4114
0%
20%
40%
60%
80%
100%
E1 E4 E1, E3 and E4 E3 and E4
Relative link cost
end-to-end grooming
intermediate consolidation
Figure 22: Relative link cost comparison for different traffic scenarios
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%
E1 E4 E1, E3 and E4 E3 and E4
Average total filling
End-to-end
grooming
Intermediate
consolidation
Figure 23: Total link filling for different traffic scenarios
The relative node cost for different traffic scenarios is depicted in Figure 24.
In case only E1 traffic is present, NS-4 has the highest node cost, because of the high
cross-connecting cost (all traffic is LO traffic and needs to be cross-connected). NS-1 is 12%
cheaper than NS-4 but still the second expensive option, because a LO ADM can only access
25% of the LO traffic on an STM-16, thus a lot of LO ADMs need to be installed in tandem.
For this traffic scenario NS-2 is the cheapest option, because it optimally uses the
functionality of the ADMs: LO ADMs are used solely for the add/drop of LO traffic and HO
ADMs only for cross-connecting. Thus using LO ADMs only for LO traffic when a lot of
COMPARISON OF SDH NETWORK ARCHITECTURES 115
such traffic is present, is cheaper than using HO ADMs in combination with LO multiplexers.
Indeed, a HO ADM fully equipped with LO multiplexers is more expensive than a LO ADM.
In case the traffic matrix only contains E4 traffic, all node scenarios except NS-1
have the same cost. Indeed, in this case NS-2 uses no LO ADMs and is thus equivalent to NS-
5. Also NS-3 and NS-4 only use the DXC for LO traffic, so no DXCs are needed either. The
cost of NS-1 is 92% higher than the other node scenarios, because the LO ADMs are more
expensive and have restricted access capabilities and thus a lot of these ADMs need to be
installed in tandem to obtain full access to the available bandwidth on an STM-16.
For the traffic matrix with a mix of E1, E3 and E4 traffic, NS-5 has the lowest node
cost as already explained in section 4.4.1.1. A similar evolution can be witnessed in case no
E1 traffic is present.
0%
20%
40%
60%
80%
100%
E1 E4 E1, E3 and E4 E3 and E4
Relative node cost
NS-1
NS-2
NS-3
NS-4
NS-5
Figure 24: Relative node cost for different traffic scenarios
Putting node and link cost together gives us the total cost as depicted in Figure 25.
Generally, the less E1 traffic is present, the lower the cost for each node scenario. Overall,
NS-5 is the cheapest scenario for all traffic scenarios. NS-2 is a competitive scenario when a
high amount of LO traffic is present. NS-1 and NS-4 are most sensitive to the presence of LO
traffic.
CHAPTER 4116
0%
20%
40%
60%
80%
100%
E1 E4 E1, E3 and E4 E3 and E4
Relative total cost
NS-1
NS-2
NS-3
NS-4
NS-5
Figure 25: Relative total cost for different traffic scenarios
4.4.2 Results for MS-SPRing
In this paragraph we discuss the results for traffic protected using only
interconnected MS-SPRings. First we present results for the different node scenarios applied
to the core network using the most realistic traffic matrix containing both E1, E3 and E4
traffic. Drop & continue is used for protecting inter-ring traffic, except for NS-4, since the use
of LO cross-connecting and drop & continue (which works on the HO level) is incompatible.
Afterwards we will again study the impact of a larger network topology (the extended
network) and traffic matrixes with different granularities.
4.4.2.1 Comparison between the different node scenarios
The relative total cost (divided in link and node cost) of the different node scenarios
is shown in Figure 26. As in the SNCP case, the link cost takes up 40-50% of the total cost
and a large portion of this link cost is again a fixed cost (independent of the traffic on the
link). The link cost is equal for all end-to-end grooming scenarios (NS-1, 2, 3 and 5), while
this link cost is 22% cheaper in case intermediate consolidation (NS-4) is used, because of the
better link fillings obtained (see Figure 27). In case of MS-SPRing, we can only talk about
primary link filling, and not about total link filling (as for SNCP), since no protection paths
are explicitly set up. The link fillings are lower than for SNCP, and the cost savings by using
intermediate consolidation are larger, because MS-SPRings have to be designed 'per ring'. If
one link on the ring has to carry a high amount of traffic, not only this link has to be designed
appropriately, but all other links on the ring have to follow as well. This results in a primary
link filling of a mere 13% in case of end-to-end grooming, and double this amount in case of
intermediate consolidation (which is more in line with SNCP).
The node cost is lowest for NS-4, closely followed by NS-5. Because of the large
savings in link filling obtained through intermediate consolidation, the average amount of
wavelengths needed per link drops from 12.4 (in case of end-to-end grooming) to 6.3. Thus
COMPARISON OF SDH NETWORK ARCHITECTURES 117
much less (about half) ADMs are required to access these wavelengths in NS-4. In addition,
the DXCs 4/3/1 needed in NS-4 are fairly small, because in case of MS-SPRing only one
(working) path per connection needs to be set up and cross-connected, while in case of SNCP
two paths (working + protection) need to be cross-connected resulting in larger DXCs. As
such, fewer HO ADMs and relatively small DXCs lead to a very competitive node cost for
NS-4. Because also the link cost is lowest for NS-4, this is clearly the most cost efficient
solution in case of MS-SPRing. On the downside, NS-4 can not protect failures of the DXC
4/3/1 because traffic is protected at the MS-SPRing level, and also no drop & continue can be
used for LO inter-ring traffic, leading to a lower overall reliability. NS-2, 3 and 5 require the
same amount of HO ADMs (because the HO ADM is completely non-blocking for MS-
SPRing), about twice the amount needed for NS-4. Still, NS-5 is competitive with NS-4 with
respect to node cost (because no DXC is used), but not with respect to total cost which is
about 15% more expensive. On the other hand, NS-5 has a higher reliability than NS-4, but
then again it does not offer any flexibility for LO traffic. NS-3 does offer this flexibility but at
a 27% node cost (and 12% total cost) increase over NS-5. NS-1 is most expensive, because it
requires twice the amount of LO ADMs than the amount of HO ADMs required in other end-
to-end grooming options. Because of the additional expensive LO ADMs required, the node
cost of NS-2 is 33% higher than for NS-5, and the total cost is 15% higher, making this the
second most expensive option overall. In general the use of NS-2 makes less sense for MS-
SPRing, because the HO ADMs are non-blocking, and can thus be reused for groomed LO
traffic (using a LO multiplexer) instead of installing additional LO ADMs.
0%
20%
40%
60%
80%
100%
NS-1 NS-2 NS-3 NS-4 NS-5
Relative cost
Node cost
Link cost
Figure 26: MS-SPRing cost
CHAPTER 4118
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
30.00%
End-to-end grooming Intermediate consolidation
Average primary link filling
Figure 27: MS-SPRing primary link filling
To assess the impact of drop & continue on the results for MS-SPRing, the
calculations for NS-5 were done with and without the use of drop & continue. The use of drop
& continue leads to an additional link cost of 7% and node cost of 12%. The total cost as such
increases by 9% and the link filling decreases by 13% compared to the case without drop &
continue. This is a result of the additional capacity used between the gateway nodes on the
rings. In fact this protection capacity is transported over the working capacity of the MS-
SPRing and is thus protected twice. To avoid this extra protection cost, part of the drop &
continue traffic can be carried on the protection bandwidth of the ring, using ring
interworking on protection (see chapter 3).
4.4.2.2 Results for the extended network
For the extended network, we only focus on NS-4 and NS-5, the cheapest end-to-end
grooming and intermediate consolidation option. The total cost of the extended network for
these node scenarios is shown in Figure 28, along with a comparison with the results for the
core network. For the extended network, the positive effect of intermediate consolidation is
even stronger than for the core network. The link cost of NS-4 is 50% cheaper than for NS-5,
because of the much better filling of the links, thus requiring only 6 wavelengths per link on
average compared to an average of 32 wavelengths in case of end-to-end grooming. As can be
seen from Figure 29, the average primary link filling in case of end-to-end grooming is only
4%, because a large amount of connections, with rather low capacities (especially at the edges
of the network) have to be groomed in separate VC-4s. In contrast, when using intermediate
consolidation, the same VC-4s can be shared for different connections, thus leading to a much
better primary link filling of 22%. The node cost of NS-5 is even 130% higher than that of
NS-4, because of the much lower amount of HO ADMs required in NS-4. This compensates
largely for the extra cost of the DXC 4/3/1 required in NS-4. As such, the total cost of NS-4 is
54% cheaper than NS-5, which shows that intermediate consolidation is clearly more
advantageous as the size of the network grows. As for the core network, the additional
COMPARISON OF SDH NETWORK ARCHITECTURES 119
advantage of NS-4 is the automated flexibility for cross-connecting LO traffic, while the
disadvantage is the lower reliability.
0%
20%
40%
60%
80%
100%
NS-4 NS-5 NS-4 NS-5
Relative cost
Node cost
Link cost
Core Network Extended Network
Figure 28: Total cost of core network versus extended network
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
30.00%
Core network Extended network
Average primary link filling
End-to-end grooming
Intermediate consolidation
Figure 29: Primary link filling of core network versus extended network.
4.4.2.3 Results for different traffic scenarios
The four different traffic scenarios described in section 4.2.2, have also been applied
to the core network for all 5 node scenarios using MS-SPRing. The results on the link costs
are presented in Figure 30, which shows results for end-to-end grooming (NS-1, 2, 3 and 5)
and intermediate consolidation (NS-4). As in the SNCP case, the relative link cost is largely
CHAPTER 4120
independent on the traffic scenario in case of intermediate consolidation, because the total
amount of traffic is the same for all scenarios and only the granularity is different. For the
traffic scenario with only E4 traffic, end-to-end grooming is slightly more expensive than
intermediate consolidation because the former uses drop & continue. Because all VC-4s are
optimally filled in this case, this is the optimal scenario for the end-to-end grooming case.
Intermediate consolidation has the largest benefit over end-to-end grooming in case a
heterogeneous traffic mix with different granularities has to be transported (E1, E3 and E4
traffic), which can also be witnessed from the link fillings shown in Figure 31.
0%
20%
40%
60%
80%
100%
E1 E4 E1, E3 and E4 E3 and E4
Relative link cost
end-to-end grooming
int ermed iate consolidatio n
Figure 30: Relative link cost comparison for different traffic scenarios
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
30.00%
35.00%
E1 E4 E1, E3 and E4 E3 and E4
Average primary link filling
End-to-end
grooming
Intermediate
consolidation
Figure 31: Primary link filling for different traffic scenarios
COMPARISON OF SDH NETWORK ARCHITECTURES 121
The relative node cost for the different traffic scenarios can be found in Figure 32.
Overall, NS-1 always has the highest node cost, independent of the traffic scenario (although
it gets worse as less LO traffic is present). This is because the LO ADM is blocking (both for
LO and HO traffic) in MS-SPRing configuration, while the HO ADM is fully non-blocking
for HO and groomed LO traffic in MS-SPRing configuration. As already noted, NS-2, NS-3
and NS-5 require the same amount of HO ADMs. In addition NS-2, uses additional LO
ADMs (except in the case of only E4 traffic) while NS-3 and NS-4 can reuse the non-
blocking HO ADMs for LO traffic using a simple LO multiplexer or the DXC 4/3/1. Overall
NS-5 has the lowest node cost. When only E1 traffic is used, all traffic needs to be cross-
connected in NS-4, making the node cost of NS-4 25% more expensive than NS-5. However,
the node cost of NS-4 is 2% cheaper in case of a heterogeneous traffic mix with different
granularities (E1, E3 and E4 traffic). In case the traffic matrix contains only E4 traffic, the
node cost of NS-2, NS-3 and NS-5 is equal, because they require the same amount of HO
ADMs and no DXCs. The node cost of NS-4 is slightly lower in this case because no drop &
continue is used.
0%
20%
40%
60%
80%
100%
E1 E4 E1, E3 and E4 E3 and E4
Relative node cost
NS-1
NS-2
NS-3
NS-4
NS-5
Figure 32: Relative node cost for different traffic scenarios
The total cost for the different traffic scenarios is shown in Figure 33. Generally, the
less E1 traffic is present, the lower the cost for each node scenario. NS-4 is the cheapest
option, certainly when the traffic matrix consists of a heterogeneous mix of granularities.
Otherwise, NS-5 is competitive and in addition it has a higher reliability.
CHAPTER 4122
0%
20%
40%
60%
80%
100%
E1 E4 E1, E3 and E4 E3 and E4
Relative total cost
NS-1
NS-2
NS-3
NS-4
NS-5
Figure 33: Total cost for different traffic scenarios
4.4.3 Comparison between the different architectures
In the previous paragraphs we have studied the impact of node scenarios, network
size and traffic granularities on the results for SNCP and MS-SPRing architectures separately.
In this paragraph we will make a head-to-head comparison between the results of both
network architectures and also study the third architecture, which combines SNCP with MS-
SPRing. To compare the different architectures, we will only focus on NS-4 and NS-5 since
these turned out to be most interesting based on the aforementioned results for SNCP and
MS-SPRing.
We start by comparing the cost of the SNCP and MS-SPRing architectures for the
core network. The link costs for both architectures are depicted in Figure 34. Both for NS-4
and NS-5 SNCP appears to be the solution with the lowest link cost. For NS-4, the gap is
small (4%), but for NS-5 the MS-SPRing architecture is 13% more expensive and 18% in
case drop & continue is used. Although MS-SPRings are capable of sharing protection
bandwidth, while SNCP can not do this, the link cost of the SNCP architecture is the lowest.
The main reason for this is that the traffic pattern on the rings is not in favour of MS-SPRing,
because a limited number of nodes on each ring attract most of the traffic (e.g. large business
centers and the gateway nodes between rings). In addition, each ring has to be designed worst
case (i.e. the ring capacity is dependent on the link carrying most traffic), while in the SNCP
architecture each link can be individually dimensioned to optimality.
COMPARISON OF SDH NETWORK ARCHITECTURES 123
0%
20%
40%
60%
80%
100%
NS-4 NS-5
Relative link cost
SNCP
MS-SPRing (without D&C)
MS-SPRing (with D&C)
Figure 34: Link cost comparison between SNCP an MS-SPRing for core network
Concerning the node cost for the core network, depicted in Figure 35, MS-SPRing is
the cheapest architecture. The explanation for this is twofold. First, the HO ADM is strictly
non-blocking for MS-SPRing operation, which is not the case for SNCP operation (thus
sometimes HO ADMs need to be installed in tandem). Second, in case of SNCP, each
connection has a working and protection path in the network, while MS-SPRings only require
a working path and the protection occurs in the rings. The protection path for SNCP (which is
typically longer than the working path) thus also contributes significantly to the amount of
traffic to be cross-connected in the nodes. This is particularly evident for NS-4, which
requires expensive DXC 4/3/1 for the cross-connecting of LO traffic. The node cost for SNCP
in NS-4 is 54% more expensive than for MS-SPRing. In case no DXCs are used for the cross-
connecting, the relative difference is smaller: in NS-5 the SNCP node cost is 18% more
expensive in case no drop & continue is used, and only 5% more expensive when drop &
continue is used.
CHAPTER 4124
0%
20%
40%
60%
80%
100%
NS-4 NS-5
Relative node cost
SNCP
MS-SPRing (without D&C)
MS-SPRing (with D&C)
Figure 35: Node cost comparison between SNCP and MS-SPRing for core network
The combination of link and node cost gives the total cost, as depicted in Figure 36.
The higher link cost for MS-SPRing is compensated by the lower node cost, rendering MS-
SPRing as the cheapest overall architecture for the core network. In NS-4, the MS-SPRing
architecture is 20% cheaper than SNCP. In NS-5, MS-SPRing is 3% cheaper when no drop &
continue is used, but 6% more expensive when drop & continue is used. On the other hand,
interconnected rings (in particular when using drop & continue) are more reliable than end-to-
end SCNP, because multiple failures occurring in different rings can be survived from (see
also chapter 6 for interconnected WDM rings).
0%
20%
40%
60%
80%
100%
NS-4 NS-5
Relative total cost
SNCP
MS-SPRing (without D&C)
MS-SPRing (with D&C)
Figure 36: Total cost comparison between SNCP and MS-SPRing for core network
COMPARISON OF SDH NETWORK ARCHITECTURES 125
We can do the same cost comparison between the SNCP and MS-SPRing
architecture for the extended network. The total cost (and constituting link and node cost) is
represented in Figure 37. NS-4 is clearly the cheapest solution for both SNCP and MS-
SPRing, which are almost equally expensive (MS-SPRing is 1.3% cheaper). On the other
hand, for NS-5, MS-SPRing is 38% more expensive than SNCP. This is because in the
extended network, with low traffic requirements at the edges, it is very difficult to find a
suitable set of rings that can be efficiently utilized when no intermediate grooming is allowed.
0%
20%
40%
60%
80%
100%
NS-4 NS-5 NS-4 NS-5
Relative cost
Node cost
Link cost
SNCP MS-SPRing with D&C
Figure 37: Cost comparison between SNCP and MS-SPRing for extended network
Besides the pure SNCP and MS-SPRing architecture, we will now also take into
account the hybrid architecture, combining both SNCP and MS-SPRings. The hybrid design
was performed for several combinations of rings in the core network, and the result with the
lowest overall cost was kept. This design included 3 rings, that were chosen such that the
amount of inter-ring traffic was not too high (in order to avoid the extra cost of drop &
continue). The results for the core network, are depicted in Figure 38. With the hybrid
architecture, a lower cost design can be obtained than with both single-architecture designs.
The hybrid design is 23% cheaper than SNCP and 4% cheaper than MS-SPRing for NS-4. For
NS-5, mixing SNCP and MS-SPRing, gives a cost advantage of 5% over pure SNCP and 11%
over pure MS-SPRing. The amount of traffic transported over the MS-SPRings in the hybrid
architecture is 47%, while the remainder is handled by SNCP.
CHAPTER 4126
0%
20%
40%
60%
80%
100%
NS-4 NS-5
Relative total cost
SNCP
MS-SPRing
Mix
Figure 38: Cost comparison with hybrid architecture for core network
Also for the extended network, we found a hybrid design that was cheaper than both
SNCP and MS-SPRing designs. This hybrid design included 4 rings, covering the core part of
the network. The results are shown in Figure 39. For NS-4, the hybrid design was 11%
cheaper than the SNCP design and 10% cheaper than the MS-SPRing design. For NS-5, the
hybrid design can do 2.2% better than SNCP and 29% better than MS-SPRing, because only
few rings are used in the hybrid case in the core of the network. The MS-SPRings in the
hybrid design for the extended network convey 74% of the total network traffic, while the
remainder is treated by SNCP.
0%
20%
40%
60%
80%
100%
NS-4 NS-5
Relative total cost
SNCP
MS-SPRing
Mix
Figure 39: Cost comparison with hybrid architecture for extended network
COMPARISON OF SDH NETWORK ARCHITECTURES 127
4.5 Conclusion
In this chapter we have described a framework for dimensioning SDH network
architectures, based on SNCP path protection and/or interconnected MS-SPRings. We
considered different alternatives for equipment configurations in the nodes of the network,
using SDH network elements such as ADMs and DXCs in different combinations. This has
been translated in so-called node scenarios, which describe the functional architecture of a
node, taking into account the capacity limitations of each network element. We developed a
network dimensioning process that takes the different recovery strategies and node scenarios
into account, and thus allows to study the impact of different combinations of recovery
strategies, node scenarios, network topology and traffic matrices on the network design.
For SNCP based network designs we concluded that the node cost is most sensitive
to the node scenarios. Due to the large fixed cost of each link, the potential savings in link
cost are more limited, although a substantial impact on link filling can be witnessed when
using intermediate consolidation (NS-4). Regarding link filling, we also noticed that a large
portion of the link capacity is dedicated to protection traffic. For the different node scenarios,
NS-1 was found to be most expensive due to the excessive amount of expensive LO ADMs
that are required (because these ADMs have limited connectivity). NS-5, using only HO
ADMs and no DXC yielded the lowest cost, while NS-2, using both LO and HO ADMs is a
competitive solution, especially when a lot of E1 traffic is present. In addition, we witnessed
that the more E1 traffic is present, the higher the node cost for all options. This is very
apparent for the node cost of NS-4. The link cost of NS-4 on the other hand is independent of
the traffic mix but only depends on the traffic volume. In general NS-4 becomes more
interesting compared to other options when a heterogeneous traffic mix with a lot of
granularities needs to be transported. Also for larger network, with less dense traffic, NS-4
offers considerable cost advantages.
For MS-SPRing, both node and link cost are sensitive to the considered node
scenario. Intermediate consolidation (NS-4) significantly improves the link cost. Also the
node cost is lowest for NS-4, rendering this design the cheapest overall option for MS-
SPRing. This option makes use of DXCs 4/3/1, which on the upside give higher routing
flexibility, but on the downside result in a lower reliability, because failures of the DXCs can
not be recovered from. NS-5 is a competitive node scenario for small networks with a dense
traffic matrix. NS-5 is more interesting when only one granularity of traffic is present, while
NS-4 is more suitable for a heterogeneous traffic mix. However for larger networks with less
dense traffic, NS-4 is definitely the most cost-effective solution. NS-1 and NS-2 are less
suited for MS-SPRing, because the LO ADM is superfluous due to the non-blocking nature of
HO ADMs for MS-SPRing (while the LO ADM is still 50% blocking). The influence of drop
& continue on the result can not be ignored. For the core network the total cost increases by
9% and the average link filling decreases by 13%.
When comparing SNCP with MS-SPRing, it seems that in case of end-to-end
grooming (e.g. NS-5) SNCP is the cheapest solution. This is especially true for larger
networks, as it is difficult to find a suitable set of rings, covering the whole network and
yielding a good filling on each ring. This filling can be considerably improved when using
intermediate consolidation (NS-4). In this case, MS-SPRing becomes more cost effective than
SNCP, also because the cross-connecting cost is lower for MS-SPRing, since only one path
CHAPTER 4128
per connection exists, while SNCP requires a working and protection path. For each node
scenario and network topology, hybrid networks combining SNCP and MS-SPRing, yield a
lower cost than both single technology architectures. This is achieved by placing rings in
strategic positions, with suitable traffic patterns (e.g. in the network core).
COMPARISON OF SDH NETWORK ARCHITECTURES 129
4.6 References
[1] P. Golden, "Carriers set for a battle in cross-border services", Fibre Systems, Vol. 3, No. 1, pp. 43-46,
February 1999.
[2] K. Struyve, P. Arijs, N. Wauters, D.Colle, P. Demeester, P. Falcao, P. Lagasse, "Application, design
and evolution of WDM in GTS's pan-European tranport network", IEEE Communications Magazine,
Vol. 38, No.3, pp. 114-121, March 2000.
[3] P. Falcao, "Pan-European multi-wavelength transport networks", Proceedings of DRCN'98, Paper P3,
Brugge (Belgium), May 1998.
[4] T.-H. Wu, D.J. Kolar, R.H. Cardwell, "Survivable network architectures for broadband fiber optic
networks: model en performance comparison", Journal of Lightwave Technology, Vol. 6, No. 11, pp.
1698-1709, November 1988.
[5] R.D. Doverspike, J.A. Morgan, W. Leland, "Network design sensitivity studies for use of digital cross-
connect systems in survivable network architectures", IEEE Journal on Selected Areas in
Communications, Vol. 12, No. 1, pp. 69-78, January 1994.
[6] C.-C. Shyur, S.-H. Tsao, Y.-M. Wu, "Survivable network planning methods and tools in Taiwan", IEEE
Communications Magazine, Vol. 33, No. 9, pp. 100-107, September 1995.
[7] D. Johnson. P. Veitch, N. Hayman, "Core transport network redesign", Proceedings of DRCN'98, Paper
O2, Brugge (Belgium), May 1998.
[8] M. Bettin, G. Ferraris, G. Pignari, "Comparison of protection and restoration schemes for SDH
networks", Proceedings of DRCN'98, Paper O1, Brugge (Belgium), May 1998.
[9] J.D. Allen, S. Nathan, J. Huang, "Rings in a highly-connected network - an economic comparison",
Proceedings of DRCN'98, Paper O43, Brugge (Belgium), May 1998.
[10] B. Doshi, P. Harshavardana, "Broadband network infrastructure of the future: roles of network design
tools in technological deployment strategies, IEEE Communications Magazine, Vol. 36, No. 5, pp. 60-
71, May 1998.
[11] P. Arijs et al., "SDH protection in long-distance networks: A practical case study", Proceedings of
DRCN, Paper O3, Brugge (Belgium), May 1998.
[12] J.W. Suurballe, R.E. Tarjan, "A quick method for finding shortest pairs of disjoint paths", Networks,
Vol. 14, pp. 325-336, 1984.
[13] P. Arijs, M. Claeys, P. Demeester, "The design of SDH ring networks using tabu search and simulated
annealing", Proceedings of 5th International Conference on Telecommunications Systems: Modeling
and Analysis, pp. 287-294, Nashville (TN), March 1997.
[14] P. Arijs, "Development of algorithms for optimal ring selection within an SDH network topology”, M.
Sc. Thesis (in Dutch), Ghent University, 1995-1996.
[15] S. Cosares, I. Saniee, "An optimization problem related to balancing loads on SONET rings",
Telecommunications Systems, Vol. 3, No. 2, pp. 165-181, November 1994.
CHAPTER 4130
CHAPTER 5
Dimensioning and configuration of a
WDM ring
5.1 Introduction
In the previous chapter we have discussed architectural design options for an SDH-
based WDM network. We assumed WDM was merely used in point-to-point configurations
to economically upgrade the transmission capacity of transmission links. All networking
functionalities, such as routing the traffic, recovery from failures and management of the
network, were performed in the SDH layer. However as WDM technology is evolving and
traffic keeps increasing, definite opportunities exist to process part of the traffic in the nodes
at the optical level. Such optical processing lowers the amount of optical-to-electrical
conversions, as well as the required amount of electrical processing capacity and thus
increases the throughput of the network (for a same amount of electrical equipment). Since
optical processing is expected to become much cheaper than electrical processing, optical
networking provides many opportunities for decreasing the per unit transmission cost. A first
step towards optical processing is already witnessed through the use of fixed optical add/drop
multiplexers (OADMs), which are available today from some vendors. Such OADMs allow to
add/drop a fixed subset of the optical channels, while transparently passing through the
remainder of the channels transiting a node. The real value proposition for OADMs will be
met once they can be completely reconfigured to add/drop between 0 and 100% of the traffic,
and when they are able to support ring-based protection mechanisms as explained in section
3.6.3 of Chapter 3. Indeed, the throughput growth possible by employing WDM technology
will also increase the impact of physical failures, since much more traffic will be affected
simultaneously. As connections in the optical layer (also called lightpaths) may support
higher client layer links between nodes far from each other, one single fiber potentially carries
many different higher layer links. Optical recovery schemes treat multiple affected logical
links at once and entail less coordination than higher layer recovery (enabling faster recovery
times). Providing survivability at the WDM layers thus becomes inherently attractive as the
network throughput increases [1][2]. Indeed, besides the fact that optical recovery of physical
failures is simpler and more effective than when it would be done by an equivalent higher
layer scheme, WDM survivability may also turn out to be cheaper, as discussed in section
3.6.6 of Chapter 3.
The evolution from current SDH-based transport networks relying on point-to-point
WDM systems to all-optical networks will not occur following a ‘big bang’ scenario, but it
will be a smooth transitional process. Initially, WDM rings using fixed OADMs might be
configured such to support SDH rings. A next phase might be witnessed through the
CHAPTER 5132
introduction of WDM rings, protecting optical channels, in leopard spots of the network with
high traffic requirements. The introduction of a WDM ring, poses specific design and
configuration problems to be solved, some of which are quite similar to current SDH rings,
others which are specific to WDM rings. In this chapter, we describe the different design and
configuration problems related to a WDM ring. We only focus on design and configuration
problems related to a single topological ring (potentially consisting of multiple stacked rings).
The specific design problems related to deploying a network based on interconnected WDM
rings are discussed in Chapter 6.
The remainder of this chapter is structured as follows. First we will describe the
different ring planning problems. Afterwards, a solution method for each problem will be
proposed, and some results will be presented. Finally, the most relevant conclusions will be
summarized. The main results of this chapter can also be found in [3][4][5][6][7].
5.2 WDM ring planning
5.2.1 Long term planning
5.2.1.1 Determination of the architecture
The major decision in the long-term is to determine which type of WDM ring to use:
dedicated protection rings (DPRing) or shared protection rings (SPRing). This decision may
be distinct for different parts of the network and may even include a combination of both
rings (e.g., in a stack of rings). The choice of ring type is important, since it results in the
adoption of a certain type of OADM for a whole lot of traffic. It may be impossible to change
the protection scheme at a later stage: for instance, an OADM whose internal design is
optimized for OCh-DPRing operation may be incapable of effectuating OMS-SPRing
protection. Also, changing the protection configuration may involve disruption of live traffic,
which is highly undesirable for the customers and thus also the network operator.
The decision on the ring type depends amongst others on the following factors:
• The maturity of either solution in the time frame for deployment of WDM rings.
• A DPRing is the preferred ring type if high flexibility is required, because the required
ring capacity only depends on the total demand quantity. In contrast, when the SPRing
is optimized for a certain demand pattern, it may not be capable of supporting a
different demand pattern. For instance, this could be the case when the ring requires
dynamic reconfiguration in order to accommodate different demand patterns over time.
• For the same reason, DPRings are more robust when it is hard to predict the dispersion
of future transmission demands.
• The SPRing shares each protection channel amongst different lightpaths that are setup
in the corresponding working channel. The potential for wavelength sharing and the
corresponding capacity advantage gets larger when the ring collects more nodes and
with adjacent and uniform demand patterns (see section 5.3.2).
• If it is critical to maintain an efficient usage of the base infrastructure - such as fibers
and optical amplifiers - the SPRing is the preferable ring type since it generally requires
less wavelengths. For the same reason, SPRings induce the lowest overall transmission
cost in long-distance networks, where the cost of fibers and optical amplifiers is
DIMENSIONING AND CONFIGURATION OF A WDM RING 133
dominant. On the other hand, on shorter rings (e.g., metropolitan) where the OADMs
themselves form the major share of the transmission cost, the node cost will have a
more decisive impact. The impact of these costs is further discussed in section 5.2.2.3.
5.2.1.2 Determination of the ring topology
Another crucial long-term planning task is to decide on the topology of the rings.
While rings are conceptually easy, building a network based on a set of interconnected rings
is far from a straightforward task. The problem consists of finding a minimum cost
combination of interconnected rings, such that all demands in the network can be routed. The
main tasks in this design process consist of finding communities of interest between which to
deploy the rings. Such communities of interest are clusters of nodes that have considerable
amounts of traffic between them and are closely geographically affined. While we do not
tackle the topological dimensioning of WDM rings in this chapter (we do so in Chapter 6), we
provide some algorithms and guidelines for solving problems that are typically encountered
during the course of this topological optimization process (such as estimating the amount of
rings required in a community of interest or optimal routing of intra-ring traffic).
5.2.2 WDM ring dimensioning
Once the architecture and the topology of the rings have been chosen, it has to be
determined how many transmission resources are required to accommodate the upcoming
transmission demands over time. Some transmission resources have to be planned well in
advance before they can actually be commissioned to carry traffic. For instance, the
installation of optical fibers and amplifiers may have to be planned about 2 to 3 years in
advance. Other resources can be installed in a shorter timeframe. For example, some modular
components in the OADM can quickly be installed on a wavelength-by-wavelength basis.
Nevertheless, in the mid-term, the needed amount of wavelengths and consequent equipment
– to carry future traffic – must already be estimated.
In the short term, the most important planning task is to configure the ring: that is,
every time a new optical demand arrives, lightpaths have to be configured on the ring to
accommodate the demand. For this purpose, we need configuration routines that are capable
of optimizing the use of the resources. These routines should ensure optimal flexibility and
maximum remaining capacity at any time, in order to accommodate future demand growth.
Besides, in the mid-term dimensioning phases the assumption is often made that the short-
term ring configuration will make optimal use of provisioned capacity. Ring dimensioning
and configuration both require to address the following problems:
• Ring loading: “which way will we route our traffic?”
• Wavelength assignment: “which particular wavelength will we use for each path?”
These two problems will be highlighted in the following sections.
5.2.2.1 Ring loading
The objective of ring loading is to accommodate all optical demands between the
different nodes on the ring while using as few wavelengths as possible. The routing of bi-
directional protected traffic on a DPRing is trivial, since the associated protection mechanism
requires one dedicated unit of capacity on every section (i.e., link) of the ring for each unit of
CHAPTER 5134
demand. The utilization of every section (and thus the required ring capacity) is equal to the
total demand quantity, as shown in Figure 1 for a DPRing with 3 demands.
Figure 1: Loading on a DPRing
On a SPRing, each demand can be routed either clockwise or counter-clockwise in
the working channels of the ring. No explicit protection paths are set up. For calculating the
required number of wavelengths on a SPRing, the link of the ring (i.e., the optical multiplex
section) carrying the most working traffic is determinative (see Figure 2). Hence, the routing
strategy that minimizes the utilization of the most loaded link results in a minimal number of
required wavelengths. Since half of the wavelengths are required as protection channels, the
amount of wavelengths required on a SPRing is twice the utilization of the most loaded link.
It is important to note that routing the demand along the shortest side of the ring (e.g., shortest
in number of hops) does not necessarily lead to the optimal loading.
most loaded
link
Figure 2: Loading on a SPRing
The ring loading problem has been discussed before in the literature in the context of
SDH/SONET ring planning. Two variants of this optimization problem exist:
• A binary variant where it is not permitted to split a transmission demand on the ring
(i.e., each demand between two nodes has to be routed either clockwise or counter-
clockwise in its entirety). This variant of the problem has been extensively studied for
SDH/SONET rings, for which mathematical models and solution methods are described
in [8][9][10][11][12]. It has been shown in [8] that this optimization problem is NP-
complete.
• An integer variant where each transmission demand can be split in integer parts which
are routed in different directions (i.e., one part routed clockwise and the other part
routed counter-clockwise). It has been shown in [9] that this optimization problem can
be solved in polynomial time.
DIMENSIONING AND CONFIGURATION OF A WDM RING 135
In a WDM ring, there is little reason why all lightpaths supporting a multi-
wavelength node-to-node demand should follow the same route. The total demand between
two nodes can be split up in two parts (as long as each part remains an integer number of
wavelengths) and both parts can be routed along different sides of the ring. In section 5.3, we
therefore focus on the integer version of the ring loading problem. We provide theoretical
bounds on the wavelength requirements for certain demand patterns that are characteristic for
a ring topology. In addition, we describe a deterministic optimization method for the ring
loading problem for more general demand patterns. Furthermore, we examine the wavelength
requirements for SPRings compared to DPRings for a number of different scenarios.
5.2.2.2 Wavelength assignment
The ring loading problem does not consider to which particular wavelength each
lightpath should be assigned on each section of the ring. When the nodes are able to perform
wavelength conversion (see Chapter 3), the wavelength assignment problem is trivial to solve,
since we can change the wavelength of each lightpath in intermediate nodes to avoid
wavelength conflicts. When wavelength conversion is unavailable or undesirable, each
lightpath has to be assigned a unique wavelength on all the sections it crosses. In case of
DPRings, the wavelength assignment problem is straightforward to solve, since working and
backup path ‘use up’ the same wavelength along the ring. Consequently, it is impossible to
share a wavelength amongst multiple demands (and there is no potential for wavelength
conflicts). In a SPRing, wavelength conflicts can occur if no wavelength conversion is
available, and to resolve these conflicts some additional wavelengths might be required
compared to the case with wavelength conversion. Consider for example the 6-node ring in
Figure 3 that needs to accommodate three lightpaths between diametrically opposite nodes.
Without wavelength conversion, the ring needs to be equipped with at least three wavelengths
(no matter how the paths are routed). Wavelength conversion in intermediate nodes can
reduce the requirement to two wavelengths (e.g., with the loading of the figure).
Figure 3: Wavelength assignment on a 6-node ring
Having the possibility to interchange the wavelength of optical paths in intermediate
nodes provides the following advantages:
• The wavelength assignment problem is easier to solve.
• It may result in a higher utilization of the ring capacity.
• The ring is more flexible and scalable.
On the other hand, wavelength conversion in rings may be inappropriate because:
CHAPTER 513
6
• It increases the complexity and the cost of the OADMs because wavelength conversion
requires more expensive optical components, and the OADMs require a higher internal
connectivity.
• It may affect the transparency of the network (some wavelength conversion techniques
do not work all-optically).
• It may introduce single points of failures, since a failure of an intermediate OADM
converting the wavelength of a particular optical path that passes through this OADM
cannot be resolved with SPRing protection.
Furthermore, it has been shown in [13][14] that wavelength conversion in general
does not dramatically improve the utilization of WDM networks.
Optimal wavelength assignment algorithms for SPRings have been devised for
certain specific demand patterns (e.g., uniform pattern) as in [15][16][17]. For more general
demand patterns heuristics have been developed [18][19]. In section 5.4, we present an
optimization method that is not restricted to specific demand patterns and allows to obtain
optimal results. We consider the wavelength assignment problem separated from the routing
problem (i.e., we start from a given routing), in contrast to [20], where both problems are
solved simultaneously. We will show that in practice, the outcome of the ring loading
problem hardly poses any wavelength conflicts (that is, the routes and number of wavelengths
resulting from the loading problem generally allow a unique wavelength assignment), such
that it can be justified to solve both problems separately. At the same time, we show that
wavelength conversion does usually not reduce the wavelength requirement on optical rings.
5.2.2.3 OADM cost influence
The previous ring planning problems focus on the optimization of the wavelength
requirement of optical rings. Optimizing the wavelength requirement relates to minimizing
the cost of cable ducts, fibers, optical amplifiers, etc. However, the total transmission cost also
depends on the cost of the OADM, which itself depends a lot on the complexity of the
protection protocol. Since the SPRing protocol is more complex than DPRing operation (for
instance, the OCh-DPRing does not require any co-ordination nor signaling between
switching nodes), the implementation and operational costs of OADMs featuring SPRing are
potentially larger [21]. As such, a SPRing has a higher per wavelength cost than a DPRing.
The better wavelength utilization of a SPRing needs therefore to be traded off with its higher
cost.
In section 5.5, we compare the cost of a DPRing solution (= amount of DPRings
required x cost of DPRing) with that of a SPRing solution (= amount of SPRings required x
cost of SPRing) for stacked WDM rings, considering several cost scenarios. When the
wavelength requirement is higher than the capacity of a single WDM system, multiple WDM
rings have to be stacked on top of each other (see Figure 4). The amount of rings required in
the stack is determined by the amount of wavelengths on each single WDM ring system (e.g.,
32 wavelengths). The type of rings to be used depends on the demand pattern and relative cost
difference between both ring types. While the wavelength requirement of DPRings is
typically higher than that of SPRings, but the OADMs of a DPRing are likely to be cheaper
(for a same channel count), the DPRing might still be the preferred solution in certain cases.
DIMENSIONING AND CONFIGURATION OF A WDM RING 137
Cable topology
Fiber
topology
OADM
Building
Figure 4: Three stacked WDM rings
Finally, when deploying a stack of optical rings, we are not confined to using only
one type of ring in the stack. Indeed, a stack can contain both DPRings and SPRings. The
advantage of such a hybrid solution is that each ring type can support the demand pattern for
which it is best suited [6]. This allows to obtain even more cost-efficient solutions than the
pure DPRing and SPRing solutions. In section 5.5 we present a model, capable of optimizing
such hybrid DPRing and SPRing architectures. We present results for different demand
patterns under different cost scenarios, and thereby assess the benefits of both ring types
under these conditions.
5.2.3 SDH-on-WDM ring dimensioning
WDM rings can also be configured such to support SDH rings. In this case,
wavelengths around the ring act as 'virtual fibers' for an SDH ring (see Figure 5). These
wavelengths are dropped at every node, and do not need to be protected (because protection
switching occurs in the SDH layer). This is a near-term solution, which does not require fully-
fledged OADMs. This can thus be a viable architecture for fixed OADMs that do not support
optical protection switching yet. Alternatively, this can be realized without OADMs, using
point-to-point WDM systems between adjacent nodes on the ring. Another reason to opt for
this architecture might be because the traffic requirements on the ring are not high enough to
set-up dedicated wavelengths between the nodes of the ring. In this case, the traffic matrix
between the nodes of the ring is expressed in the SDH layer, e.g. as the amount of VC-4s to
be transported between each two nodes.
CHAPTER 5138
SDHSDH SDH
SDHSDH SDH
OADM
Figure 5: SDH-on-WDM ring
When deploying SDH-on-WDM rings, several design decisions have to be taken.
Dimensioning of SDH rings, involves similar issues as for WDM rings (i.e. ring loading and
timeslot assignment). In addition, for the stacked SDH rings (on the different wavelengths), a
decision must be made on which ring(s) each connection is routed. By optimally assigning
connections to the appropriate stacked ring(s), the amount of required SDH ADMs can be
substantially lowered. Indeed, the configuration as shown in Figure 5, can be wasteful in
terms of required ADMs, because each SDH ring requires an ADM at every node. By
installing SDH ADMs only at these nodes on the ring where traffic needs to be terminated,
some ADMs on some stacked rings can be omitted and replaced by a pass-through
wavelength in the OADM (or between 2 WDM multiplexers in the node). This reflects the
main application area of an OADM: providing access to a selected set of wavelengths while
transparently passing through the other wavelengths. In case all wavelengths would have to be
accessed, it is better to use WDM terminal multiplexers.
Two ways exist to minimize the amount of required SDH ADMs in stacked rings. A
first option is to groom SDH traffic between the end-nodes in the appropriate SDH ring, such
that traffic passed through a certain node is transported as much as possible on the same ring,
and the ADM in that node can be avoided and replaced by a pass-through wavelength. In this
scheme, no traffic is exchanged between different rings in the stack. This problem has already
been solved for dedicated protection rings [22][23][24][25] or shared protection rings with a
uniform traffic pattern [26] or for more general traffic patterns using heuristics [27]. In section
5.6, we present an integer linear programming model for optimally solving this problem for
shared protection rings with general demand patterns.
A second way of minimizing the amount of required ADMs also grooms traffic in
the appropriate SDH ring, but in addition allows to cross-connect traffic between different
rings of the stack at so-called hub nodes [28]. While this may result in additional ADM
savings over the previous approach, it may also incur additional stacked rings and expensive
cross-connects. Hence, this approach is only applicable in case the SDH ring cost (i.e. cost per
wavelength on the WDM ring) is low compared to the cost of the SDH ADMs and in case the
cross-connecting cost is low (e.g. if it is performed hard-wired). Such cross-connecting also
results in a lower reliability, because a single point of failure is introduced (which might be
avoided by using drop & continue, but this again incurs an additional cost). In addition it has
been shown in [29] that cross-connecting between rings is less useful if a demand is allowed
to be split in integer parts, and each part may be routed on a different ring in the stack. For
these reasons we do not consider cross-connecting between stacked rings in this chapter.
DIMENSIONING AND CONFIGURATION OF A WDM RING 139
5.3 Ring loading
5.3.1 Problem formulation
We consider a single SPRing containing a set of nodes N labeled 1,2,…,n (following
a clockwise proceeding). The links of the ring are labeled accordingly: a link between node i
and i+1 is labeled i. These notations are also shown in Figure 6.
1 2 ...
nn-1 ...
i
i+1
12
i
n-1
n
Figure 6: Node and link notations on the ring
The ring is to be used to accommodate a set K of node-to-node demands. Every
demand k in the set K comes with a pair of distinct endpoints s(k) and t(k) and a quantity of
d(k) (expressed as an equivalent amount of wavelengths). The demands are assumed to be bi-
directional. Without loss of generality, we can assume that s(k)<t(k) for all k. In other words,
to satisfy each demand k, a total amount of d(k) bi-directional optical paths are to be routed
between s(k) and t(k). Each of these paths is routed either clockwise or counter-clockwise. We
denote the amount of clockwise paths related to demand k as x(k). Conversely, the amount of
paths routed counter-clockwise is d(k)=− x(k). If we define z as the decision variable
representing the working ring capacity (i.e., amount of wavelengths installed on every
working section), the problem can be formally stated as follows.
Minimize: z
Subject to:
,)]()([)(
)()()()(
zkxkdkx
iktiks
Kk
ktiks
Kk
≤−+
≤∨>
∈
<≤
∈
∀=i∈N
0 ≤ x(k) ≤ d(k)∀=k∈K
x(k) integer ∀=k∈K
The first constraints states that on each link i (between nodes i and i+1), less traffic
must be routed, than the total ring capacity. The amount of traffic routed on link i is calculated
by adding the clockwise and counter-clockwise routed traffic on this link. The other two
constraints are trivial.
CHAPTER 5140
5.3.2 Bounds on the wavelength requirements
In this section, we will present some theoretical bounds on the wavelength
requirement of the ring loading problem. Although these bounds do not tell us how to route
(and split) the individual demands to obtain an optimal ring loading, they are very useful for
ring planning. First of all, they can be used to speed up the optimization process drastically.
Second, in the mid-term planning phase, the individual node-to-node demands may not be
forecasted with enough accuracy to use in a deterministic optimization model. However, often
some demand characteristics are known (such as the demand pattern, the total volume of
demand, or the most dominant demands on the rings) from which theoretical bounds on the
ring capacity can be derived without knowledge of the exact demand dispersion. As such, the
bounds can also be used when only an estimate on the required ring capacity is required, e.g.
during the determination of the ring topology. In those cases, a rough ring dimensioning
process can be carried out based upon these bounds.
We will apply the bounds on a number of characteristic demand patterns. We define
TD as the total quantity of demand to be accommodated:
∈
=
Kk
kdTD )( . In the following, we
will compare the value of the bound for the SPRing with TD (which also represents the
wavelength requirement of the DPRing to accommodate the same demand set K) for several
demand patterns.
5.3.2.1 General demand pattern
The removal of any two links cuts the ring into two components (see Figure 7). Let
Kpq be the set of demands whose endpoints are separated by the removal of links (p,p+1) and
(q,q+1), with p < q. We define Dpq as the total quantity of the demands in Kpq.
≤<
≤
∈
>
≤<
∈
+=
qktp
pks
Kk
qkt
qksp
Kk
pq kdkdD
)(
)(
)(
)(
)()(
1 p p+1
nq+1 q
...
...
Dpq
Figure 7: Demand cut on a ring
Dpq represents the quantity of demand that needs to be routed across the cut
{(p,p+1),(q,q+1)}. As such, the total amount of working traffic on link (p,p+1) and link
(q,q+1) can not be lower than Dpq. We define the maximum demand cut as the pair of links
whose removal from the ring separates the largest demand quantity D*. Clearly, D* provides
a lower bound for the total required capacity (i.e. working and protection capacity) on the
ring. In fact, the optimal value of the linear relaxation of the ring loading problem (only
working capacity) is equal to D*/2 (see [11][20]). Because of the integer character of the
DIMENSIONING AND CONFIGURATION OF A WDM RING 141
capacity, we can state the following lower bound 2/*Dzopt ≥. Consequently, any integer
loading of the demands resulting in a maximum link utilization of 2/*
D is optimal.
In addition, Schrijver [11] determined an upper bound for ring loading based upon
the optimum solution of the linear relaxation. We refer to [11] for the details and the proof,
but in the case of SPRing where integer demand splitting is allowed, it follows from [11] that
theoretically 2/)3*( +≤ Dzopt . Based upon empiric observations, Schrijver even guarantees
that 12/* +≤ Dz opt . Concluding, the maximum demand cut value D* bounds the optimal
amount of the wavelengths required for working traffic on SPRings in the following way:
1
2
*
2
*+≤≤
ù
ê
ê
éD
z
Dopt
In the following paragraphs, we will translate this bound to some specific demand
patterns on a ring.
5.3.2.2 Hub demand pattern (single star)
A demand set is said to follow a hub pattern (or a single star pattern) when each
demand terminates in a single hub node (say node 1), that is s(k) = 1 , ∀ k∈K (we assume t(k)
> s(k), ∀ k∈K), as shown in Figure 8.
Figure 8: Hub demand pattern
It is easy to see that the maximum demand cut disconnects the hub node from the
other nodes (i.e., resulting in the partition })1{\,1( N) and its value is TD. We can thus easily
deduct that D* = TD. If the hub demand pattern is homogeneous, d(k) = Dh, ∀ k∈K, then
h
DnTDD ).1(* −==
The lower bound 2/TD is also the optimal value for z, because it can be achieved
with the following routing of demands. Find node j on the ring for which
>∈<∈
ù
ê
ê
é
≤
ù
ê
ê
é
≤
})(|{})(|{ 2
)(and
2
)(
jktKkjktKk
TD
kd
TD
kd
Consequently,
• Route all demands {k∈K | t(k) < j } in their entirety clockwise from 1 to t(k).
• Route all demands {k ∈K | t(k) > j } in their entirety counter-clockwise from 1 to t(k).
• Route
<∈
−
})(|{
)(2/
jktKk
kdTD units of demand between node 1 and node j clockwise and
route the remaining units of demand between node 1 and j counter-clockwise.
CHAPTER 5142
The total amount of (working and protection) wavelengths required by the SPRing is
thus 2/2 TD⋅, while in case of the DPRing, the amount of required wavelengths is TD.
Thus, when the demands follow a hub pattern, a SPRing does not provide any capacity
advantage compared to a DPRing. Also, the lower bound does not improve when the demands
terminate in multiple adjacent hubs, following a so-called multiple star pattern. In this pattern,
each demand k∈K terminates in one of h adjacent hub nodes, that is s(k) ∈ {1,…,h}, ∀k∈K
(assuming there is no demand between hub nodes).
5.3.2.3 Long demand pattern
A demand set is said to follow a long pattern when all demands are between nodes
that are located diametrically opposite to each other on the ring (see Figure 9).
Figure 9a: Long demand pattern (n even) Figure 9b: Long demand pattern (n odd)
In case n is even, the maximum demand cut (which is composed of 2 opposite links)
comprises all demands of the ring: D* = TD. In case n is odd, the maximum demand cut
comprises all demands except the smallest demand. If we define Dmin as the smallest demand
on the ring: D* = TD – Dmin. When the demands follow the long pattern, the SPRing does
hardly offer any capacity advantage versus the DPRing. In some cases the DPRing even has a
capacity advantage over the SPRing (if 12/* += Dz opt ).
5.3.2.4 Uniform homogeneous demand pattern
A demand set is said to follow a uniform pattern when there exists a demand for each
pair of nodes. With this demand pattern, the demand graph is fully meshed (see Figure 10).
The demand pattern is said to be homogeneous when d(k) = Du for each k∈K.
Figure 10: Uniform demand pattern
Each demand cut Kpq, separates the node set N into a partition (A,B), with:
{}
qipiNiA >∧≤∈= | and
{}
qipNiB ≤<∈= |. The demand across this demand cut
equals u
DBA ...
DIMENSIONING AND CONFIGURATION OF A WDM RING 143
It is easy to determine that the maximum demand cut is formed by the partition
which divides N in two parts that are each as large as possible. The number of demands cut by
such a partition is ú
ê
ë
ê
ú
ú
ù
ê
ê
é
=
ú
ú
ê
ë
ê
ú
ú
ù
ê
ê
é
2
.
22
||
.
2
|| nnNN
This gives us the capacity of the maximum demand cut as
ë
2/.2/* nnDD u
=. In
comparison with the DPRing - which requires TD = Du.n.(n-1)/2 wavelengths - the lower
bound on the wavelength requirement of the SPRing is almost two times better than the
wavelength requirement for the DPRing (assuming n >> 1).
5.3.2.5 Adjacent demand pattern
A demand set is said to follow an adjacent pattern when s(k) and t(k) are adjacent
nodes on the ring, for each k∈K (see Figure 11).
Figure 11: Adjacent demand pattern
With an adjacent demand pattern, each demand cut comprises only two individual
demands. The value of the maximum demand cut is equal to the sum of the largest and second
largest demand quantity. Let D' be the largest amongst all demand quantities, and D'' the
second largest, then:
)"()(max"and)(maxarg"
)'()(max'and)(maxarg'
}'\{
}'\{
kdkdDkdk
kdkdDkdk
kKk
kKk
Kk
Kk
===
===
∈
∈
∈
∈
Thus, the maximum demand cut separates a demand D* = D'+D'' and the lower
bound on the amount of working channels is ú
ú
ù
ê
ê
é+
≥2
"' DD
zopt .
The lower bound is also the optimum, since it can for instance be achieved as
follows. Route all demands except k' in their entirety along the shortest one-hop path. From
the largest demand k', route 2/*D units on the one-hop path between s(k') and t(k'). The
remaining
ë
2/)'''(2/*' DDDD −=− units are to be routed on the long path between s(k')
and t(k') over the ring.
Since D'+D'' ≤ TD, the wavelength requirement of an adjacent demand pattern is
never higher than a hub demand pattern with same TD value (and thus also never higher than
for the DPRing). If the adjacent demand pattern is homogeneous (that is, d(k) = Da, ∀ k ∈ K)
then:
a
DnTD .= and a
DDD == "'
Thus:
CHAPTER 5144
n
TD
DD a.2.2* ==
n
TD
Dz a
opt == .
A homogeneous adjacent demand pattern is the most efficient demand pattern for the
SPRing. The above formula shows that the wavelength requirement of the SPRing with n
nodes can be up to n/2 times better than that of the DPRing.
5.3.3 Comparison of demand patterns
Figure 12 shows the evolution of the relative value of the largest demand cut (and
thus the lower bound on the wavelength requirement) versus the total demand for the patterns
described previously. In other words, the value set out in the y-axis represents approximately
how many wavelengths are required per unit of optical demand carried on the SPRing.
0%
20%
40%
60%
80%
100%
120%
345678910111213141516
Amount of nodes on the ring
D*/TD
hub
long
uniform
adjacent
Figure 12: Evolution of D*/TD versus the ring size
The working channels on the different sections of a SPRing can be reused by
multiple non-overlapping lightpaths, and the potential for wavelength reuse increases with
ring size. When more demands share the same working channel, a SPRing requires fewer ring
capacity. This is particularly evident with the homogeneous adjacent pattern, where each
working channel is shared between n lightpaths. For DPRings, one wavelength is dedicated to
each unit of optical demand. Consequently, Figure 12 can also be used for a rough
comparison of the wavelength requirements of DPRings versus SPRings for different demand
patterns. As explained before, when demands follow a uniform, or adjacent pattern, the
SPRing definitely requires less wavelengths than the DPRing (up to n/2 times less). However,
when the demands follow a hub pattern or a long pattern, there is hardly any potential for a
better wavelength requirement with the SPRing as compared to the DPRing.
DIMENSIONING AND CONFIGURATION OF A WDM RING 145
5.3.4 Mathematical programming algorithm
In the previous paragraph, we presented theoretical bounds that enable to get a good
idea of the wavelength requirement of a SPRing. In addition, the routing of the demands to
obtain the optimal ring loading can be rather trivial for specific demand patterns. In the
general case, a more sophisticated approach is required to solve the loading problem. A
mathematical approach is to implement the model presented in section 5.3.1 in an integer
programming solver such as CPLEX [30]. Such an approach has also been followed in [12],
to compare results obtained by CPLEX and a genetic algorithm. It was concluded that the
genetic algorithm is much slower than CPLEX and does not always yields the optimal result.
In addition, it was stated the CPLEX approach can be speeded up by enhancing the
mathematical formulation of the ring loading problem.
To reduce the computational efforts, we can enhance the mathematical model in the
following ways. First of all, we can include the bounds mentioned in section 5.3.2.1 for
decision variable z. This is done by calculating all possible 2
n
C demand cuts to find the
maximum value D*. Secondly, we can add a lightpath routing cost to the objective function:
Minimize: z=+=)]()()].[()([)]()().[(. ksktnkxkdksktkxR
Kk
+−−+−
∈
The coefficient R represents the relative cost of routing one wavelength on one link
of the ring. Its value can influence the bias towards shortest path routing (e.g., because shorter
paths induce less delay and are more reliable). For large values of R, the routing cost is more
important than the wavelength requirement and the optimal loading routes all demands along
their shortest paths. Even when we do not want to take such a routing cost explicitly into
account, the incorporation of R into the optimization model can be beneficial to the decisions
made by the solver. In fact, the solution to the ring loading problem often has a lot of
equivalent routings (i.e., solutions with a same wavelength requirement but a different
routing). This increases the computational complexity (and the associated run-time). For this
purpose, R can be chosen very small but non-zero (that is, 0 < R << 1) to break the tie
amongst these equivalent solutions, by choosing the one with the lowest routing cost.
5.3.5 Comparison of DPRing and SPRing
We have solved the ring loading problem - as described above - with CPLEX for
random demand patterns and different ring sizes n (= amount of nodes on the ring). In Figure
13, we set out the relative number of wavelengths required for a SPRing as compared to the
total demand TD. Results were generated for 100 different demand patterns for each ring size.
The demand sets were generated with individual values for each demand drawn randomly out
of a uniform distribution between 0 and 8 wavelengths. For each sample, the optimal amount
of wavelengths required by a SPRing was calculated and the relative requirement compared to
TD was averaged over all results. For DPRings, the relative wavelength requirement is always
100%. Figure 13 sets out the average and standard deviation for the set of outcomes for each
particular ring size.
CHAPTER 514
6
0%
20%
40%
60%
80%
100%
4 6 8 10121416
Amount of nodes in the ring
Relative wavelength requirement
Figure 13: Number of wavelengths required for SPRing vs. DPRing
As can be seen from Figure 13, a SPRing can save between 20% and over 40% of the
wavelengths that would be required by a DPRing. The reason is again that a SPRing is able to
share wavelengths amongst multiple working paths on the ring. The maximum amount of
working paths that can share the same protection channel is proportional with the size of the
ring (i.e., number of links), thus the relative benefit of the SPRing (versus the DPRing)
increases as the ring size increases.
5.4 Wavelength assignment
5.4.1 Problem formulation
We consider a single SPRing with a set of nodes N numbered clockwise from 1 to n.
A set K of demands k, each with d(k) wavelengths, has been routed on the ring. The route of
each connection k is thus fixed and designated by r(k,i):
r(k,i) = 1, if the route of connection k includes the link i between node i and i+1
= 0, if not.
If a demand has been split (i.e. partly routed clockwise and partly routed counter-
clockwise), this can be represented by 2 separate demands between the same end nodes, but a
different routing, in the above formulation.
We consider a set of M wavelengths to be available for demands on the SPRing, and
denote the binary variable x(k,m) to designate whether one unit of demand k is assigned to
wavelength m. The binary variable y(m) indicates whether wavelength m has been deployed
on the ring. The wavelength assignment problem can then be stated formally as:
DIMENSIONING AND CONFIGURATION OF A WDM RING 14
7
Minimize:
m
my )(
Subject to:
=
m
kdmkx )(),( ∀=k∈K
≤
k
myikrmkx )(),().,( ∀=i∈N, ∀=m∈M
}1,0{),( ∈mkx ∀=k∈K, m∈M
}1,0{)( ∈my ∀=m∈M
The first constraint states that each demand should be fully wavelength-allocated.
The second constraint indicates that each wavelength can only be used once on each section
of the ring. The remaining constraints are trivial. A lower bound on the amount of required
wavelengths is the amount of wavelengths resulting from the solution of the ring loading
problem (i.e., the solution with wavelength conversion).
The problem can be simplified by searching for demands that have been split up in
their loading on the ring (that is, both sides of the ring carry parts of traffic between the same
end nodes). In our formulation, such a splitted demand is expressed as two demands. If two
demands k' and k" are routed between the same end-nodes but both along a different side of
the ring (i.e. r(k',i) ≠ r(k",i), ∀=i∈N) we can reduce both d(k') and d(k") with min{d(k'), d(k")},
which is the common amount of wavelengths routed both clock-wise and counter-clockwise.
Because this common amount of wavelengths occupies the entire circumference of the ring,
the wavelength assignment can be done in a straightforward way, by assigning min{d(k'),
d(k")} wavelengths to this common part. Afterwards, this common part can be removed from
the problem formulation. In the resulting formulation, maximum one of both demands k' and
k" appears, reduced with min{d(k'), d(k")}.
A computational inefficiency of the above formulation is that it can result in a large
number of solutions that are all permutations of each other. Indeed, in the model all
wavelengths are equivalent, and one can obtain an equivalent solution by swapping the
assignment to any two wavelengths (that is, the solutions x and x’ with x’(k,i) = x(k,j) and
x’(k,j) = x(k,i) are equivalent). To address this problem of permutable solutions, we can
already assign wavelengths to certain connections before starting the actual optimization.
Once all split demands have been removed, we fix the variables x(k,m) of the demands that
are routed on the most loaded link (l,l+1) of the ring. Indeed, demands routed over the same
link should all have a distinct wavelength, and by choosing the most loaded link, we can fix
the largest amount of variables.
Thus ∀=k∈K | r(k,l) = 1, we can fix:
x(k,m) = 1 m = 1 …
∈Kk
lkrkd ),().(.
By assigning fixed wavelengths to these demands in advance, the amount of possible
permutations can be limited, thereby enhancing the computational efficiency.
CHAPTER 5148
5.4.2 Results
The above formulation has been modeled into CPLEX to solve the wavelength
assignment problem. Using this model, we solved 200 problem instances for a set of
randomly generated demand sets (where the individual demand quantities varied between 0
and 4 wavelengths). We investigated rings with different number of nodes n. We used the
routing of the demands resulting from the optimal solution of the ring loading problem (see
section 5.3.4). The results in Table 1 indicate the number of extra wavelengths that are
required for the different problem instances to resolve all wavelength conflicts. This can also
be interpreted as the extra amount of wavelengths that are required to reduce the need for any
wavelength conversion on the ring.
N
Average
number of
wavelengths
Examples
with 0 extra
wavelengths
Examples
with 1 extra
wavelength
Examples
with 2 extra
wavelengths
Examples
with more
than 2 extra
wavelengths
4 5 100.0 % 0.0 % 0.0 % 0.0 %
6 11 98.0 % 2.0 % 0.0 % 0.0 %
8 19 98.0 % 1.5 % 0.5 % 0.0 %
10 29 99.0 % 1.0 % 0.0 % 0.0 %
12 40 99.0 % 0.5 % 0.5 % 0.0 %
14 53 99.5 % 0.5 % 0.0 % 0.0 %
16 69 100.0 % 0.0 % 0.0 % 0.0 %
Table 1: Extra wavelengths needed without wavelength conversion
The results show that in almost all cases the number of wavelengths required with
and without wavelength conversion is the same. In a limited number of cases, a small amount
of extra wavelengths might be required when no wavelength conversion is available. This
certainly does not justify the extra cost for providing wavelength conversion in all the nodes
of the ring. Alternatively, a limited amount of wavelength conversion facilities might be
established in the OADMs [31]. We have done similar simulations for demands varying
between 0 and 8 wavelengths, and the results were along the same line. These conclusions are
in line with [19], where a heuristic algorithm was used and the wavelength savings through
wavelength conversion were also found to be minimal for static traffic patterns. For dynamic
traffic however, wavelength conversion turned out to be more useful.
In principle, the wavelength assignment problem and the ring loading problem are
interrelated. In other words, rerouting certain lightpaths (in another way than they are routed
in the solution of the ring loading problem) may potentially save a few wavelengths because it
might lead to a better wavelength assignment. To obtain the absolute minimum amount of
wavelengths, both problems of routing and wavelength assignment have to be tackled in an
integrated model. However, this results in a problem that is computationally much harder to
address [20].
In contrast, we start by solving the simpler ring loading problem, where we obtain
the minimal wavelengths required (the wavelength requirement of the integrated problem
cannot be better than the optimum of the loading problem). By optimizing the wavelength
assignment for this optimal routing afterwards, we can usually conclude that the minimal
amount of wavelengths resulting from the ring loading problem, suffices for the wavelength
DIMENSIONING AND CONFIGURATION OF A WDM RING 149
assignment and thus no extra wavelengths are required. Our results thus show that in practice
it is justified to solve both problems of ring loading and wavelength assignment sequentially.
This decoupling of both problems hardly affects the quality of the solution compared to an
integrated solution approach.
5.5 Hybrid DPRing/SPRing architecture
5.5.1 Problem formulation
Consider a stack of rings (e.g., see Figure 4) to accommodate a certain demand
pattern. The hybrid DPRing/SPRing planning problem consists of choosing the right ring type
(shared or dedicated) for each ring in the stack and routing the traffic on these rings, such to
minimize the total installation cost. The problem can be stated in the same way as the ring
loading problem (we refer to section 5.3.1 for the notations). Two kind of decisions can be
made for each unit of demand: whether to accommodate it on a SPRing or on a DPRing. In
the case of the SPRing, each lightpath can be routed clockwise or counter-clockwise. The
variables x+(k) and x−(k) denote the amount of lightpaths of demand k which are routed
clockwise and counter-clockwise respectively on SPRings in the stack. Conversely, the
amount of lightpaths of k accommodated on DPRings in the stack is d(k) − x+(k) − x−(k).
We characterize SPRings by a cost CS and capacity WS (amount of wavelengths
available for working traffic). We characterize the DPRings by a cost CD and a capacity WD
(total amount of wavelengths). For the ring loading and wavelength assignment problem, the
objective was to minimize the amount of wavelengths on the ring. Now, we consider 2 ring
types, and the amount of wavelengths on a ring system is given. The objective of the hybrid
ring planning problem is to minimize the total installation cost of all the ring, while
accommodating all node-to-node demands. If we define y as the amount of DPRings and z as
the amount of SPRings to be optimized, the problem can be formulated as follows:
Minimize: CD . y + CS . z
Subject to:
zWkxkx S
iktiks
Kk
ktiks
Kk
.)()(
)()()()(
≤+
≤∨>
∈
−
<≤
∈
+ ∀=i∈N
yWkxkxkd D
Kk
.)]()()([ ≤−− −
∈
+
x+(k) ≥ 0∀=k∈K
x−(k) ≥ 0∀=k∈K
0 ≤ x+(k) + x−(k) ≤ d(k)∀=k∈K
x+(k), x−(k), y, z integer ∀=k∈K
The first constraint is similar to the first constraint of the ring loading problem. It
states that the demands routed on each link of the SPRings should not exceed the total amount
of capacity available for the SPRings. No wavelength assignment is performed. The second
CHAPTER 5150
constraint states that enough DPRings should be available to route all DPRing traffic. Since
DPRing traffic is routed on all links of the ring, there is only one constraint for the entire ring.
The other constraints are trivial. The above formulation can be enhanced by adding bounds on
the amount of DPRings and SPRings and a wavelength routing cost (as in section 5.3.4).
5.5.2 Results
We have compared 3 different stacked ring architectures: a design based solely on
DPRings, a design based solely on SPRings (which involves solving the ring loading
problem) and the above described hybrid design (using both SPRings and DPRings). The
SPRing and hybrid design problems were solved by implementing the developed models in
the CPLEX solver. In the remainder of this section, we present results for different demand
patterns and various cost scenarios.
In order to make meaningful comparisons, we assumed that WD = 2.WS (which
means the total amount of wavelengths on both ring types is the same). In the different cost
scenarios, we assumed that the SPRing was never cheaper than the DPRing with the same
amount of wavelengths. In section 3.3, we have argued that this is indeed a realistic
assumption. In fact, it is only under this assumption that the optimization makes sense: since
in general, SPRings make better use of the available wavelengths on the ring, it would always
be the preferred solution if a SPRing were to be cheaper than a DPRing (with the same
amount of wavelengths).
A first set of results was obtained for random demands. In Figure 14 and Figure 15,
we present the average results for 200 demand patterns on a 10-node ring. We assumed that
each ring is equipped with 16 wavelengths. The node-to-node demand was drawn randomly
from a uniform distribution between 0 and 5 wavelengths. In Figure 14, we represent the
average relative extra cost of the pure DPRing design and the pure SPRing design, compared
to the optimized hybrid ring design. In Figure 15 the relative amount of the different ring
types in the hybrid design is presented. We present results for different cost scenarios, with
the cost of a DPRing ranging from equally expensive as a SPRing down to half the price of a
SPRing (with the same amount of wavelengths).
DIMENSIONING AND CONFIGURATION OF A WDM RING 151
0.0%
10.0%
20.0%
30.0%
40.0%
50.0%
60.0%
100% 90% 80% 70% 60% 50%
Cost DPRing vs. SPRing
Average extra cost compared to hybrid design
Only SPRing
Only DPRing
Figure 14: Relative cost comparison
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
100% 90% 80% 70% 60% 50%
Cost DPring vs. SPRing
Amount of
DPRings
Amount of
SPRings
Figure 15: Contribution of DPRings and SPRings in hybrid design
When the cost of both ring types is equal (100% in Figure 14), the optimal hybrid
design results in a pure SPRing based solution (i.e., all rings in the stack are of the SPRing
type as can be seen in Figure 15). However, when the DPRing cost decreases (even slightly)
compared to the SPRing, the hybrid design optimum includes both SPRings and DPRings.
The cost of the hybrid design can be substantially lower than the single-ring-type designs.
Also the amount of DPRings in the optimized hybrid design can be substantial, even if the
cost of a DPRing is only slightly lower than that of a SPRing. E.g. when the cost of the
CHAPTER 5152
DPRing is 10% lower than that of the SPRing, the hybrid network design already contains
more DPRings than SPRings, and this hybrid design is on average 6.4% cheaper than the pure
SPRing design. As the cost of a DPRing further decreases, the hybrid design contains an
increased amount of DPRings (see Figure 15). This is also reflected in Figure 14 as the cost of
the pure DPRing design option decreases and the cost of the pure SPRing design option
increases compared to the hybrid design option. We noticed in this case study that when the
cost of a DPRing is about 60-70% of the cost of a SPRing, a pure DPRing based design
becomes more attractive than a pure SPRing based design. Still, both designs are more than
20% more expensive than the hybrid design option. On the other hand, even for a low cost of
the DPRing compared to the SPRing, there are still some of these SPRings present in the
optimal hybrid design. Even though the SPRings are far more expensive than the DPRing in
this case, they still help to lower the cost of the hybrid design compared to a pure DPRing
based design, e.g. for accommodating the adjacent node demands.
A second set of results was obtained for different special demand patterns on a 12-
node ring. The quantities of the demands were chosen in such a way that the total demand was
the same for each demand pattern. Hereby, each demand pattern required the same amount of
DPRings, which enables to make meaningful comparisons. The used values for the demands
on a 12-node ring are shown in Table 2.
Amount of
demand pairs
Quantity of each
individual demand
Total demand
quantity
Adjacent demand 12 11 132
Uniform demand 66 2 132
Hub demand 11 12 132
Long demand 6 22 132
Table 2: Amount of demand pairs and wavelengths per demand pair on 12-node ring
A summary of the results for each design option is shown in Table 3. We
investigated different scenarios, again with the cost of a DPRing ranging from 100% to 50%
of the cost of a SPRing. The values in the table represent the relative cost compared to the
cost of the DPRing solution with the cost ratio of a DPRing being 100% of a SPRing. Notice
that for the hub and long demand pattern the bandwidth requirements of the DPRing and the
SPRing are equal. This means that the DPRing based design is the optimal solution even
when the cost of a DPRing is the same as that of a SPRing and certainly when the DPRing is
cheaper. For the adjacent node demand pattern, the bandwidth usage of the SPRing is much
better, such that at equal costs the SPRing based design (which is equal to the hybrid solution)
is certainly the preferred solution. For the uniform demand pattern, the SPRing based design
is also the preferred solution in the equal cost scenario.
DIMENSIONING AND CONFIGURATION OF A WDM RING 153
Cost ratio DPRing 50% 60% 70% 80% 90% 100%
Hub + Long pattern
SPRing only 100% 100% 100% 100% 100% 100%
DPRing only 50% 60% 70% 80% 90% 100%
Hybrid 50% 60% 70% 80% 90% 100%
Uniform pattern
SPRing only 56% 56% 56% 56% 56% 56%
DPRing only 50% 60% 70% 80% 90% 100%
Hybrid 39% 42% 46% 49% 52% 56%
Adjacent pattern
SPRing only 22% 22% 22% 22% 22% 22%
DPRing only 50% 60% 70% 80% 90% 100%
Hybrid 22% 22% 22% 22% 22% 22%
Table 3: Comparison of design options for different costs and demand patterns
For the above described special demand patterns, the hybrid solution offers no cost
reductions compared to the best of both DPRing and SPRing solutions. However, for the
uniform demand pattern, the optimal design is the hybrid design when the cost of the DPRing
is (even slightly) lower than that of the SPRing, because some SPRings can be replaced by
cheaper DPRings. As the cost of the DPRing further decreases, the amount of DPRings in the
hybrid design also increases and the total cost will decrease. When the cost of the DPRing
drops to about 50-60% of the SPRing, the full DPRing solution becomes the preferred
solution over the full SPRing solution for this uniform demand pattern, while the hybrid
design option is still the best overall solution. This again proves that the hybrid network
design results in the lowest overall cost and that even at a higher cost per wavelength unit, the
SPRing can still help to lower the total network cost. This is due to the fact that the SPRing is
particularly advantageous for adjacent demand patterns. Indeed, even if the SPRing is twice as
expensive as the DPRing, the SPRing based solution is the preferred solution for the adjacent
demand pattern. The cost of the DPRing would have to decrease to 20% of the cost of the
SPRing (which is very unlikely in practice) before the DPRing based solution would be
preferred for the adjacent demand pattern in this case study.
From the above, we can conclude that SPRings provide a cost-efficient means for
protecting the optical layer, even if their unit-cost per wavelength might be higher than that of
the DPRing. Cheap DPRings can further help in lowering the cost of hybrid ring design
solutions to accommodate selective demands that are not suited for wavelength reuse
(demands following a hub or long pattern).
5.5.3 Extension for rings with drop & continue
In the previous section, we have shown the suitability of a hybrid model, as soon as
the DPRing is slightly cheaper than the SPRing. In this section we will show that a hybrid
model can be more cost effective in cases that the amount of inter-ring traffic is substantial,
and needs to be protected using drop & continue.
The model from section 5.5.1 can be extended for drop & continue traffic in the
following manner. We assume that the gateway nodes that provide inter-ring connectivity are
the adjacent nodes n and n-1. We consider a set K of intra-ring demands k, and a set L of
inter-ring demands l. While the intra-ring demands have a source and termination node s(k)
CHAPTER 5154
and t(k) on the ring, the inter-ring demands can be specified by a source node s(l) only. The
termination nodes of the inter-ring demands on the ring are always n-1 and n. From Figure 16,
it can be seen that on a shared protection ring, both clockwise and counter-clockwise inter-
ring connections make use of the section between both gateway nodes. This is in contrast to
dedicated protection rings, in which a connection uses up all sections on the ring (thus also
between both gateway nodes), independent whether drop & continue is used or not (see
Chapter 3). In the specific case of Figure 16, the shared protection ring requires 4 wavelengths
(2 working + 2 protection) to accommodate both inter-ring demands, while the same demand
pattern could be supported on a dedicated protection ring using only 2 wavelengths. This
already shows the benefit of dedicated protection rings for inter-ring traffic protected by drop
& continue.
1 3
n
2
n-1
Figure 16: Drop & continue on shared protection ring
The formulation of the hybrid ring design problem now becomes:
Minimize: CD . y + CS . z
Subject to:
zWlxkxlxkx S
iktiks
Kk
niils
Ll
nils
Ll
ktiks
Kk
.)()()()(
)()(1)()()()(
≤+++
≤∨>
∈
−≥∨>
∈
−−
<≤
∈
+
<≤
∈
+ ∀=i∈N
yWlxlxldkxkxkd D
LlKk
.)]()()([)]()()([ ≤−−+−− −
∈
+−
∈
+
x+(k), x−(k), x+(l) and x−(l) ≥ 0∀=k∈K, ∀=l∈L
0 ≤ x+(k) + x−(k) ≤ d(k)∀=k∈K
0 ≤ x+(l) + x−(l) ≤ d(l) ∀=l∈L
x+(k), x−(k), x+(l), x−(l), y, z integer ∀=k∈K, ∀=l∈L
The first constraint again limits the demands routed on each link of the SPRings. The
first term represents the clockwise intra-ring demand, and the third term the counter-
clockwise intra-ring demand on the link. This is similar to the model in section 5.5.1. The
second term represents the clockwise routed inter-ring demand, and the fourth term the
counter-clockwise inter-ring demand on the link. The link n-1 between nodes n-1 and n,
contains all (clockwise and counter-clockwise routed) inter-ring demands. The second
DIMENSIONING AND CONFIGURATION OF A WDM RING 155
constraint limits the demands routed on the DPRings. The remaining constraints are again
trivial.
Again, the model was implemented in CPLEX, and the solution of the hybrid design
has been compared with the pure DPRing and SPRing solution. We again assumed 16
wavelength rings (i.e. WD = 16 and WS = 8) and now assumed that the cost of a DPRing was
equal to a SPRing. We present the average results over 100 demand patterns for a 10-node
ring with inter-ring and intra-ring demands chosen randomly from different uniform
distributions. The distributions were chosen such that the amount of inter-ring traffic could be
tuned between 0 and 100%. In Figure 17, we depict the average extra cost of the pure DPRing
and pure SPRing solution over the optimal hybrid design. From this figure, we can conclude
that the pure SPRing solution performs well as long as the majority of the ring traffic is intra-
ring traffic. As soon as some inter-ring traffic is present, the hybrid design becomes the
optimal solution, even in the equal cost scenario, which we consider here. E.g. in case 25% of
the traffic leaves the ring, the hybrid design consists for about 30% of DPRings (see Figure
18). Thus if the DPRing would be cheaper than the SPRing, the cost savings could be more
explicit. When 50% of the traffic is inter-ring traffic (which is realistic in practice), the
difference between the pure DPRing solution and pure SPRing solution is not that large, but
still both designs are more than 20% more expensive than the optimal hybrid solution. The
hybrid solution consists of an almost equal amount of SPRings and DPRings in this case (see
Figure 18). For an even higher amount of inter-ring traffic, the DPRing solution is clearly
more interesting than the SPRing solution. In this case the difference between the hybrid
solution and the DPRing solution is not that large, as most of the rings are DPRings (see
Figure 18). In case all traffic leaves the ring, the hybrid solution is equal to the DPRing
solution. In this case the SPRing solution requires twice as much wavelengths (see example in
Figure 16) and is thus almost twice as expensive as the DPRing solution. The reason why the
SPRing solution is not 100% more expensive than the DPRing solution, is because of the
granularity of the rings (e.g. if the DPRing solution requires 40 wavelengths, thus 3 rings of
16 wavelengths, the SPRing requires 80 wavelengths thus 5 rings of 16 wavelengths). We can
thus conclude that the hybrid design is particularly interesting in the equal cost scenario,
when about half of the traffic leaves the ring and is protected with drop & continue. In case
the DPRing would be cheaper than the SPRing, even larger savings could be obtained.
CHAPTER 515
6
0%
20%
40%
60%
80%
100%
0% 25% 50% 75% 100%
Amount of inter-ring traffic
Relative extra cost compared to hybrid design
only DPRing
only SPRing
Figure 17: Relative cost comparison
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0% 25% 50% 75% 100%
Amount of inter-ring traffic
Amount of
DPRings
Amount of
SPRings
Figure 18: Contribution of DPRing and SPRing in hybrid design
DIMENSIONING AND CONFIGURATION OF A WDM RING 15
7
5.6 SDH-over-WDM ring design
5.6.1 Problem formulation
We consider the design of a stack of SDH MS-SPRings on an unprotected WDM
ring. In such a configuration, each SDH MS-SPRing is supported by one wavelength in each
direction around the WDM ring, as in Figure 5. Fixed OADMs can be used for adding/
dropping the wavelength from the WDM ring towards the SDH ADM. However, as already
explained in section 5.2.3, not all SDH MS-SPRings might require an ADM in each node of
the ring: when we can route all SDH connections terminating in a certain node on one single
SDH MS-SPRing, there is no need for the other stacked MS-SPRings to have access to this
node. This means that there is no SDH ADM needed in this node for the other rings, and the
wavelength can simply be passed through in the OADM instead of being added/dropped. This
results in substantial cost savings in SDH equipment, as well as less add/drop required on the
OADMs.
The problem we consider is to route the SDH traffic in such a way as to minimize the
total amount of SDH ADMs required on all stacked rings, thereby reducing the SDH
investment costs as much as possible. As an example, we consider a stack of STM-16 MS-
SPRings on the wavelengths of the WDM ring. We assume the required number of MS-
SPRings and the traffic pattern of bidirectional VC-4 connections (155 Mb/s) between nodes
as given. This means the optimization of the amount of required rings has already been
performed, and we know for each connection in which direction it is routed (e.g. using the
ring loading solution from section 5.3.4). We still need to decide on which stacked ring each
connection has to be routed, in order to minimize the amount of ADMs required on all stacked
rings. An example of such a traffic pattern is given in Figure 19, for a 5-node ring with 9
connections (represented by their capacity and an alphabetical identifier).
A connection can be split up among several stacked rings, i.e. half of the connection
can be routed on the first stacked ring, while the other half is routed on the second stacked
ring. However the connection must always be split in integer parts and each part must be
routed end-to-end within the same stacked ring.
1
5 2
34
6 (F)
7 (I)
6 (H)
9 (G)
5 (D)
3 (B)
5 (C)
11 (A)
4 (E)
Figure 19: Given demand pattern on a ring
CHAPTER 5158
We again assume the ring consists of a set of nodes N to be labeled consecutively
from 1 to n. Furthermore there are K connections k, each with a demand of d(k) wavelengths.
The route of each connection k is fixed and designated by r(k,i), with:
r(k,i) = 1, if the route of connection k includes the link between node i and i + 1,
= 0, if not.
We assume the set M of stacked rings is also given and the amount of stacked rings is
m = |M|. We will later describe how to obtain m. The variables to be optimized are:
x(k,j) = amount of connection k which is routed on ring j.
y(i,j) = 1 if an ADM is installed in node i of ring j,
= 0 if not.
Under these assumptions we can formulate the problem mathematically as follows:
Minimize:
∈∈NiMj
jiy ),(
Subject to:
∈
≤
Kk
Sikrjkx ),().,( ∀=i∈N, ∀=j∈M
∈
=
Mj
kdjkx )(),( ∀=j∈M, ∀=k∈K
∈
≤−−
Kk
jiyPjkxikrikr ),(.),(.)1,(),( ∀=i∈N, ∀=j∈M
x(k,j) integer ∀=j∈M, ∀=k∈K
0 ≤ x(k,j) ≤=d(k)∀=j∈M, ∀=k∈K
{}
1,0),( ∈jiy ∀=i∈N, ∀=j∈M
The objective is to minimize the total amount of SDH ADMs. The amount of rings
does not need to be optimized but is given. The first constraint expresses that on each link of a
ring, only S slots are available for working traffic. For an STM-16 MS-SPRing, S equals 8.
The second constraint states that no demand of a connection should remain unassigned to a
ring. The third constraint restricts the amount of traffic that can be added/dropped in a node of
a ring if an ADM is installed in that node. If no ADM is present in that node, no traffic can be
added/dropped. P represents the number of add/drop ports on an ADM. For non-blocking HO
ADMs, which are most commonly used in STM-16 MS-SPRing configurations, P equals 16.
The remaining constraints are straightforward and restrict the values of the variables.
5.6.2 Bounds and additional constraints
As already noted in [23] (for dedicated protected rings) a minimal ILP formulation,
such as the above, is not a very effective one. The bound on the objective value obtained from
the linear programming relaxation is weak. In addition, all the rings are identical, so each
feasible solution is repeated an enormous number of times, just by permuting the number of
rings. To resolve this, we added a number of extra constraints and fix some variables in
advance in order to tighten the formulation.
A first restriction we can add is that each ring must contain at least 2 ADMs. Thus:
DIMENSIONING AND CONFIGURATION OF A WDM RING 159
∈
≥
Ni
jiy 2),( ∀=j∈M
Again, this is assuming we know the amount of stacked rings m in advance. We can
devise a lower bound for m from the given routing. The amount of traffic routed on the link
between node i and i + 1 is denoted by Li. The amount of traffic add/dropped at node i is
denoted as ADi. Both values can be easily calculated as follows:
∈
=
Kk
ikdikrL )().,(
≠−
∈
=
),()1,(
)(
ikrikr
Kk
ikdAD
As such the minimum amount of required stacked rings mmin is bounded by:
ù
ê
ê
é
=
∈),max(max
min P
AD
S
L
mii
Ni
In most cases the value m = mmin results in a feasible solution for the stacked ring
problem, however in some cases it is not possible to assign all the traffic to all rings (even
without minimizing the amount of ADMs). In this case, the optimization with m = mmin will
not result in a feasible solution, and we will try the optimization again with m = m + 1, until a
feasible solution can be reached.
For each node, we can also calculate a lower bound on the amount of required
ADMs, based on the amount of add-drop traffic in a node:
=
=−
∈
=
0),(
1)1,(
)(
ikr
ikr
Kk
W
ikdAD
=
=−
∈
=
1),(
0)1,(
)(
ikr
ikr
Kk
E
ikdAD
E
i
W
ii ADADAD +=
Where ADi again represents the total amount of add-drop traffic in node i,
while W
i
AD and E
i
AD represent the amount of traffic that is add-dropped towards the west and
east side respectively of the ring in node i.
Then a lower bound on the amount of ADMs in node i is:
ö
ç
ç
è
æù
ê
ê
éù
ê
ê
éù
ê
ê
é
=P
AD
S
AD
S
AD
LB i
E
i
W
i
i,,max
=
Because the add-drop capacity is not only limited by the capacity on the ring (S
capacity slots are available at both west and east side of the ADM), but also by the port
limitations (only P add-drop ports are available per ADM).
Thus, adding this lower bound to the formulation gives:
CHAPTER 5160
∈
≥
Mj
i
LBjiy ),( ∀=i∈N
To limit the number of possible permutations we can also fix some of the ADMs at
the node where most ADMs are required, given by:
)(max
max i
Ni LBLB ∈
=
)(argmax
max i
Ni
LBn
∈
=
Fixing LBmax ADMs in node nmax gives:
y(nmax,j) = 1 j = 1 … LBmax
We can even do a little better by also fixing the ADMs in a second node. If nmax* is
the node where most ADMs are required to exchange traffic with nmax and LBmax* is the
amount of ADMs required (both nmax* and LBmax* can be calculated as nmax and LBmax, but
only taking into account the traffic exchanged with nmax), then we can also fix:
y(nmax*,j) = 1 j = 1 … LBmax*
5.6.3 Results
The formulation of section 5.6.1 and the additional constraints and fixing of the
variables as in section 5.6.2 can now be implemented in a solver, such as CPLEX to solve
problems. In this section we present some results obtained in this way. We always consider
STM-16 rings (S = 8) with non-blocking ADMs (P = 16). As a first example, we give the
result for the ring depicted in Figure 19. In this case, we can easily calculate the amount of
required stacked rings nmin = 3. This would require 15 (= 3×5) ADMs if an ADM were to be
installed in each node of every ring. The optimal result however shows that only 9 ADMs are
really needed and 6 ADMs can be replaced by a pass-through wavelength. In this case, an
efficient planning can thus result in savings of 40% with respect to the amount of ADMs
required. In Figure 20 we represent a possible ring assignment for the optimal solution. The
nodes shaded in grey are the ones where no access is required to the ring, while in the other
(non-shaded) nodes an ADM is installed.
1
52
34
6 (F)
7 (I)
1 (G)
8 (A)
1
52
34
8 (G)
3 (B)
3 (A)
4 (E)
1
52
34
6 (H)
5 (D)
5 (C)
Figure 20: Possible ring assignment for the optimal result
DIMENSIONING AND CONFIGURATION OF A WDM RING 161
In Table 4 we present some results for different ring sizes. Different numbers of
nodes n as well as different numbers of stacked MS-SPRings m are considered for random
traffic patterns. Demand sets were generated with individual values for each demand drawn
randomly out of a uniform distribution between 0 and Dmax VC-4s. In Table 4, the amount of
ADMs required in the optimized solution is represented. The last column of the table reflects
the savings that can be achieved by installing only the minimum amount of ADMs needed,
compared to the situation for which an ADM were to be installed in each node of each MS-
SPRing.
nD
max m# ADM Savings
5 3 2 7 30.0 %
6 3 2 8 30.0 %
6 4 2 9 25.0 %
7 3 2 10 28.6 %
7 4 3 14 33.3 %
8 4 3 15 37.5 %
8 5 3 14 41.7 %
8 6 4 18 43.7 %
9 4 4 20 44.4 %
9 5 5 23 48.9 %
9 6 6 30 44.4 %
Table 4: Results for different ring sizes
The results show that for small rings, with a limited number of nodes and stacked
rings, 25 to 30 % of the ADMs can be saved. For medium sized rings (more than 7 nodes and
4 stacked rings) we witness savings of 40 to 50 %. For larger rings, comparable or even better
results can be expected, although we were not able to solve these cases to optimality using the
CPLEX solver.
These savings in SDH equipment result in substantial cost savings, since the
expensive SDH ADMs can be replaced by a simple wavelength pass-through in the OADM at
virtually no cost. In addition, maximizing the amount of pass-through traffic minimizes the
amount of drop channels in the OADM, leaving room for cheap OADMs with limited add-
drop capabilities. Another advantage is that longer SDH rings are possible by reducing the
amount of ADMs. In order to assure fast protection switching times, the standardization
bodies have limited the MS-SPRing protocol to rings with up to 16 ADMs. By replacing
ADMs by a pass-through wavelength, the total number of nodes on the rings can be increased.
We have also studied the impact of different kinds of demand patterns on the
solution, as shown in Table 5. All results are obtained on a 6-node ring. We have however
chosen the sizes of the demands in such a way that the amount of stacked rings required was
always 4, in order to make meaningful comparisons. This implies that the maximum amount
of ADMs (when all nodes are fully equipped) is 24. The routing of the connections has been
done optimally. For the long demand pattern this means that the connections have been routed
by splitting them in half and routing them along both sides of the ring, to make better use of
the bandwidth.
CHAPTER 5162
Demand pattern # ADM Savings
Adjacent 24 0.0 %
Uniform 18 25.0 %
Hubbed 14 41.7 %
Long 12 50.0 %
Mainly adjacent 20 16.7 %
Mainly uniform 17 29.2 %
Mainly hubbed 14 41.7 %
Mainly long 13 45.8 %
Table 5: Results for different demand patterns
As expected, the adjacent demand pattern requires an ADM in each node of each
ring, so no savings are possible. The uniform demand pattern yields moderate savings, while
for the hubbed and long demand pattern the largest savings are possible. Since the extreme
traffic patterns are not always realistic, we have also performed simulations in which the
extreme traffic pattern was distorted by adding extra demand (with a random demand pattern).
This resulted in traffic patterns with mainly adjacent, uniform, long and hubbed demands. The
results are also shown in Table 5. We again see that the best savings are possible for the
mainly long demand pattern and the mainly hubbed demand pattern. For these two demand
patterns, dedicated protection rings have traditionally been favoured [21]. However, our
optimization of the number of ADMs in shared protection rings results in substantial savings,
such that the use of shared protection rings also becomes attractive for these demand patterns.
In fact, by splitting up the demand pattern over several stacked rings, the optimal demand
pattern within each stacked ring will be mainly an adjacent demand between the ADMs that
still reside on the ring.
A drawback of replacing ADMs with equipment without add/drop capabilities, is the
reduced flexibility of the network. When dealing with rapidly changing traffic patterns, the
optimal design (which is optimized for one specific traffic pattern) might not be able to
accommodate such changes. Hence, our approach, which optimizes the amount of ADMs, is
best suited for static traffic patterns.
The same model can be used for the optimization of pass-through traffic in stacked
WDM rings, in which some OADMs can be replaced by a fiber connector (or optical
amplifier) if there is no need for wavelength add/drop in certain nodes of the ring.
5.7 Conclusion
This chapter discussed various planning issues related to WDM rings. For the ring
loading problem, we derived theoretical bounds and described a model that can be used to
obtain optimal results. The mathematical bounds can be used to make a rough prediction of
the required ring capacity. Using these bounds, we compared the wavelength requirement of
the SPRing and DPRing for some characteristic ring demand patterns. The SPRing utilizes the
wavelengths better than the DPRing, especially in case of uniform and adjacent demand
patterns. The potential for capacity savings with SPRing gets higher as the ring collects more
nodes. In case of hub or long demand patterns, the SPRing hardly provides any capacity
advantage over DPRing.
The deterministic optimization algorithm for the ring loading problem is based on an
enhanced ILP formulation. The model can be used to compare the wavelength requirement of
DIMENSIONING AND CONFIGURATION OF A WDM RING 163
the SPRing with the DPRing for any particular set of demands and to determine how each
demand should be accommodated exactly on the ring. A case study revealed that for random
traffic patterns, a SPRing saves between 20% and over 40% of the wavelengths that would be
required by a DPRing.
We examined the need for wavelength conversion in OADMs of a SPRing by
solving the wavelength assignment for a predetermined routing. The used routing was the
optimal routing (as a result of solving the ring loading problem). For most of the demand
patterns we studied, the benefit of wavelength conversion in terms of wavelength
requirements turned out to be minimal. As such, the additional cost and complexity of
providing wavelength conversion in the nodes of the ring may not be justified.
In practice, the ring loading problem and the wavelength assignment are interrelated
and have to be solved in an integrated model to obtain the best result for the overall planning
problem. However, as a secondary result of our simulations, we conclude that in practice, the
ring loading problem and wavelength assignment problem can be solved separately, without
any significant loss of optimality.
To combine the advantages of DPRings and SPRing, we have devised a hybrid
network architecture merging both ring types in a stack of rings. A model has been developed
to optimize such network architectures, taking into account different cost models for DPRings
and SPRings. Over a wide range of cost values, both ring types can be advantageously
combined in a hybrid network design. SPRings provide a cost efficient means for
accommodating adjacent and uniform demands, even if their unit cost per wavelength is
higher than that of the DPRing. On the other hand, cheap DPRings help in lowering the cost
of hybrid ring design solutions to accommodate selective demands that are not very well
suited for sharing of protection capacity (demands following a hubbed or long pattern). Such a
hybrid scheme is particularly advantageous in case the rings use drop & continue, and the ring
contains a mixture of inter- and intra-ring traffic.
Finally, we have studied the problem of efficiently assigning connections to stacked
SDH rings on a WDM ring in order to minimize the amount of SDH ADMs required, and
optimize the amount of optical pass-through traffic in the nodes of the ring. Results were
obtained for different ring sizes as well as for different demand patterns. For medium sized
rings with random demand patterns, savings of 40% and more were observed in terms of the
required number of ADMs. The largest savings are possible for demand patterns with a lot of
long connections or a lot of hubbed demand.
CHAPTER 5164
5.8 References
[1] O. Gerstel, "Opportunities for optical protection and restoration", Proceedings of OFC'98, Session ThD,
pp. 269-270, San Jose (CA), February 1998.
[2] R. Batchellor, "Putting the lens to optical-layer protection", Ligthwave, Vol. 15, No. 10, September
1998.
[3] P. Arijs, P. Demeester, "Efficient design of stacked SDH multiplex section shared protection rings",
Proceedings of 3th European Conference on Networks and Optical Communications (NOC'98), pp. 164-
171, Manchester (UK), June 1998.
[4] P. Arijs, P. Demeester, "Design issues for WDM rings", 6th International Conference on Optical
Communications and Networks, Paris (France), January 1999.
[5] P. Arijs, M. Gryseels, P. Demeester, "Planning of WDM ring networks", Photonic Network
Communications, special issue on WDM transport networks: key elements and architectures, Vol. 2,
No. 1, pp. 33-51, January-March 2000.
[6] P. Arijs, P. Demeester, " The merit of shared and dedicated protection WDM rings in a hybrid network
design", Proceedings of OFC 2000,Paper FE6, Baltimore (MD), March 2000.
[7] P. Arijs, D. Colle, P. Demeester, "Optimization models for cost savings in stacked ring network
design", Proceedings of 5th INFORMS Telecommunications Conference, Boca Raton (FLA), March
2000.
[8] S. Cosares, I. Saniee, "An optimization problem related to balancing loads on SONET rings",
Telecommunications Systems, Vol. 3, No. 2, pp. 165-181, November 1994.
[9] Y.-S. Myung, H.-G. Kim, D.-W. Tcha, "Optimal load balancing on SONET bidirectional rings",
Operations Research, Vol. 45, No. 1, pp. 148-152, January-February 1997.
[10] M. Dell'Amico, M. Labbé, F. Maffioli, "Exact solution of the SONET ring loading problem", Working
paper SMG96/05 Service des Mathématiques de la Gestion, Université Libre de Bruxelles (Belgium),
May 1996.
[11] A. Schrijver, P.D. Seymour, P. Winkler "The ring loading problem", SIAM Journal of Discrete
Mathematics, Vol. 11, No. 1, pp. 1-14, February 1998.
[12] N. Karunanithi, T. Carpenter, "SONET ring sizing with genetic algorithms", Computers and Operations
Research, Vol. 24, No. 6, pp. 581-591, 1997.
[13] R. Ramaswami, K.N. Sivarajan, "Routing and wavelength assignment in all-optical networks",
IEEE/ACM Transactions on Networking, Vol. 3, No. 5, pp. 489-500, October 1995.
[14] N. Wauters, P. Demeester, "Design of the optical path layer in multiwavelength cross-connected
networks", IEEE Journal on Selected Areas in Communications, vol.14, no. 5, pp. 881-892, June 1996.
[15] G. Ellinas, K. Bala, G.-K. Chang, "A novel wavelength assignment algorithm for 4-fiber WDM self-
healing rings", Proceedings of IEEE ICC'98, pp. 197-201, Atlanta (GA), June 1998.
[16] D.K. Hunter, D. Marcenac, "Optimal mesh routing in 4-fiber WDM rings", IEE Electronic Letters, Vol.
34, No. 8, pp. 796-797, April 1998.
[17] C. Law, K.-Y. Siu, "On-line routing and wavelength assignment in WDM rings", Proceedings of SPIE
Conference on All-Optical Networking, pp. 276-288, Boston (MA), September 1999.
[18] C. Félicité, D. Marcenac, "The wavelength allocation problem in WDM networks", Proceedings of
NOC'99, pp. 18-24, Delft (Netherlands), June 1999.
[19] D. Marcenac, C. Félicité, "Practical benefits of 4-fibre WDM rings with realistic traffic", Proceedings
of OFC'99, Paper TuF5, San Diego (CA), February 1999.
DIMENSIONING AND CONFIGURATION OF A WDM RING 165
[20] T. Carpenter, S. Cosares, I. Saniee, "Demand routing and slotting on ring networks", DIMACS
Technical Report 97-02, January 1997.
[21] T.-H. Wu, R.C. Lau, "A class of self-healing ring architectures for SONET network applications", IEEE
Transactions on Communications, Vol. 40, No. 11, pp. 1746-1756, November 1992.
[22] J. Harder, "SONET ring design with cutset inequalities and other speedups: some computational
experience", Proceedings of 4th International Conference on Telecommunications Systems, pp. 477-491,
Nashville (TN), March 1996.
[23] A. Sutter, F. Vanderbeck and L. Wolsey, "Optimal placement of add/drop multiplexers: Heuristic and
exact algorithms", Operations Research, Vol. 46, No. 5, pp. 719-728, 1998.
[24] M. Armory, J.G. Klincewicz, H. Luss, M.B. Rosenwein, "Design of stacked self-healing rings using a
genetic algorithm", Proceedings of 4th INFORMS Telecommunications Conference, Boca Raton (FLA),
March 1998.
[25] A.L. Chiu, E.H. Modiano, "Reducing electronic multiplexing costs in unidirectional SONET/WDM
ring networks via efficient traffic grooming", Proceedings of Globecom'98, Sydney (Australia),
November 1999.
[26] J.M. Simmons, E.L. Goldstein, A.A.M. Saleh, "Quantifying the benefit of wavelength add/drop in
WDM rings with distance-independent and dependent traffic", Journal of Lightwave Technology, Vol.
17, No. 1, pp. 48-57, January 1999.
[27] O. Gerstel, P. Lin, G. Sasaki, "Wavelength assignment in WDM ring to minimize cost of embedded
SONET rings", Proceedings of IEEE Infocom'98, San Francisco (CA), March 1998.
[28] O. Gerstel, P. Lin, G. Sasaki, "Combined WDM and SONET network design", Proceedings of
Infocom'99, New-York (NY), March 1999.
[29] D. Colle, P. Arijs, P. Demeester, K. Struyve, "Comparison of architectures for stacked ring networks
featuring compact add/drop multiplexers", Proceedings of DRCN 2000, Munich (Germany), April
2000.
[30] ILOG Inc., CPLEX Division: Using the CPLEX callable library, 1997.
[31] O. Gerstel, R. Ramaswami, G.H. Sasaki, "Fault tolerant multiwavelength optical rings with limited
wavelength conversion", IEEE Journal on Selected Areas in Communications, Vol. 16, No. 7, pp. 1166-
1178, September 1998.
CHAPTER 516
6
CHAPTER 6
Planning of interconnected WDM rings
6.1 Introduction
In Chapter 4 WDM was merely used in the network as a point-to-point transmission
technology, to economically boost the capacity of fiber links. A first step in the direction of
optical networking was considered in Chapter 5, which described the specific design issues
related to introducing a single WDM ring in leopard spots of the network. The next logical
step is to extend the WDM networking functionality throughout the entire meshed network.
Such meshed networks can rely on optical cross-connects (OXC), to switch wavelengths
between multiple fiber pairs. However, at the moment technologies to manufacture cost-
effective, scalable large-size OXCs are only starting to appear [1]. On the other hand, optical
add-drop multiplexers (OADMs) as used in WDM rings, require far less internal connectivity
than OXCs, while manifold technologies for OADMs have already been demonstrated
successfully [2]. Therefore, in the near term, all optical networks can rely on WDM rings,
covering the underlying meshed fiber infrastructure.
For designing a network based on interconnected WDM rings, several strategic
decisions need to be made. First of all, the protection mechanism of the rings needs to be
determined. A network can either be based on a homogenous set of dedicated or shared
protection rings. While shared protection rings yield a higher throughput due to the sharing of
protection bandwidth (see chapter 5), this effect has shown to diminish in an interconnected
ring network, as shown in chapter 4 for SDH MS-SPRings. Alternatively, a network can also
be based on a heterogeneous set of shared and dedicated protection rings, using the best suited
ring type according to the specific demand pattern on the ring.
A second strategic decision that needs to be taken, concerns the architecture of the
interconnected WDM rings. When the network is small in terms of the amount of nodes, it
might be possible to choose the rings in such a way that each two nodes of the network are
contained in at least one ring [3]. As such, all traffic can be routed within one single ring, and
no ring interconnection is required. As soon as the network contains a significant amount of
nodes, such an approach requires a high amount of large-size rings. In addition, the free
capacity on one ring, can not be reused by a connection of which the source and termination
node are not contained in the ring. Such an architecture thus makes inefficient use of the
resources. Therefore, large-scaled networks based on WDM rings, require ring
interconnection, in order to result in an economical, reliable and manageable network design.
The rings can be interconnected logically in two ways, either by defining a hierarchy as in
[4][5], in which several rings of a lower hierarchical level are connected to a ring in the next
hierarchical level, until the highest level, consisting of one ring, is reached. The other
alternative is to use a flat architecture (i.e. without defining a hierarchy), in which all rings are
CHAPTER 6168
on the same hierarchical level, and can be interconnected with each other freely [6]. In a
hierarchical architecture, the routing is trivial because the routing possibilities are constrained
by the hierarchy. Routing of traffic between two rings consists of climbing up in the hierarchy
until a ring is reached that connects these two rings through one or more levels. This also
results in a simple management model. On the other hand, in a flat architecture the routing
decisions are more complex, but connections can be routed more freely, resulting in a
potentially higher throughput.
Finally, a decision needs to be taken for interconnecting the rings on the physical
level. As described in Chapter 3, hard-wired interconnection, through back-to-back OADMs
is suitable for network architectures conveying a static traffic matrix. If dynamic
reconfiguration is needed, an intermediate OXC can be deployed or a collapsed OXC, which
contains the OADMs as well. However, for these OXCs, the same comments regarding
market availability hold, as described earlier for meshed WDM networks. Failures of the ring
interconnection gateways can be survived from by using dual node interconnection strategies,
such as drop & continue or other matched node mechanisms, as described in chapter 3. One of
the differences with SDH/SONET and WDM ring design is the analogue medium of the
latter, which imposes physical impairments such as noise accumulation and dispersion that
have to be taken into account. In standardization bodies many contributions have suggested
the use of transponders (3R regenerators) at sub-network interfaces (SNI) to effect a clean
hand-off of signals between domains and provide OCh performance monitoring [7]. In the
interconnected WDM ring network, the interconnection points between rings are natural
places to put such transponders.
In this chapter we will describe the design of a flat interconnected WDM ring
network, using optical channel dedicated protection rings (OCh-DPRings) with hard-wired
ring interconnections (using transponders). The different ring network design issues such as
ring selection, dimensioning of rings and routing over these rings, will be described. For these
issues, an overview of related work in literature will be given. Afterwards, we describe and
compare our solution methods for planning the considered ring network architecture. We use
both optimal as heuristic techniques, and compare these methods with each other. We will
also use these planning methods to study the impact of network and ring size on the cost and
availability of the ring network. Finally, we also compare interconnected OCh-DPRings, with
meshed WDM architectures based on OXCs in order to prove the cost-effectiveness of
interconnected rings.
The main results of this chapter can also be found in [6], [8], [9], [10], [11] and [12].
PLANNING OF INTERCONNECTED WDM RINGS 169
6.2 Planning issues for interconnected WDM rings
Ring identification
and interconnection
Inter- and intra-ring
routing and
ring dimensioning
Ring network design
• Traffic
• Costs
• Ring type
• Ring interconnection
• Failure scenarios
• Physical limitations
• ...
Figure 1: WDM ring network design phases
Once the decision has been made about which ring type to use, and which
interconnection strategy between rings, the process of designing an interconnected ring
network can be considered, which typically consists of 2 main sub-problems (Figure 1):
1. Ring identification (greenfield or not, hierarchical or not)
2. Ring dimensioning based on inter- and intra-ring routing (and wavelength
assignment)
No specific spare capacity planning phase (as in meshed networks [8]) is required for
interconnected rings, because the spare capacity is embedded in the rings.
The ring identification problem consists of finding a minimum cost combination of
interconnected rings, such that all demands in the network can be routed and protected. The
problem can be stated in a greenfield scenario (i.e. with no cables installed yet) or in an
existing underlying meshed network topology. As there exist a large number of potential ring
positions (even in an existing network topology) and even more combinations of rings, this is
a very complicated problem. For evaluating the cost of a combination of rings, it is required to
dimension the rings, by routing the traffic across and within the rings in the network.
Therefore the ring dimensioning process is typically executed a large number of times during
the ring identification process, using some kind of feed-back mechanism to adapt the current
ring selection. Hence, it is of paramount importance that the dimensioning algorithm is
sufficiently fast
Hierarchical ring network design was already considered in [13] for SDH/SONET
rings, using a clustering heuristic based on topological and traffic affinity, starting from the
lowest level in the hierarchy and subsequently working its way up. A somehow similar but
elaborated approach has been taken in [4] and [5] for WDM rings.
The non-hierarchical ring network design problem has received most attention in
literature, mainly for SDH/SONET rings. Because of the complexity of the problem, most
authors use a heuristic optimization method. A first kind of heuristic solution method adds
rings to the network in a sequential fashion (based on certain traffic metrics), until enough
rings are present such that all traffic can be routed. This approach was taken in [14] and [15].
A second kind of heuristic consists of several phases in which potential rings are generated,
and several suitable ring combinations are evaluated by routing traffic on the selected rings
[16]. Other authors have used heuristic search techniques, such as simulated annealing [17] or
CHAPTER 6170
genetic algorithms [18] to find suitable ring combinations. Also integer linear programming
techniques (ILP) [19] have been used, but mostly make some simplifying assumptions to
facilitate the problem complexity. E.g. [20], make more or less abstraction of the routing
problem. In [21] an optimal routing is used, but only considering a very limited subset of
possible rings. In [22] on the other hand, the problem is split up in two phases (ring
assignment and routing) and solved both using ILP.
After the planning phase, an additional routing phase can be considered. Either to
route different traffic matrices than the matrix for which the network was dimensioned (to
evaluate the robustness of the dimensioning under varying traffic scenarios), or to verify
whether additional traffic can be routed in an already dimensioned ring network.
6.3 Problem formulation
In this paragraph we state the considered WDM ring network design problem in a
more formal way. We consider three main network planning problems in this chapter:
• Ring identification: given a network topology without any rings present, determine the
optimal placement and interconnection of OCh-DPRings in order to route all traffic.
This is the planning problem an operator faces in the initial stage of deploying a
network based on WDM rings.
• Ring dimensioning: given a network topology, the placement and interconnection of the
OCh-DPRings, and already routed traffic, dimension the rings such that all future traffic
can be transported. This is the planning problem an operator faces, when he has to
dimension/upgrade his existing WDM ring network to meet future traffic growth.
• Ring routing: given a network topology, a dimensioned set of interconnected OCh-
DPRings, and current traffic routes, route the maximal amount of new traffic in these
rings. This is the problem an operator faces for accommodating new traffic in today's
dimensioned WDM ring network.
The detailed requirements and optimization objectives of each planning problem are
stated in section 6.3.2. First we will give an overview of the input parameters to these
planning problems.
6.3.1 Input parameters
6.3.1.1 Network topology
We consider the physical network topology to be given as a set of cables (or links) L,
containing sufficient fibers, and a set of offices (or nodes) N.
When considering the ring identification problem, there are no rings present yet. In
other words, we want to build a ring network on top of an already existing cable network
topology.
In case of ring dimensioning, we also know the ring-sites R, i.e. the positions were
OCh-DPRings can be deployed. If a site r∈R, passes through office n∈N, then s(r,n) = 1
(otherwise 0). In a ring-site, an OCh-DPRing can be deployed by using one fiber pair in the
cables, and one OADM in the offices along which it is routed. Multiple overlay rings
deployed on one ring-site are called stacked rings.
PLANNING OF INTERCONNECTED WDM RINGS 171
In case of ring routing, not only the positions of the ring-sites are given, but also the
amount and dimensioning of the rings on each ring-site.
6.3.1.2 Traffic
We consider a static bi-directional traffic matrix, expressing the amount of
wavelengths required between each two offices in the network: the demand dij expresses the
amount of bi-directional wavelengths that have to be routed and protected between each two
different offices i∈N and j∈N (j>i). All traffic needs to be protected in the rings.
6.3.1.3 Rings
We consider 2-fiber OCh-DPRings. We assume that all new rings (i.e. the rings to be
designed) have the same characteristics. This means all rings are OCh-DPRings with the same
amount of wavelengths W. Legacy rings (i.e. rings present in the network prior to the network
design) can have different characteristics, but must also be OCh-DPRings. However, such
rings can have a lower channel count due to outdated technology. Legacy rings can also
contain some legacy traffic. In this case, these legacy rings are characterized by the amount of
free wavelengths. The cost structure of the rings depends on the ring-site on which they are
deployed. These costs are presented in section 6.3.1.6.
6.3.1.4 Ring interconnection
When traffic needs to be routed over multiple rings, ring interconnection is required.
We assume that ring interconnection is performed hard-wired, and not using OXCs, which is
a valid assumption for static traffic. This is achieved by interconnecting the wavelength
dropped at the tributary side of the OADM in one ring to the tributary side of the OADM in
the other ring (e.g. using a fiber distribution frame). Within an office, traffic can thus be
exchanged between all the rings that have an OADM in this office.
Concerning the protection strategy for traffic that is routed over multiple rings, we
assume that traffic is protected ring-per-ring, and not end-to-end. As such, simultaneous
failures in different rings of the path can be survived from independently. Regarding the
interconnection of rings, drop & continue can be used or not. If drop & continue is used, ring
interconnection has to be done via at least 2 offices.
6.3.1.5 Transparency and wavelength conversion
We consider opaque and partly opaque networks. In the case of opaque networks, 3R
regeneration is performed in the OADMs of the rings. The network is engineered link-per-link
and transmission issues thus have no impact on the network design. The network is designed
simply on the grounds of cost optimization.
In the case of partially transparent networks, no 3R regeneration is performed in the
rings, but only at ring interconnection points. In such partially transparent networks, too long
rings (that do not have acceptable transmission performances) have to be avoided. Limitations
on the amount of nodes in a ring, fiber length and amount of amplifiers in the ring can be
specified when rings have to be identified.
While we do not consider wavelength conversion in the rings, we assume the use of
transponders at ring interconnection points, such that the wavelength can be changed at ring
interconnection points.
CHAPTER 6172
6.3.1.6 Costs
We consider a generic cost model, consisting of the following components:
• A ring cost ring
r
C for each ring to be deployed on ring-site r∈R: this cost includes the
fixed installation costs of the ring, fiber costs, amplifier costs, fixed OADM costs, …
• A routing cost route
r
C for routing one wavelength on a ring on ring-site r∈R: this costs
accounts for the modular components of the ring that can be installed on a wavelength
basis (e.g. tributary cards, transponders).
• An interconnection cost inter
C: the cost for interconnecting a wavelength between 2
rings (e.g. transponder cost).
While the first two cost components are potentially different for each ring-site
(depending on the circumference of the ring-site), the last cost component is a fixed cost.
The results in the remainder of this chapter are based on the above cost model, in
which we consider the cost of a ring r passing through Nr offices and having Ar amplification
sites to be:
AMP
r
OADMlink
r
ring
rCACCNC .).( ++= and route
r
route
rCNC .=
We thus assume a fixed link, OADM and amplifier cost for calculating the ring cost.
The ring cost is then proportional to the number of offices and amplification sites on the ring.
The routing cost is assumed to be proportional to the number of offices. This is just one cost
model. The ring and routing cost can in principle be much more complex and different for
every ring-site.
The used cost values are:
• Clink = 200
• COADM = 50
• CAMP = 40
• Croute = 20
• Cinter = 10
6.3.1.7 Availability
Besides cost, another important metric to evaluate network designs is the network
availability, which represents the probability that the network is capable of transporting the
given traffic matrix at any moment in time [23]. For calculating the network availability, we
need to know the failure rates of the network equipment. These failure rates can be expressed
as mean time between failures (MTBF) and mean time to repair (MTTR) a failure. The
unavailability of the network element is thus: MTTR/(MTBF+MTTR). The failure rates we
used in the chapter are shown in Table 1. Methods to calculate the network availability in ring
and mesh based networks are described in [23].
MTBF MTTR
Line system 1 failure per year per 300 km 24 h
OADM and OXC 105 h 6 h
In-line amplifier 5.105 h 24 h
Table 1: Failure rates
PLANNING OF INTERCONNECTED WDM RINGS 173
6.3.2 Network design objectives
The three planning problems related to interconnected rings, considered in this
chapter, all have different objectives. These objectives are shortly described below.
6.3.2.1 Ring routing
The objective in the ring routing phase, is maximization of the throughput. Given a
dimensioned network, the operator wants maximize his revenue by setting up as much as
possible connections. Thus, we have to decided which (if not all) connections have to be
routed, and along which rings they have to be routed. The throughput can then be calculated
by dividing the amount of traffic that can be routed in the dimensioned network by the total
amount of given traffic. If not all traffic can be accommodated in the network infrastructure
the throughput will be less than 100%.
6.3.2.2 Ring dimensioning
In case of ring dimensioning, the objective is to route all the traffic, while
minimizing the cost of the newly added stacked rings on the given ring-sites. If some legacy
rings are present on these ring-sites, it is first tried to reuse the remaining capacity on these
rings as much as possible, then new rings with a predefined set of wavelengths are added.
6.3.2.3 Ring identification
Again, the objective is to route all the traffic, while minimizing the cost of the newly
added rings. In this case no ring-sites are given. Ring identification thus includes determining
both the positions as the dimensions of the rings. Certain constraints must be obeyed, e.g.
limitations on the amount of nodes in a ring, fiber length and amount of amplifiers in the ring
can be specified to avoid transmission impairments. Also the amount of ring-sites can be
restricted between a minimum and maximum value, due to cost, reliability or management
considerations. If both values are equal, the amount of ring-sites to be identified is fixed.
6.4 Equivalent network
The equivalent network represents the topological relationship between the different
ring-sites and can be used to assist in the routing decisions to be taken within the ring
network. The solution methods for ring routing and ring dimensioning all rely heavily on the
adoption of the equivalent network, and therefore we present this concept first.
In the equivalent network, each office is represented by a separate node. In addition,
each ring-site is also represented by a node. Two types of links exist in the equivalent
network. A link between a node representing an office n and a node representing a ring-site r
is present if s(r,n) = 1. Such a link is called an access link and represents the fact that the ring-
site passes through this office (and thus traffic injected at this office has access to the rings of
the ring-site). A link between two nodes representing a ring-site is called an interconnection
link and represents the fact that both ring-sites are interconnected and thus traffic can be
exchanged between rings of both ring-sites. In case no drop & continue is used, rings can be
interconnected in just one office, in case drop & continue is required, rings must be
interconnected through at least two ring sites. Thus, depending on the use of drop & continue
CHAPTER 6174
or not, the topology of the equivalent network can be different. An example of a ring network
and the corresponding equivalent network is given in Figure 2.
ring 1n1
n4
n3
n2
n7
n6
n5
n8
ring 2
ring 3
ring 4
office
OADM Intercon.
link
access
link
n1
n4
n3
n2
n7
n6
n5
n
8
ring 1
ring 2
ring 3
ring 4
Ring network Equivalent network
Figure 2: Ring network and equivalent network
Traffic between two offices in the ring network can be routed by finding a path
between the corresponding nodes in the equivalent network. A path in the equivalent network
should always start and end with an access link, while all intermediate links should be
interconnection links. Afterwards, the obtained path in the equivalent network can be
translated again to a sequence of interconnected rings in the original network.
We assign weights wl and capacities cl to link l in the equivalent network in such a
way, that wl represents the cost of routing one wavelength over the ring(s) correspondent with
the end-nodes of the link, and cl represents the capacity on the ring(s).
If l is an access link between nodes representing office n and ring-site r:
2.2
route
r
ring
r
l
C
W
C
w+=
)(. rxWcl=
The cost of a ring is thus divided over the links adjacent to the corresponding node in
the equivalent network. These costs are divided by 2 because a connection routed over a ring,
will pass through the correspondent node in the equivalent network, and will thus be routed
over 2 adjacent links of this node. The fixed cost is in turn divided by the amount of
wavelengths on the ring, thereby distributing this cost equally among all channels on the ring,
such that wl represents the cost of routing one wavelength over the ring.
The capacity on the ring is accredited to the access link towards the ring. The integer
variable x(r) represents the amount of OCh-DPRings to be deployed in each ring-site r∈R.
When multiple rings are deployed in the same ring-site r (i.e. when x(r) > 1), this represents
stacked rings. W represents the number of channels on each OCh-DPRing.
If l is an interconnection link between nodes representing ring-site r1 and r2:
inter
route
r
ring
r
route
r
ring
r
lC
C
W
CC
W
C
w.
2.22.2
2211
α
++++=
{}
)(),(min. 21 rxrxWcl=
Traffic routed over the interconnection link, is routed over both rings correspondent
with the end nodes of the interconnection link. The interconnection link thus comprises half
of the per-wavelength cost of both rings. The interconnection costs are in turn accredited to
PLANNING OF INTERCONNECTED WDM RINGS 175
the interconnection links. The factor α in the last term of wl allows to express whether
interconnection between two rings takes place in one or two nodes (to allow drop &
continue). In the former case α = 1, in the latter case α = 2.
The capacity of the interconnection link, is equal the minimum of the capacity
available on both rings correspondent with the end nodes of the interconnection link. Indeed,
in order to route traffic between both rings, enough capacity should be available on both
rings.
Such a model, in which cost and capacity of a ring are divided over the links adjacent
to the corresponding node in the equivalent network, can only be used for dedicated
protection rings. In such rings a demand uses up an entire channel on the ring, and the ring
sizing does not depend on the traffic pattern, but only on the amount of traffic on the ring. It is
important to stress that these costs are only used for making routing decisions and not for
calculating the network cost, because they do not take the discrete installation steps of ring
capacity into account.
6.5 Ring routing
Ring routing starts with a given set of OCh-DPRings. The different rings can have a
different amount of wavelengths, and some traffic can already be routed on these rings. As
such it is possible to take into account legacy rings and traffic. Next to a traffic matrix, the
fact whether drop & continue is used or not, is given as input. The ring routing algorithm then
attempts to route as much traffic as possible on the given rings, without adding any new rings.
After the equivalent network has been constructed, we can perform the routing. First,
the connectivity of the equivalent network is verified. If the network is not connected, no
routing is performed. If the network is connected, the routing can be performed according to
two routing algorithms, as described below.
6.5.1 Integer linear programming
The routing problem can be stated as an integer linear programming problem. The
structure of this problem is very similar to a multi-commodity flow problem [24], translated
to the equivalent network. The amount of unused capacity on the rings on each ring-site is
given as c(r). Within the considered ring network, each demand dij can be routed along a large
number of possible paths pij∈Pij. Such a path consists of a number of interconnected ring-
sites. If ring-site r∈R is contained in path pij∈Pij, then q(pij,r) = 1 (otherwise 0). The integer
decision variable y(pij) represents the amount of traffic that is routed on path pij∈Pij.
In principle, all possible paths between each two offices must be considered in order
to obtain the optimal solution. However, typically only short paths (i.e. containing a limited
amount of rings will be considered). Indeed, if the capacity in the ring network is limited, it is
preferable to set-up as much as possible short connections, since they use capacity on the least
amount of rings, and thus allow the most of such connections to be set-up. E.g. instead of
setting up one connection across three rings, it is better to set-up one connection within each
of these 3 rings. On the other hand, if sufficient capacity is present on the rings to set up all
traffic, choosing the shortest path is most efficient, because this minimizes the ring routing
cost. Finally, the total amount of possible paths in a ring network is typically not so high.
Whereas the underlying meshed network might be dense (i.e. consist of a large number of
CHAPTER 6176
links), the interconnected ring network (or equivalent network) is typically more sparse,
because it typically consists of a limited number of rings.
The set of paths Pij we use for each pair of offices is determined by searching for the
K-shortest paths [25] between the corresponding nodes in the equivalent network. A better
solution can be found by considering more possible paths (i.e. K large), however this tends to
increase the computational complexity of the problem considerably. In practice, only modest
values of K will be considered, which can still result in very good results, as argued above and
as will be shown later on.
The ring routing problem can now be formally stated as:
Maximize:
∈
>
∈∈Ni
ij
NjPp
ij
ijij
py )(
Subject to:
ij
Pp
ij dpy
ijij
≤
∈
)( ∀=i∈N, j∈N | j > i
)()().,( rcpyrpq
Ni
ij
NjPp
ijij
ijij
≤
∈
>
∈∈
∀=r∈R
y(pij) integer and ≥ 0 ∀=i∈N, j∈N | j > i
We thus want to maximize the total amount of traffic y(pij) routed on paths
pij=between all nodes i and j in the network. The first constraint states that the amount of
traffic routed between two individual nodes i and j should not exceed the total amount of
traffic dij to be routed between both nodes. The second constraint says that the amount of
traffic y(pij) routed over paths pij=that use ring-site r (if q(pij,r) = 1) should not exceed the
remaining capacity c(r) on this ring-site. The last constraint is trivial.
This model can then be implemented in an ILP solver such as CPLEX [26] to solve
the ring routing problem. Such an ILP solution method is suitable for optimizing medium
sized problems and when the solution time is not critical. For larger problem instances, or
when fast optimization is required, it is better to use a heuristic solution method, as described
in the next section.
6.5.2 Shortest path first routing
As already mentioned, short paths (i.e. over a limited number of rings) allow to
obtain a high throughput, because they waste capacity on a limited amount of rings.
Therefore, we have developed a simple but efficient heuristic solution method, which first
tries to route all single ring connections, then all connections that can be routed on 2 rings,
etc. The algorithm works as follows.
1. All connections are split in single-wavelength connections. E.g. a connection of 3
wavelengths between two offices, is split in three connections of one wavelength
between both offices. This allows to set-up part of a connection if there is not sufficient
capacity to set-up the entire connection.
2. Initialization: l = 2. Where l represents the hoplength (i.e. amount of links) of a path in
the equivalent network. Since each path in the equivalent network should start and end
PLANNING OF INTERCONNECTED WDM RINGS 177
with an access link, and contains only interconnection links in between, the amount of
rings corresponding with a path of hoplength l is l-1.
3. The free capacity of all links in the equivalent network is evaluated.
4. Links with no free capacity are removed from the equivalent network.
5. The shortest path in the equivalent network is calculated for every connection.
6. All connections with hoplength = l in the equivalent network are routed on the
corresponding l-1 rings, as long as sufficient capacity is available, and the capacity on
the links in the equivalent network is adapted.
7. Increase hoplength: l = l+1.
8. If not all connections are routed and some capacity is still available, return to step 3.
9. Stop.
6.5.3 Results
In this section, we present results for both above described routing algorithms. Three
sample networks were considered:
• a 9-node network, containing 4 ring-sites
• a 16-node network, containing 6 ring-sites
• a 32-node network, containing 12 ring-sites
The ring-sites were chosen such that they cover the entire network. Each ring-site
contains one 16 wavelength ring.
For the 9-node network, 3 traffic matrices were considered. The matrix m1 is a
uniform random demand matrix between 0 and 1 wavelengths. This demand can be fully
accommodated in the existing ring infrastructure, as can be seen in Figure 3, Figure 3 shows
the throughput (i.e. the percentage of demand that can be routed) obtained with the different
algorithms. Both the shortest path first (SPF) heuristic and the ILP algorithm with K shortest
paths yield a throughput of 100%, even for K=1. When K=1, there is only one possible path
for each connection, thus the only decision the ILP algorithm makes, is whether to route the
connection along this path or not. For the matrix m1, all connections can thus be routed along
the shortest path. This is not the case for the matrix m2, which is a uniform random demand
matrix between 0 and 2 wavelengths. In this case, both the SPF and the ILP algorithm with
K=1 yield a throughput of 97.8%, while the ILP algorithm with K≥2, yields a throughput of
100%. Thus, by only considering the shortest and second shortest path for each connection,
the maximum throughput can be obtained. This can also be witnessed for m3, a uniform
random demand matrix between 0 and 4 wavelengths. In this case, the total demand is too
large to be accommodated in the existing ring infrastructure. However, by using the SPF or
ILP algorithm with K=1, a throughput of 67.5 % can be reached. With K≥2, the throughput
increases to 70%.
CHAPTER 6178
0%
20%
40%
60%
80%
100%
m1 m2 m3
Throughput
SPF
ILP (K=1)
ILP (K=2)
ILP (K=3)
Figure 3: Throughput in 9-node network
In the 16-node network, similar results can be witnessed as shown in Figure 4 and
Figure 5. While the demand matrix m1 (0 or 1 wavelength) can still be fully accommodated
(using the ILP algorithm with K≥2) in the 16-node network, this is no longer possible in the
32-node network. In Figure 5, we witness that both demand matrices can not be set up
completely. The SPF algorithm performs slightly worst than the ILP algorithm with K=1.
Although the SPF algorithm recalculates the path in each iteration, depending on the free
capacity, and thus has additional flexibility (compared to the ILP algorithm with K=1), the
choice of which shortest path to set up is still quasi-randomly compared to the ILP algorithm
which makes an optimal assignment. Overall, the SPF algorithm yields results in close range
with the optimal result from the ILP algorithm. Furthermore, the ILP algorithm tends to
produce optimal results even for K=2, thus considering only the shortest and second shortest
path for each connection.
PLANNING OF INTERCONNECTED WDM RINGS 179
0%
20%
40%
60%
80%
100%
m1 m2
Throughput
SPF
ILP (K=1)
ILP (K=2)
ILP (K=3)
Figure 4: Throughput in 16-node network
0%
10%
20%
30%
40%
50%
m1 m2
Throughput
SPF
ILP (K=1)
ILP (K=2)
ILP (K=3)
Figure 5: Throughput in 32-node network
6.6 Ring dimensioning
Ring dimensioning starts from a given set of ring-sites. On some ring sites, some
legacy OCh-DPRings can already be present and some traffic can already be routed. Next to a
traffic matrix, the fact whether drop & continue is used or not, is given as input. The network
dimensioning process now has to make sure that all traffic is routed and protected. If the
current ring infrastructure does not suffice, or no rings are present yet, stacked rings have to
CHAPTER 6180
be added on the given ring-sites, in order to accommodate all the traffic at a minimal
installation cost.
Two different dimensioning algorithms have been worked out, both relying on the
equivalent network.
6.6.1 Integer linear programming
In a similar way as the ring routing problem, also the ring dimensioning problem can
be stated as an integer linear program. This time, also the amount of OCh-DPRings x(r) to be
placed on the ring-sites r∈R is a decision variable. All new OCh-DPRings to be added have W
wavelengths. The amount of free capacity on the legacy rings on each ring-site is given as
c(r). Each demand dij can again be routed along a large number of possible paths pij∈Pij,
consisting of a number of interconnected ring-sites. If ring-site r∈R is contained in path
pij∈Pij, then q(pij,r) = 1 (otherwise 0). The set of paths Pij we use for each pair of offices can
again be determined by searching for the K-shortest paths [25] between the corresponding
nodes in the equivalent network. The integer decision variable y(pij) represents the amount of
traffic that is routed on path pij∈Pij.
The ring dimensioning problem can thus be formally stated as:
Minimize:
)(.1),(..)().,(.)(. ij
Ni
ij
NjPpRr
ij
inter
Ni
ij
NjPp
ijij
Rr
route
r
Rr
ring
rpyrpqCpyrpqCrxC
ijijijij ∈
>
∈∈ ∈∈
>
∈∈∈∈
ö
ç
è
æ−++
α
Subject to:
ij
Pp
ij dpy
ijij
=
∈
)( ∀=i∈N, j∈N | j > i
)()(.)().,( rcrxWpyrpq
Ni
ij
NjPp
ijij
ijij
+≤
∈
>
∈∈
∀=r∈R
y(pij) integer and ≥ 0 ∀=i∈N, j∈N | j > i
x(r) integer and ≥ 0 ∀=r∈R
The objective function consists of 3 terms. The first term accounts for the fixed ring
cost contribution and is proportional with amount of rings x(r) on each ring-site r. The second
term represents the cost of routing wavelengths on the rings, and is proportional with the
amount of traffic y(pij) routed over paths pij=that use ring-site r (if q(pij,r) = 1). The third term
adds the cost of interconnecting wavelengths between two rings. This cost is proportional
with the amount of traffic y(pij) routed over paths pij and the amount of ring interconnections
in this path. This amount of ring interconnections equals the amount of rings r in path pij
minus one. The factor α in this third term allows to express whether interconnection between
two rings takes place in one or two nodes (to allow drop & continue). In the former case α =
1, while in the latter case α = 2. Note that the optimization with and without drop & continue
can give different results. First of all, the structure of the equivalent network might be
different in case of drop & continue, because certain rings are only interconnected through
PLANNING OF INTERCONNECTED WDM RINGS 181
one node. Second, the cost structure is different, such that also the weights of the links in the
equivalent network are different, potentially leading to different routes.
The first constraint states that the amount of traffic routed between each two
individual nodes i and j should be equal the total amount of traffic dij to be routed between
both nodes, such that all traffic is fully routed in the network. The second constraint says that
the amount of traffic y(pij) routed over paths pij=that use ring-site r (if q(pij,r) = 1) should not
exceed the remaining capacity c(r) on this ring-site plus the additional capacity of the newly
installed rings x(r) with W wavelengths. The last two constraints are trivial.
The above model can be implemented in an ILP solver such as CPLEX [26] to solve
the ring dimensioning problem. Such an ILP solution method is suitable for optimizing
medium sized problems and when the solution time is not critical. For larger problem
instances, or when fast optimization is required, it is better to use a heuristic solution method,
as described in the next section.
6.6.2 Heuristic solution method
When the ring dimensioning algorithm is used as a phase in the ring identification
process (see section 6.7), it has to be executed a large number of times, in order to evaluate a
large set of possible ring combinations. Therefore, it is of paramount importance that the
routing algorithm is sufficiently fast. In this section we present a fast and simple heuristic,
that has shown to yield close to optimal results, and can thus be used for this purpose. The
heuristic solution method is shown in Figure 6, and explained below.
Route traffic along
the shortest paths
Determine next ring r
with lowest utilization
Find a connection c
on the stack of ring r that
can be rerouted with the
least additional cost
reroute c
yes
If ring still contains traffic
yes
If new ring topology
without r is cheaper
than previous one
no
Restore the rerouted
connections on the rings
no
Remove ring r
no
yes
Figure 6: Heuristic solution method
The heuristic starts by routing the connections dij along their shortest path in the
equivalent network. After all connections have been routed we can determine the amount of
traffic routed on each ring-site. From this we can deduct the amount of stacked rings required
at each ring-site, as well as the utilization of each ring. In case multiple stacked rings are
CHAPTER 6182
present at a ring-site, all rings have a utilization of 100%, except for the last ring in the stack,
which might have a lower utilization. The heuristic now attempts to eliminate some of the
rings with low utilization ratios, by trying to reroute connections on this ring (or on any other
ring of the stack) on a different route, not containing this ring-site. On the stack of rings
corresponding with the ring with the lowest utilization ratio, we start with the connection that
can be rerouted on an alternative route, with the least additional cost compared to the cost of
its original route. The new route is calculated in the equivalent network (along the shortest
path), in which we first tear down the original route (and adapt the capacities), and from
which we exclude the current ring and all other rings that have no free capacity left. This
process of rerouting connections is repeated until one ring can be removed from the
considered stack, or until no more connections can be rerouted. In the first case the ring is
removed if the elimination of the ring leads to a cheaper overall solution. If the removal does
not lead to a cheaper solution, or in the second case, we restore the rerouted connections on
the current ring and do not remove the ring. In both cases the algorithm is continued by
searching for the ring with the next lowest utilization, and again we try to reroute connections
on this ring in the same way. This process is repeated until all rings are evaluated. Note that
the utilization ratios are recalculated in each iteration, because the rerouting of traffic of
removed rings also affects the traffic on the other rings.
6.6.3 Results
We compare results for the above-described dimensioning algorithms for the
following networks:
• a 9-node network, containing 4 ring-sites
• a 16-node network, containing 6 ring-sites
• the same 16-node network, containing 3 ring-sites
• a 32-node network, containing 12 ring-sites
• the same 32-node network, containing 6 ring-sites
For the 16-node and 32-node network, two ring constellations are considered to study
the impact of the amount of rings and ring size. In case the network is constituted of a small
set of ring-sites, these ring-sites typically span a large number of offices. In case the network
is composed of a large set of rings, the ring-sites are typically smaller. For the 9-node
network, the traffic matrix m3 (0-4 wavelengths) from section 6.5.3 has been chosen, because
this matrix can not be accommodated with one ring on each ring-site, and thus requires
multiple stacked rings on some ring-sites. For the same reason the matrix m2 (0-2
wavelengths) has been chosen for the 16-node network and matrix m1 (0-1 wavelength) for
the 32-node network.
In Figure 7 the cost of the 9-node network design is represented for each of the
dimensioning algorithms. The results are benchmarked against a simple dimensioning
algorithm, which routes all traffic along the shortest path in the equivalent network, and
dimensions the rings accordingly. As can be seen, the developed optimization algorithms
result in cost savings of 5% compared to the shortest path design. In addition, the heuristic
yields the same results as the optimized ILP algorithm, and this in a much shorter timeframe.
The ILP algorithm with K=2 yield the same result as the ILP algorithm with K=3 (and also for
higher values of K). This means that the optimal result can be obtained by only considering
the shortest and second shortest path in this example. The use of drop & continue results in an
PLANNING OF INTERCONNECTED WDM RINGS 183
extra cost of 1.7%, due to the additional ring interconnection cost. The amount of rings
needed and the ring dimensioning does not change when using drop & continue in this
example.
18000
18500
19000
19500
20000
20500
Shortest
path
Heuristic ILP (K=2) ILP (K=3) Shortest
path
Heuristic ILP (K=2) ILP (K=3)
Cost
Without drop & continue With drop & continue
Figure 7: Dimensioning of the 9-node network
For the 16-node network, both with 6 and 3 ring-sites, all dimensioning algorithms
yield the same result (see Figure 8 and Figure 9). Thus, in this case shortest path routing in the
equivalent network leads to the optimal ring dimensioning. Again a small difference between
the result with and without drop & continue is witnessed due to the extra cost of ring
interconnection. The 6 ring-site example turns out to be about 25% cheaper than the 3 ring-
site example. This clearly illustrates the importance of good ring identification. If the ring-
sites are chosen badly, optimal dimensioning will not help that much in reducing the overall
cost. In general we have witnessed that a ring design based on a large set of relatively small
ring-sites, results in better results than when using a small set of relatively large ring [6].
Although in a network with larger rings, less inter-ring traffic is present, and thus a lot of
traffic can be carried on a single ring, the total cost is higher. This is because connections
typically have long protection paths on the rings, thus making inefficient use of the network
capacity. Even in a single ring, a connection between neighbouring nodes takes up capacity
on the entire long ring. In addition the connections that do require to be routed on multiple
rings, consume capacity on almost the entire network. In a network with smaller rings, the
protection paths on each ring are relatively short, and even if a connection is transported on
multiple rings, the total amount of protection capacity remains reasonable. However, it could
be noted that one or two large 'overlay' rings might be beneficial for transporting connections
from one end of the network towards the other end in a network composed of smaller rings.
CHAPTER 6184
40000
40500
41000
41500
42000
42500
43000
43500
44000
Shortest
path
Heuristic ILP (K=2) ILP (K=3) Shortest
path
Heuristic ILP (K=2) ILP (K=3)
Cost
Without drop & continue With drop & continue
Figure 8: Dimensioning of the 16-node network with 6 ring-sites
56000
56200
56400
56600
56800
57000
57200
57400
57600
57800
Shortest
path
Heuristic ILP (K=2) ILP (K=3) Shortest
path
Heuristic ILP (K=2) ILP (K=3)
Cost
Without drop & continue With drop & continue
Figure 9: Dimensioning of 16-node network with 3 ring-sites
In Figure 10 and Figure 11, we present the results of the different dimensioning
algorithms for the 32-node network. The heuristic yields results that are better than those
obtained through shortest path routing and in close range with the results obtained through
integer linear programming. Again, the ILP algorithm with K=2 (thus only considering the
shortest and second shortest path) gives the same results than for larger values of K. In this
case, the cost difference between networks employing drop & continue and not, is larger than
in the previous examples. This is because the ring constellation is such that the routes used
without using drop & continue, can not be used for drop & continue, because certain rings
PLANNING OF INTERCONNECTED WDM RINGS 185
along the route are not interconnected through at least two offices. Thus in case drop &
continue is used, some connections have to take different routes, leading to a different (i.e.
more expensive) ring design. In addition, part of the extra cost also comes form the ring
interconnection cost. Again, the design based on a large set of relatively small rings results in
a lower overall cost, than when using a small number of larger rings.
100000
110000
120000
130000
140000
150000
160000
Shortest
path
Heuristic ILP (K=2) ILP (K=3) Shortest
path
Heuristic ILP (K=2) ILP (K=3)
Cost
Without drop & continue With drop & continue
Figure 10: Dimensioning of 32-node network with 12 ring -sites
120000
125000
130000
135000
140000
145000
150000
155000
160000
165000
170000
Shortest
path
Heuristic ILP (K=2) ILP (K=3) Shortest
path
Heuristic ILP (K=2) ILP (K=3)
Cost
Without drop & continue With drop & continue
Figure 11: Dimensioning of 32-node network with 7 ring-sites
Besides the fact that smaller rings result in a lower overall network cost, they also
help in increasing the availability of the network. This is reflected in the expected loss of
CHAPTER 618
6
traffic (ELT), which is the amount of traffic that the network is expected to lose per year due
to failures [23]. We express ELT in STM-16 hours per year, assuming each wavelength
transports one STM-16 signal (=2.5 Gb/s). We have calculated the ELT the following failure
scenarios:
- link failures only,
- link failures including optical amplifier (OA) failures,
- link, OA and OADM failures.
In each case, ring interconnection with and without drop & continue has been
considered. The considered failure probabilities of the network elements are those of Table 1.
The ELT values for the 16-node and 32-node network (with the different ring
constellations), are represented in Figure 12 and Figure 13 respectively. From this we can
indeed conclude that a network composed of a large amount of small rings results in a lower
ELT, and thus a higher network availability. This is particularly evident when only
considering link (+OA) failures, as multiple small rings can better survive from multiple link
failures in the network, because the chance is higher that the different link failures occur in
different rings. When considering both link and OADM failures, the effect is less strong
because an OADM failure involves the loss of all traffic originating-terminating in that
OADM, which is traffic that can not be recovered any way. This explains the much higher
ELT values for the node failures (shown in the right axis of the figures) and the small
differences between the networks with different ring sizes. In fact the difference in ELT for
link and OADM failures between large rings and small rings is very small when not
considering drop & continue. Indeed, larger rings can survive from less multiple failures, but
on the other hand, they involve less ring interconnections (single points of failure in case no
drop & continue is used). When drop & continue is used the difference in ELT between small
and large rings is more apparent. The overall effect of drop & continue on the ELT values is
remarkable, and increases for the larger network. When considering both link and node
failures, the ELT almost doubles when not using drop & continue. As drop & continue does
not require any additional ring capacity for OCh-DPRings, it can thus provide a substantial
increase in availability at a relatively low cost.
PLANNING OF INTERCONNECTED WDM RINGS 18
7
0
5
10
15
20
25
30
35
6 rings 6 rings + D&C 3 rings 3 rings + D&C
ELT (STM-16 h/y) for link (+OA) failures
0
50
100
150
200
ELT (STM-16 h/y) for link, OA + OADM failures
Only link
failures
Link + OA
failures
Link, OA +
OADM failures
Figure 12: Availability (ELT) for 16-node network
0
10
20
30
40
50
60
70
80
90
100
12 rings 12 rings + D&C 7 rings 7 rings + D&C
ELT (STM-16 h/y) for link (+OA) failures
0
100
200
300
400
500
600
ELT (STM-16 h/y) for link, OA + OADM failures
Only link
failures
Link + OA
failures
Link, OA +
OADM failures
Figure 13: Availability (ELT) of 32-node network
In conclusion, we have developed an ILP solution method based on the K shortest
paths, that gives good results for low values of K. Furthermore, a simple and fast heuristic
optimization method was developed that performs in close range with the optimal result
yielded by the ILP algorithm. The heuristic is sufficiently fast to be used as a phase in the
more complicated ring identification process, during which it has to be executed a large
number of times. When comparing routing in a network with a limited set of large rings or a
large set of smaller rings, the latter architecture is preferred both from a cost and availability
CHAPTER 6188
point of view. Also the impact of drop & continue on the overall network availability is
substantial and increases as the size of the network grows.
6.7 Ring identification
In the previous section, we already witnessed the importance of selecting an
appropriate set of ring-sites. In practice, the amount of possible ring-sites is enormous and the
number of possible combinations of ring-sites is even higher. Thus, a methodological
approach is needed to identify the best set of ring-sites for a given network topology, traffic
matrix and cost function. Due to the high amount of possible solutions, an exact optimization
method (e.g. using integer linear programming) would only work for very small examples.
Therefore, a heuristic solution approach is more suitable.
In this section, we present a pragmatic heuristic solution method that is suitable for
ring identification in large-scale networks. The heuristic is constituted of 3 main phases, as
depicted in Figure 14.
• In the ring generation phase, all possible ring positions (i.e. ring-sites) in the underlying
meshed network topology are generated, taking into account certain constraints. One
obvious constraint is the ring length. In order to avoid transmission impairments and to
minimize the protection switching delay, the physical ring length can be limited. Also
the amount of OADMs residing in the ring can be constrained for the same reasons or
due to the implementation of a signaling protocol on the ring (e.g. in SDH, only 16
ADMs can be supported on a ring, because of the limited addressing space in the
overhead).
• The goal of the subsequent ring pre-selection phase is to trim down the potentially
enormous amount of rings generated in the first phase. This phase selects the best-suited
rings from the ring generation phase, based on traffic characteristics. Handing over a
limited amount of rings to the final ring optimization phase, leads to faster optimization,
at the penalty of a potentially worse design, because less ring-sites can be considered.
• The final ring-optimization phase defines the best ring combination from the pool of all
selected rings from the previous phase. Starting from an initial ring definition, changes
can be made in an iterative way (using a heuristic search algorithm such as tabu search),
such to end up with the best ring combination. During each iteration, the different ring
combinations need to be evaluated using a dimensioning algorithm (see section 6.6). In
this phase, certain bounds may be specified regarding the minimum and maximum
amount of ring-sites allowed in the network.
PLANNING OF INTERCONNECTED WDM RINGS 189
Ring generation
Ring pre-selection
Ring identification
Ring dimensioning
Ring optimization
Max. ring length
Min. & max. OADMs/ring
Min. & max. number
of ring-sites allowed
Figure 14: Ring identification process
6.7.1 Ring generation in an existing topology
Before we can start combining rings, they first have to be generated, based on the
underlying meshed topology. In this section, we present two complementary algorithms for
generating all the possible ring-sites in an existing topology.
6.7.1.1 Generating rings containing a given node
A first question we try to solve is: how to determine all ring-sites containing a certain
node? Later on, we will use this algorithm to identify all ring-sites in the entire network.
Besides the network topology and the given node, the other input parameters are the
maximum ring length and the maximum amount of OADMs allowed on each ring.
The generation of all possible ring-sites containing a certain node, can also be looked
upon as the generation of all paths starting from this node and terminating in this node, while
not traversing any other node more than once (such a closed path is also called a loop). This is
also the main idea behind our approach, of which a schematic overview is shown in Figure
15. We start by generating a path by adding the given node (i.e. the starting node) to this path
and will gradually add other nodes to the path until we find a loop. First, we look for the
adjacent nodes to the starting node. An adjacent node is a node that can be reached from the
current node via one single link in the underlying network topology. From the moment one
such adjacent node is found, it is added to the path. Next, we will continue in the same way by
searching for the adjacent nodes of this last node of the path. Such a search, which looks for
one adjacent node of the last found node at a time, is called a depth-first search. To avoid that
the same node is encountered twice, we always verify whether the adjacent node is not yet
contained in the current path, before adding it to the path. This process is continued until one
of the following conditions is met:
• The starting node is reached. This means a ring-site has been discovered. This ring-site
will be added to the list of found ring-sites.
• There are no more adjacent nodes that are not yet contained in the current path.
CHAPTER 6190
• The amount of nodes in the path is equal to the maximum amount of OADMs allowed
on a ring, or the current length of the path exceeds the maximum allowed fiber length of
the ring.
In each of these conditions, the last node in the path is removed (e.g. in case the
starting node was reached, we remove the starting node). Thus the last node in the current
path becomes the last-but-one node from the previous path. From this node, we continue the
search for an adjacent node. The node that has just been removed is not allowed to be added
to the path in the next iteration (but can be added again in subsequent iterations). This process
is repeated until all adjacent nodes of the starting node have been subject to such a depth-first
search. The process can be made more efficient by eliminating the link between the starting
node and its adjacent node, each time this adjacent node has been subject to a complete depth
first search. In this way, it is also avoided that the same ring-site is generated twice.
n = start node
Add n to path
Find next adjacent
node n’ to n
Remove n
from path
If path
empty
n = last node
from path
End
If n’ is contained
in path no
yes
yes
If # nodes in path =
max. # nodes per ring
or path length >
max. ring length
no
yes
If n’ = start node
no
Ring found!
Add ring to ring list yes
Add n’ to pathn = n’
no
found
not
found
Figure 15: Generation of rings containing a given node
As an example, we generate all the ring-sites in the 7-node network depicted in
Figure 16, containing node 6. The maximum amount of nodes and the maximum ring length
is fixed at a very high value such that all rings will be found. In Table 2 we represent all the
rings (by their nodes) that were generated, sorted by the adjacent node of the start node (i.e.
node 6) which is first encountered in the depth first search. The first iteration of the depth first
search tree is shown in Figure 17. In Table 2, it is shown that the most ring-sites are generated
through the first adjacent node (i.e. node 4). Indeed 9 out of all 16 rings are generated through
this first adjacent node. Through the second adjacent node (i.e. node 5), just 3 ring-sites are
found, and the third adjacent node (i.e. node 3) only contributes to 2 more ring-sites. As
expected, no more ring-sites are generated through the last adjacent node (i.e. node 7),
because all ring containing node 7 have been found before.
PLANNING OF INTERCONNECTED WDM RINGS 191
1
3
5
2
6
4
7
Figure 16: Example network
Node 4 Node 5 Node 3 Node 7
6-4-7-3-5
6-4-7-3
6-4-7
6-4-2-3
6-4-2-3-5
6-4-2-3-7
6-4-2-1-3
6-4-2-1-3-5
6-4-2-1-3-7
6-5-3
6-5-3-2-4-7
6-5-3-1-2-4-7
6-5-3-7
6-3-2-4-7
6-3-1-2-4-7
6-3-7
-
Table 2: Ring-sites starting from node 6
6
45 37
72
36
65 21
Figure 17: Depth-first search tree
6.7.1.2 Generating all possible rings in a given network topology
We can use the algorithm described in the previous paragraph to generate all possible
ring-sites in the existing network topology. It is sufficient to apply this algorithm iteratively to
all nodes of the network to find all ring-sites. Such a straightforward approach is however far
from efficient because each ring-site is found as many times as it contains nodes. To avoid
this and improve the computational efficiency, we can temporarily delete each node from the
network after all ring-sites containing this node have been generated. This will result in each
CHAPTER 6192
ring-site being found only once, because once all rings containing a node have been found,
this node is deleted from the network and this node can not be used in other rings generated
from a different starting node. In addition it is guaranteed that each node will be found,
because a node is only removed after all ring-sites containing this node are found. As the
algorithm progresses and more nodes in the network have been tackled, the network will
gradually contain less nodes and less ring-sites will be generated through the remaining
nodes. The most ring-sites will thus be found at the start of the algorithm. When only 2 more
nodes remain in the network, no more rings will be found because each useful ring contains at
least 3 nodes.
An example is shown in Table 3, which shows all ring-sites that were found in the
network of Figure 16, sorted by the node through which they were generated. The algorithm
starts in node 1 through which it finds 7 of the 19 ring-sites. The subsequently evaluated
nodes are node 2 and 3.
Through node 3, 5 more ring-sites are generated, which is a relatively high amount,
considering that at this point two nodes have already been removed. This is due to the high
node degree of node 3. If the algorithm would have started through node 3, 18 out of 19 rings
would have been found in the first iteration, and only one more ring in the subsequent
iterations.
The last ring is found through node 4. This can easily be seen by removing nodes 1, 2
and 3 from the network in Figure 16. No more rings are found through nodes 5, 6 and 7.
Node 1 Node 2 Node 3 Node 4 Node 5 Node 6 Node 7
1-2-4-6-3
1-2-4-6-5-3
1-2-4-6-7-3
1-2-4-7-6-3
1-2-4-7-6-5-3
1-2-4-7-3
1-2-3
2-4-6-3
2-4-6-5-3
2-4-6-7-3
2-4-7-6-3
2-4-7-6-5-3
2-4-7-3
3-6-4-7
3-6-5
3-6-7
3-5-6-4-7
3-5-6-7
4-6-7 - - -
Table 3: All ring-sites in the network
6.7.2 Ring pre-selection
The ring generation routine does not take into account the traffic matrix, but
identifies rings only on the basis of the fiber topology of the network. When the network is
very dense (i.e. when a lot of links exist) or when the constraints for the ring generation
routine are not very limiting, the amount of possible ring-sites is very large. To restrict the
amount of possible rings (to speed up the optimization) we only wish to retain those rings
which have good traffic characteristics. Therefore, all possible ring-sites are evaluated on the
basis of intra-ring traffic. The ring-sites capable of capturing the highest amount of traffic
within the ring, relative to the amount of nodes on the ring, are the best candidates to be
included in the final interconnected ring design. Indeed, traffic routed across multiple rings
(i.e. inter-ring traffic) consumes capacity on all rings passed, whereas intra-ring traffic only
uses capacity on a single ring.
PLANNING OF INTERCONNECTED WDM RINGS 193
We choose to keep the best K rings per node. Alternatively, we could also choose an
equivalent amount of best rings across the entire network (not bound to specific nodes), but in
the latter case we could end up with a lot of rings in one region of the network (with high
traffic) and no rings at all in another region (with low traffic). Therefore, to spread the
possible ring-sites evenly over the network, the K best rings per node are kept. Therefore,
each node is evaluated and the best K rings not yet selected through other nodes are kept. As
such, if N is the amount of nodes of the network, the amount of selected rings equals K.N. The
parameter K can be tuned, to limit the amount of rings that are considered in the optimization
process, and thus to limit computation time. If K is large, a lot of possible rings are used in the
optimization and the optimization will also take longer because more ring combinations have
to evaluated. If K is small, only a small number of rings is retained per node (the ones with
the best intra-ring traffic characteristics) and the optimization will be faster, at the expense of
a potentially less optimal result.
6.7.3 Ring optimization
After we have restricted all possible rings to a limited sub-set, we now try to choose
an optimal combination of rings, such to cover all nodes in the network (not necessarily all
links) and to transport all traffic on these rings. The amount of rings to be identified can be
restricted between a minimal and maximal value, because of cost, management or reliability
considerations. If this minimal and maximal value are equal, the amount of rings to be
identified is fixed. If the maximal value is larger than the minimal value, the amount of rings
is to be optimized by the algorithm between both values. Two algorithms for ring
optimization have been implemented: a heuristic algorithm, based on tabu search and an
exhaustive search algorithm, which tries all possible ring combinations. The exhaustive
search algorithm gives the optimal result, but requires brute force, thus extensive calculation
time and can not be used for larger examples.
6.7.3.1 Tabu search heuristic
Tabu search [27] is a heuristic search procedure known from operations research,
capable of solving a wide range of optimization problems. Tabu search starts from a non-
optimal initial solution to the studied problem, and evaluates all so-called neighbouring
solutions, by making some small changes to the current solution. The best neighbouring
solution is chosen as the new current solution, and this process is repeated a certain number of
times. To avoid backtracking and getting trapped locally in a sub-optimal solution, the last
made changes are stored in a so-called tabu list, and these changes may not be considered
anymore in the next iterations (they are called tabu). The time during which these changes
remain in the tabu list is also called the tabu tenure.
The tabu search process applied to ring optimization is depicted in Figure 18.
The initial solution (i.e. ring combination) is obtained by ranking all pre-selected rings
according to the amount of intra-ring traffic relative to the amount of nodes on the ring. A
depth-first search through this list of rings is performed such to obtain an initial ring cover
that includes all nodes of the network and is within the bounds of minimal and maximum
amount of rings allowed in the network.
The evaluation of this initial ring combination (and other ring combinations, later on)
is performed by dimensioning these rings, such that all traffic can be transported. This
CHAPTER 6194
dimensioning incurs determining the amount of stacked rings to be placed on each ring-site. It
is essential that the used dimensioning algorithm is sufficiently fast, due to the fact that it has
to be performed a large number of times during each iteration, and for a large number of
iterations. For this purpose we use the heuristic dimensioning algorithm proposed in section
6.6.2.
After the initial ring combination has been evaluated, different ring combinations can
be explored using tabu search, trying to reduce the overall cost. For our application of ring
optimization, a change to the solution is made by adding or removing a ring-site, or by
replacing one ring-site by another ring-site. In making such changes, it has to be ensured that
the minimum/maximum amount of allowed ring-sites is not violated (e.g. when the minimum
and maximum amount of ring-sites are equal, it is only possible to exchange ring-sites and not
to add/remove rings). In addition, a lot of these changes will result in a ring combination that
can not cover all nodes of the network. In this case, the cost of this ring combination does not
need to be evaluated of course.
Initial solution
(Initial ring design)
Perform new change
(add/remove/exchange ring)
Evaluate new solution
(ring dimensioning)
Keep best new solution
(keep new ring design)
Update tabu list
Tabu list
Figure 18: Tabu search applied to ring optimization
Tabu search thus moves from one solution to a neighbouring solution by adding
and/or removing rings that are not tabu. The tabu time of the ring-sites can be different for a
ring-site, depending on whether is was added or removed. We do not want the recently added
rings to hold the tabu status for too long, otherwise they can not be removed. Indeed, the
amount of rings to be potentially removed is a lot smaller than the large pool of rings to be
potentially added. Therefore, the tabu-time for the removed rings can be larger (i.e. the time
before a removed ring becomes a candidate to be added again), to favour other solutions and
to prevent cycles in the solution trajectory.
The tabu search algorithm can be halted after a pre-described number of iterations.
The larger this number, the more iterations, the longer the optimization takes, but the more
chance of finding a better solution. Typically however, some good solutions can already be
found after a small number of iterations (see further). Alternatively, we can also stop the
optimization after a number of iterations during which no further improvements have been
made.
If no better solutions are found after a certain number of iterations, the optimization
process can also try to perform intensification (this means looking for better solutions in the
PLANNING OF INTERCONNECTED WDM RINGS 195
neighborhood of the best solutions found so far) or diversification (this means looking for
better solutions in previously unexplored regions of the solution space) [27]. Intensification is
implemented by remembering the 5 best solutions found so far, and constructing a new ring
network based on the rings of these 5 solutions in the same way as the initial solution was
constructed. Diversification is implemented by remembering the rings that are least used in all
the solutions so far and by combining these rings into a new solution, again in the same way
as the initial solution was constructed. The number of iterations without better solutions after
which to perform intensification and diversification is to be defined.
6.7.3.2 Exhaustive search
The exhaustive search algorithm evaluates all possible ring combinations containing
rings between the minimal and maximal number of rings that is specified. The routing
algorithm used to evaluate the ring combination is again the heuristic described in section
6.6.2. Depending on the amount of rings considered and the minimal and maximal number of
ring-sites allowed in the solution, exhaustive search can result in a very large number of
evaluations. If R ring-sites are considered, and we allow a maximum of Rmax and a minimum
of Rmin ring-sites in the solution, the total amount of possible combinations is given by:
=−
=
max
min )!(!
!
R
RR ii
iRRR
R
C
The number R depends on the maximum size of the rings, and the amount of ring-
sites retained in the pre-selection phase. If R = 30 and Rmin = 6 and Rmax = 8, the total number
of possible ring combinations equals 7,888,725. Again, a large amount of these combinations
does not result in feasible ring combinations that cover all nodes of the network and thus do
not need to be dimensioned. Still, it is clear that the exhaustive search algorithm can only be
used for small problem instances, e.g. to benchmark the results of the tabu search heuristic.
6.7.4 Results
6.7.4.1 Performance of the tabu search heuristic
For relatively small examples it is possible to verify the results of the tabu search
algorithm with optimal results obtained through exhaustive enumeration of all possible ring
combinations. Whereas exhaustive enumeration will always yield the optimal result after
sufficiently long calculation, the tabu search algorithm searches in a much more intelligent
way such that a good (sub-optimal) solution can already be found after a limited amount of
iterations. A typical evolution of the best solution in each iteration of the tabu search
algorithm is shown in Figure 19 for a 16-node network with a uniform random demand matrix
between 0 and 2 wavelengths. The size of the rings was restricted to 10 nodes in the ring
generation phase. We selected 3 rings per node in the pre-selection phase and a minimum of 4
rings and maximum of 6 rings was required in the final ring optimization phase. This resulted
in 310,960 possible combinations for rings that had to be evaluated by exhaustive search
(which took more than 6 hours on a 166 MHz desktop PC). In contrast, 50 iterations of the
tabu search algorithm were performed in a few minutes and good results were already found
in the first iterations, which were within 2% of the optimal solution found through
enumeration. The optimal result was also found with the tabu search algorithm after 32
CHAPTER 619
6
iterations. Also for other examples, we found results close to the optimal results in short time
frames.
40000
41000
42000
43000
44000
45000
46000
47000
48000
1
4
7
10
13
16
19
22
25
28
31
34
37
40
43
46
49
Iterations
Cost
Figure 19: Evolution of tabu search algorithm
6.7.4.2 Impact of the main parameter settings
The main parameters to be set during ring design are the following:
- Maximum amount of nodes on the ring (N) during ring generation.
- The amount of rings selected per node (K) during ring pre-selection.
- Minimum/maximum amount of ring-sites (Rmin and Rmax) during final ring optimization.
The impact of the above parameters will be shown in this section. As we have
already shown it is possible to obtain good results after a limited number of iterations, we
limit the maximum number of iterations for all simulations to 50. The settings for the different
simulations are show in Table 4 for the 16-node network.
Simulation NR
min Rmax K
110463
210461
310443
410553
516463
Table 4: Parameter settings
The results for these simulations are shown in Figure 20.
PLANNING OF INTERCONNECTED WDM RINGS 19
7
40000
41000
42000
43000
44000
45000
46000
12345
Simulation
Cost
Figure 20: Results for different simulations
The main conclusions to be derived from these results are summarized below:
The main difference between simulation 1 and 2 is the amount of rings allowed per
node. When more rings are considered per node, a better result is obtained. However,
considering less rings per node, also gives acceptable results at much shorter calculation
times. This shows that it can be allowed to work with a limited set of rings with very good
traffic characteristics instead of working with a larger set of rings. Although working with a
larger set of rings can eventually yield better results as demonstrated, it has also shown to be
counterproductive in case only a limited number of optimization iterations is executed.
From simulations 3 (allowing only 4 rings) and 4 (allowing only 5 rings) it can be
concluded that the results become better when more ring-sites are allowed in the solution.
Hence the number of ring-sites should not be too small. More rings allow more intra-ring
traffic and thus minimization of the ring interconnection cost. This ring interconnection cost is
not only the cost of the interconnection itself but also the cost of using capacity on multiple
rings instead of using capacity on only one ring (in case of intra-ring traffic). In addition using
more rings typically also results in a higher reliability because additional rings allow to
recover from more simultaneous failures as shown in section 6.6.3.
Simulation 5 considers also larger rings with up to 16 nodes. This does not decrease
the cost of the eventual ring design however. Too large rings are thus not always interesting.
Although in a network with larger rings, less inter-ring traffic is present, and thus a lot of
traffic can be carried on a single ring, the total cost is higher as also shown in section 6.6.3.
This is because connections typically have long protection paths on the rings, thus making
inefficient use of the network capacity (e.g. it is better to have two different small rings for
transporting intra-ring traffic than two large stacked rings overlaying both small rings).
6.8 Comparison of ring and mesh based architectures
We can now use the above described algorithms to perform ring-based network
designs and compare these designs with mesh-based designs. We present results for a 16-
node/21-link and a 32-node/42-link network mentioned above. For both networks random
traffic matrices were generated, with the demand between two nodes varying between 0 and 2
wavelengths
CHAPTER 6198
The network has been designed for meshed path protection (using WP or VWP
OXCs), meshed link and path restoration (using VWP OXCs) and interconnected OCh-
DPRings (using OADMs and hardwired back-to-back ring interconnection) and to cope with
single link failures. For meshed path protection, the network was designed using a heuristic
based on the shortest pair of disjoint paths [29] and the wavelength assignment in case of WP
networks was done using a heuristic. For meshed restoration the working capacity was
assigned using a heuristic based on the shortest path algorithm of Dijkstra [28] and the spare
capacity was assigned using integer linear programming based on the K-shortest paths [25].
For interconnected OCh-DPRings, the rings were identified based on the tabu search
algorithm presented in section 6.7.3.1, and dimensioned using the heuristic from section 6.6.2.
Comparisons will be made based on link and node cost, and on the network
availability. The used cost values for rings are mentioned in section 6.3.1.6 and the failure
rates in section 6.3.1.7. For the cross-connects the used costs are presented in Table 5. The
OXC cost depends mostly on the size of the switching matrix. The OXCs are assumed 100%
non-blocking. The OXC cost has been estimated by the amount of basic switch elements
required to realize a VWP OXC in a clos configuration [30]. The failure rates for OXCs are
the same than those used for the OADMs (see Table 1).
Installation cost
OXC 8x8 fiber 640
OXC 16x16 fiber 1600
OXC 32x32 fiber 3840
OXC 64x64 fiber 8960
Table 5: Cost values OXCs
6.8.1 Cost comparison
First we compare the link and node cost of the different recovery schemes. These
costs are presented in Figure 21 for the 16-node network and in Figure 22 for the 32-node
network.
PLANNING OF INTERCONNECTED WDM RINGS 199
0
10000
20000
30000
40000
50000
60000
WP
protection
WP
protection
+tunability
VWP
protection
VWP link
restoration
VWP path
restoration
OCh-
DPRing
OCh-
DPRing
+D&C
Cost
Node cost
Link cost
Figure 21: Cost of different recovery schemes for 16-node network
0
50000
100000
150000
200000
250000
300000
350000
400000
WP
protection
WP
protection
+tunability
VWP
protection
VWP link
restoration
VWP path
restoration
OCh-
DPRing
OCh-
DPRing
+D&C
Cost
Node cost
Link cost
Figure 22: Cost of different recovery schemes for 32-node network
6.8.1.1 Link cost
As can be seen, the network design based on interconnected OCh-DPRings results in
the highest link cost. For the 32-node network, the use of drop & continue requires an
additional link cost of about 5%, because some connections have to take longer routes
(compared to the case without drop & continue) to ensure that all rings on the route have dual
node interconnectivity. The link cost of interconnected OCh-DPRings is about 20-25% more
expensive than that of path protection in a meshed network. The difference in link cost
between the WP and VWP path protection scheme is not that high. The WP scheme requires
about 15% more fibers compared to VWP to resolve wavelength conflicts in the 16-node
network. In the 32-node network this is only 5%, because a lot more fibers per link are
CHAPTER 6200
required in the 32-node network, making the wavelength assignment easier. When
wavelength tuneability at transmitter and receiver is provided, enabling a different wavelength
for working and protection path, the WP scheme requires less than 5% extra fibers compared
to VWP for both networks. As the fixed link cost is not that high compared to cost per
wavelength, the WP scheme results in a link cost increase of only 2-5% without tuneability at
transmitter and receiver and 1-2% with tuneability.
VWP link restoration yields a link cost about 25% lower than VWP path protection.
The link cost of path restoration compared to link restoration is even 15% lower, making this
the cheapest scheme with regard to link cost. When comparing VWP path protection with
VWP path restoration, about 35% of the link cost can be saved through sharing of protection
capacity.
6.8.1.2 Node cost
The node cost is lowest for the interconnected OCh-DPRing architecture due to the
relatively cheap OADMs (note that the cost of terminal multiplexers is included in the link
cost). The use of drop & continue requires a 10-15% higher node cost, because traffic is
exchanged at two nodes between two rings. The node cost for mesh path protection is the
highest, due to the expensive cross-connects, making this the most expensive option overall.
The difference in node cost between WP (both with and without tuneability at transmitter and
receiver) and VWP protection is small (less than 5%), which is partly due to the discrete node
cost used. For VWP only a few nodes tend to be smaller hence only a limited node cost
reduction is observed. As also the difference in link cost is small between WP and VWP
protection, the overall cost difference is also small. Note however that in this comparison the
same cost was used for a WP and VWP OXC, while it is clear that the VWP OXC installation
cost will be larger (because wavelength converters and more internal connectivity is required),
thus favoring the WP solution. On the other hand, the total cost of ownership, including
controlling and management could be lower for a VWP OXC since the link capacity can be
controlled and assigned independently, and no wavelength assignment protocols and
algorithm are required. Due to a lack of cost figures for these aspects, the comparison is
limited to installation cost. The difference in total cost between VWP meshed path protection
and interconnected OCh-DPRings is about 20%. The cost of the OXCs (as in Table 5) should
decrease by about 50% to make the total cost of the path protection approach competitive
with interconnected OCh-DPRings.
Link restoration has a node cost, about 50% less than path protection but still more
than double than that of interconnected OCh-DPRings. Path restoration has about the same
node cost as link restoration for the 16-node network, because due to the discrete sizes of the
OXCs, no OXCs can be reduced in size. For the 32-node network some OXCs can be reduced
in size and the node cost of path restoration is about 30% lower than that of link restoration,
but still much higher than interconnected OCh-DPRings. Because of the low link cost, path
restoration is the cheapest overall solution: about 10-20% cheaper than link restoration, 30%
cheaper than interconnected OCh-DPRings and 40-45% cheaper than path protection.
6.8.2 Availability comparison
A second means for comparing the different recovery schemes is the availability of
the network. As a comparative measure, we again use the expected loss of traffic (ELT),
PLANNING OF INTERCONNECTED WDM RINGS 201
which is the amount of traffic that the network is expected to lose per year due to failures
[23]. We express ELT in STM-16 hours per year (h/y), assuming each wavelength transports
one STM-16 signal (=2.5 Gb/s). We have calculated the ELT both for link failures only and
for link and node failures. The ELT figures versus the cost of the ring and meshed protection
schemes is depicted for both networks in Figure 23 and Figure 24 (for only link failures) and
in Figure 25 and Figure 26 (for both node and link failures). We did not calculate the ELT
values for restoration, because it requires extensive simulations considering manifold failure
scenarios and the result depends too much on the restoration algorithm assumed and the
amount of spare capacity that is available.
0
5
10
15
20
25
40000 45000 50000 55000 60000
Total cost
ELT (STM-16 h/y)
OCh-DPRing + D&C
OCh-DPRing
Mesh protection
Figure 23: ELT versus cost for 16-node network (link failures only)
0
20
40
60
80
100
120
140
160
180
250000 275000 300000 325000 350000
Total cost
ELT (STM-16 h/y)
OCh-DPRing + D&C
OCh-DPRing
Mesh protection
Figure 24: ELT versus cost for 32-node network (link failures only)
CHAPTER 6202
When only considering link failures, the meshed path protection scheme has the
worst availability. Especially in the 32-node network, the ELT is almost doubled compared to
the interconnected OCh-DPRings, because long paths are protected end-to-end in contrast to
the interconnected OCh-DPRings, where sub-paths per ring are individually protected. As
such, interconnected OCh-DPRings (even without drop & continue) allow to survive from
multiple link failures occurring in different rings and are therefore more reliable with regard
to link failures. With drop & continue, the interconnected OCh-DPRings can survive from
slightly more link failures (e.g. double link failures in one ring, affecting a link between both
gateway nodes and a different link), but the difference in ELT is not that large.
0
50
100
150
200
250
40000 45000 50000 55000 60000
Total cost
ELT (STM-16 h/y)
OCh-DPRing + D&C
OCh-DPRing
Mesh protection
Figure 25: ELT versus cost for 16-node network (link and node failures)
0
200
400
600
800
1000
1200
1400
250000 275000 300000 325000 350000
Total cost
ELT (STM-16 h/y)
OCh-DPRing + D&C
OCh-DPRing
Mesh protection
Figure 26: ELT versus cost for 32-node network (link and node failures)
PLANNING OF INTERCONNECTED WDM RINGS 203
When considering both node and link failures, the ELT values are considerably
larger (although the unavailability of the nodes is not that high). The main reason for this is
that when a node failure occurs, all traffic terminating in that node is irrevocably lost, no
matter what recovery scheme is used. In this case, the interconnected OCh-DPRing scheme
without drop & continue has the worst unavailability. This is because traffic routed over
multiple rings is unable to recover from failures of the OADMs in which ring interconnection
occurs. Therefore it certainly makes sense to use drop & continue when node failures come
into play. Indeed, the difference in ELT between interconnected OCh-DPRings with and
without drop & continue is significantly larger in case both node and link failures are
considered. Meshed path protection can also survive from single node failures, because main
and protection path are considered to be node disjoint. Consequently, meshed path protection
has a higher availability than interconnected OCh-DPRings without drop & continue in case
link and node failures are considered. Meshed path protection has only slightly higher ELT
values than interconnected OCh-DPRings with drop & continue for the 16-node network.
This is because most connections in the 16-node network are routed over a single ring, and
thus the influence of drop & continue is not that large. In the 32-node network, where more
connections are routed over multiple rings, the difference in ELT between meshed path
protection and interconnected OCh-DPRings with drop & continue is already more
significant.
Overall, the interconnected OCh-DPRings with drop & continue result in the best
combination of cost and availability. Considering the restoration schemes, link restoration
might be able to recover from multiple link failures, but it can not recover from node failures.
Therefore path restoration is expected to have a better availability. In combination with its
low cost, path restoration is thus an attractive alternative if fast recovery times can be
achieved and scalable large-size OXC can be implemented within the appropriate cost range.
By making sensitivity analysis on the obtained results, we can obtain boundaries on the OXC
cost range, to render the meshed architectures cheaper than the ring architecture, as shown in
[10].
6.9 Conclusion
In this chapter, the different tasks of the interconnected OCh-DPRing design problem
have been outlined, and solution methods, based on heuristics or optimal methods, have been
proposed.
For the ring routing problem, an efficient heuristic, based on an iterative application
of shortest path routing has been proposed. Next to the heuristic, an optimized integer linear
program based on the K-shortest paths was developed. We have shown that the heuristic
performs in close range to the ILP algorithm, and optimal solutions can be found with ILP
using small values of K.
For the ring dimensioning problem, we also developed a heuristic and an integer
linear programming algorithm. The heuristic has shown to yield near-optimal results in short
calculation times. By applying these dimensioning algorithms to sample networks with
different ring constellations, we have found that a network composed of a relatively large
amount of small rings has a lower cost and higher availability than a network composed of a
relatively small amount of larger rings.
CHAPTER 6204
For the ring identification process, a three phased solution method has been
developed, consisting of ring generation, ring pre-selection and ring optimization. The last
phase can use exhaustive search or can apply more intelligent search methods such as tabu
search. During this last phase, different ring combinations have to be evaluated using a fast
dimensioning algorithm. The heuristic dimensioning algorithm we have developed for this
purpose has proven very suitable for this. Overall, the tabu search heuristic yields very good
results, even after a limited amount of iterations.
Finally, the algorithms have also been used to compare ring-based designs with
mesh-based designs. We have shown that interconnected OCh-DPRings are competitive with
mesh protected designs in terms of both cost and availability. Mesh restorable designs can
yield lower costs, if economical large-scale OXCs can be manufactured.
PLANNING OF INTERCONNECTED WDM RINGS 205
6.10 References
[1] "Cross-connects for WDM networks", Session TuM, 25th Optical Fiber Conference (OFC'2000),
Baltimore (MD), March 2000.
[2] "OADM Workshop", Scheveningen (Holland), April 1998.
[3] L. Wuttisittikulkij, M.J. O'Mahony, "Design of a WDM network using a multiple ring approach",
Proceedings of IEEE Globecom'97, pp. 551-555, Phoenix (AZ), November 1997.
[4] B. Van Caenegem, P. Demeester, "Design of interconnected WDM ring networks", Proceedings of
ONDM'99, pp. 61-68, Paris (France), February 1999.
[5] A. Proestaki, M.C. Sinclair, "Interconnection strategies for dual-homing multi-ring networks",
Proceedings of ITC'16, pp. 169-181, Edinburgh (UK), June 1999.
[6] P. Arijs, P. Demeester, P. Achten, W. Van Parys, "Dimensioning of non-hierarchical interconnected
WDM ring networks", Proceedings of ONDM 2000, pp. 147-160, Athens (Greece), February 2000.
[7] M.J. Soulliere, "Interconnection of Optical Network Administrative Domains", T1X1.5 Contribution
98-118R1, September 1998.
[8] P. Arijs, W. Van Parys, B. Van Caenegem, P. Achten, P. Demeester, P. Lagasse, "Design of ring and
mesh based WDM transport networks", forthcoming in Optical Networks Magazine, 2000.
[9] P. Arijs, P. Demeester, W. Van Parys, P. Achten, R. Meersman, "Planning of survivable WDM
transport networks based on interconnected optical channel protected rings", Proceedings of NOC 2000,
Stuttgart (Germany), June 2000.
[10] W. Van Parys, P. Arijs, P. Demeester, "Cost boundaries for economical OXCs", Proceedings of
OFC'2000 (Baltimore, MA), Session ThO, March 2000.
[11] P. Arijs, W. Van Parys, P. Demeester, "Cost and availability comparison of WDM mesh and ring
network architectures", Proceedings of DRCN 2000, Munich (Germany), April 2000.
[12] P. Arijs et al., "Architecture and design of optical channel protected ring networks", to appear in Journal
of Lightwave Technology, January 2001.
[13] J. Shi, J.P. Fonseca, "Hierarchical self-healing rings", IEEE/ACM Transactions on Networking, Vol. 3,
No. 6, pp. 690-697, December 1995.
[14] S. Cosares, D.N. Deutsch, I. Saniee, O.J. Wasem, "SONET toolkit: A decision support system for
designing robust and cost-effective fiber-optic networks", Interfaces, Vol. 25, No. 1, pp. 20-40,
January-February 1995.
[15] J.B. Slevinsky, W.D. Grover, M.H. MacGregor, "An algorithm for survivable network design
employing multiple self-healing rings", Proceedings of IEEE Globecom'93, pp.1568-1573, Houston
(TX), November 1993.
[16] B. Doshi, P. Harshavardana, "Broadband network infrastructure of the guture: Roles of network design
tools in technological deployment strategies", IEEE Communications Magazine, Vol. 36, No. 5, pp. 60-
71, May 1998.
[17] P. Semal, K. Wirl, "Optimal clustering and ring creation in the network planning system PHANET",
Proceedings of Networks'94, pp. 303-308, Budapest (Hungary), September 1994.
[18] L.M. Gardner, I.H. Sudborough, I.G. Tollis, "NetSolver: A software tool for the design of survivable
networks", Proceedings of IEEE Globecom'95, pp. 926-930, Singapore, November 1995.
[19] G. Nemhauser, L. Wolsey, "Integer and combinatorial optimization", John Wiley & Sons, New York,
1998.
[20] H. Luss, M.B. Rosenwein, R.T. Wong, "Topological network design for SONET ring architecture",
IEEE Transactions on Systems, Man. and Cybernetics, Vol. 28, No. 6, pp. 780-790, November 1998.
CHAPTER 620
6
[21] A. Fumagalli et al., "Survivable networks based on optimal routing and WDM self-healing rings",
Proceedings of IEEE Infocom'99, New York (NY), March 1999.
[22] M. Herzberg, F. Shliefer, "Optimization models for the design of bi-directional self-healing ring based
networks", Proceedings of ITC'16, pp. 183-194, Edinburgh (UK), June 1999.
[23] D. Vercauteren, "Modeling and analysis of the integrity of telecommunications networks", Ph.D. thesis
(in Dutch), Ghent University, 1997-1998.
[24] R.K. Ahuja, T.L. Magnanti, J.B. Orlin, "Network flows: Theory, algorithms and applications", Prentice-
Hall, New Jersey, 1993.
[25] J. Y. Yen, "Finding the K Shortest Loopless Paths in a network", Management Science, Vol. 17, No.
11, pp 712-716, July 1971.
[26] ILOG Inc., CPLEX Division: Using the CPLEX callable library, 1997.
[27] F. Glover, M. Laguna, "Tabu search", Kluwer Academic Publishers, Boston, 1997.
[28] E. Dijkstra, "A Note on two problems in connection with graphs", Numeriche Mathematics, Vol. 1, pp.
269-271, 1959.
[29] J.W. Suurballe and R.E. Tarjan, "A quick method for finding shortest pairs of disjoint paths", Networks,
Vol. 14, pp. 325-336, 1984.
[30] P. Demeester, "Telecommunicatienetwerken A", course at Ghent University, 1999-2000.
CHAPTER 7
Topological planning of the access
network
7.1 Introduction
The previous chapters have focussed exclusively on planning problems in the
transport network. In this chapter, we will consider the access network, which also poses
challenging planning problems. Whereas the transport network is responsible for provisioning
large capacity connections between several tens or hundreds of offices, the access network
has to interconnect all hundred thousands or millions of individual subscribers to the network.
Hence, the costs involved in building out an access network are several orders of magnitude
larger than the costs related to the transport network. Whereas most incumbent operators
already have a legacy access network, using copper or coax technology, they face
considerable planning challenges when upgrading this existing access network for future
high-bandwidth services. Also new-entrant operators, building out a completely new access
network, often targeting only large business customers, are confronted with similar planning
problems.
Most of the costs in the access network are related to the installation of transmission
equipment. Because of the large amount of subscribers to be connected, a huge amount of
transmission equipment has to be installed. In addition, the installation costs in the access
network are much higher than in the transport network. In the transport network, rights of way
are relatively easy to obtain and the amount of required civil works is limited, because often
cables can be installed along railways, highways or public utility facilities. In contrast,
installation of transmission facilities in the access network typically requires extensive civil
works in cities, incurring difficult to obtain rights of way and very high costs. Therefore, one
of the main attention points in access network planning is minimization of the cost related to
installing transmission equipment.
In this chapter, we will focus on the topological planning of the transmission
facilities in the access network, in order to minimize the installation cost. We will briefly
discuss the different technologies and migration strategies for future access networks. Based
on the "fiber in the loop" migration scenario, we will then propose several access network
topologies, such as star, tree, ring and hybrid architectures. Different topological planning
algorithms will be proposed for each of these architectures. These different algorithms are
integrated in a geographical information system (GIS) to take detailed geographical
information into account during the planning process. The main results of this chapter can
also be found in [1][2][3][4][5][6].
CHAPTER 7208
7.2 Access network evolution
7.2.1 Current technologies
When looking at the installed base, we witness three important technologies used in
the public access network [7]:
Twisted copper pairs, being an analogue access line to the customer in a star
topology, originally installed for voice telephony with a 3.4 KHz bandwidth. At present,
about 700 million telephone lines are installed worldwide, mostly owned by incumbent
operators. In recent years the digitization of this network has commenced in order to support
(broadband) data services.
Coaxial access networks, originally installed for cable TV (CATV) broadcasting, are
analogue in nature and characterized by a tree and branch structure. A large downstream
bandwidth can be achieved but due to noise accumulation effects in the upstream path, the
upstream bandwidth is limited. Penetration rates of this access network vary widely from
country to country (e.g. about 99% in Belgium, but only 30% in the UK).
Wireless technologies form another type of competing access networks. Their main
advantage is their high speed of deployment. Terrestrial systems can be either fixed (e.g. point
to point microwave transmission) or mobile. For this rapidly growing “mobile market”, this
wireless technology is the obvious solution. In Europe the digital GSM standard has been
adopted, and will be succeeded by the UMTS standard. Also satellite systems can play a role
in the wireless access space. Although penetration rates vary largely, digital satellite TV is a
clear competitor for the CATV technology for broadcast applications. Due to he large latency,
such systems are less suited for interactive applications.
Also asymmetric technology architectures can be considered using one technology in
the upstream direction (e.g. twisted pair) and another technology in the downstream direction
(e.g. broadband wireless) to leverage existing infrastructure.
A schematic illustration of these access technologies is shown in Figure 1
implemented for the twisted pair, CATV (coax), mobile telephony and satellite (wireless)
networks. Bearing in mind that up to 80% of the network cost is in the access network (in the
case of twisted pair and coax), the optimal (re)use of the installed base in future evolution
scenarios is clearly of crucial economic importance.
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 209
LEX twisted pair
POTS
POTS
wireless
BTS
Mobile
Mobile Satellite
Satellite
HE coax
CATV
CATV
Figure 1: Access technologies installed today in the public network
7.2.2 Evolution towards broadband access networks
The liberalization of the telecommunications market and the creation and adoption of
new services, is creating a user pull both at the level of increasing customers per service and
increasing bandwidth needs per customer. As such, access networks have to evolve towards
multi-service broadband access platforms, in order to cope with these requirements. As
explained in the previous paragraph, optimal reuse of the installed infrastructure base is of key
economic importance. However, the existing infrastructure does not use the latest
technologies and has often been optimized for one specific service. This imposes important
limitations that have to be considered [8]. Twisted pairs can accommodate higher bandwidths
per customer with specific modulation techniques (e.g. ADSL, VDSL), although the
achievable bandwidth is inversely proportional with the distance to be traversed. Coax-based
access technologies (using cable modems) and wireless solutions (e.g. microwave technology,
LMDS, …) make use of a shared bandwidth medium, and thus the available bandwidth per
subscriber is inversely proportional to the number of active users.
To overcome the scaling issues inherent with current technologies, there is a need to
introduce more advanced and future proof technologies in the access network. However a
complete redesign of the existing access network based on new technologies is no feasible
business case. A more gradual evolution scheme retains the existing infrastructure in the so-
called last drop section (i.e. the last kilometer(s) to the customer). As such the access network
is divided in small-size islands of hundreds or thousands of customers, in which the existing
infrastructure can be reused for broadband purposes without major scaling issues. The
different islands are in turn connected using new infrastructure, typically optical fiber [9][10].
As such the introduction of optical fiber is limited to the so-called feeder part, in a fiber-to-
the-cabinet (FTTC) configuration, which requires far less investments, than bringing fiber-to-
the-home (FTTH) of each individual subscriber [11].
Whereas incumbent operators are migrating their existing access network to a
broadband infrastructure, bringing fiber gradually closer to the customer, competitive local
CHAPTER 7210
exchange carriers (CLEC) are deploying a totally new access infrastructure. To minimize
costs and maximize revenue, these CLECs only focus on large business and corporate
customers, for which it is feasible to bring fiber-to-the-building (FTTB).
7.3 Access network planning
In this paragraph we describe the considered problem related to access network
planning. In a second paragraph, a framework is described to tackle this planning problem.
The third paragraph discusses different solutions in terms of network topologies, that are
considered in this planning framework.
7.3.1 Considered planning problem
Both for incumbent operators, migrating towards a FTTC scenario, and CLECs using
a FTTB architecture, the investments are still very high. The advantages of efficient network
planning in an environment with high investments and decreasing margins has already been
discussed in Chapter 2. The cost structure of the planning problem in the access network is
quite different than for the transport network. In the transport network, the fiber topology is
typically very sparse. In addition, rights of way to install new fiber are relatively easy to
obtain, or fiber can be leased or bought from third party bandwidth resellers. In the access
network, network topologies are much denser, less installed fiber base is present, rights of
way more difficult and expensive to obtain, and also the costs of installing the fiber is much
higher. Indeed, fiber can not be installed along railway routes or highways, as in the long haul
network, but it has to be buried under ground (see Figure 2). The civil works required for such
fiber installation are extensive, and the associated cost of labour is also very high. As this cost
factor is dominating the entire planning problem related to installing fiber in the local loop,
we will focus on the topological planning of fiber in the access network. Other authors have
focussed on different planning problems, e.g. placement of the splitter locations in a known
network topology for a passive optical network (PON) [12][13], the positioning of
concentrator points in the access network [14][15], optimal location of base stations in mobile
networks [16][17], capacity planning in the access network [18][19] or techno-economic
analysis of access network architectures [20].
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 211
Figure 2: Fiber installation in Ghent
7.3.2 Access network planning framework
The considered topological access network planning problem consists of finding a
minimum cost fiber topology interconnecting the local exchange (LEX), i.e. the gateway node
towards the transport network, with so-called street cabinets (SC) located in the access area.
Starting from the street cabinet, the existing infrastructure can be reused to interconnect to
residential or small business customers. The locations of the street cabinets and LEX are
given. In case of network upgrades, these are existing street cabinets and their location
depends on the topology of the legacy access network. In case a completely new access
network has to be build, the locations of the street cabinets can be determined based on
demographic studies.
An important requirement to be taken into account, is that the optical fiber in the
feeder part needs to be installed along the streets of a city. This street map poses important
topological constraints, which have to be taken into account during the planning problem.
Therefore, accurate geographical information is of paramount importance. Indeed, not all
streets or regions are equal with respect to installing equipment: in some streets civil works
are difficult and expensive (e.g. bridges, tunnels or streets with high traffic), while on other
street segments certain ducts or dark fiber might already be available. Traditionally, operators
maintained such geographical information on paper maps, which needed to be redrawn each
CHAPTER 7212
time the situation changed. In addition, the topological planning was largely done by hand.
Therefore experts used these street maps and decided, based on experience and intuition,
where to install the equipment. During the last decade, information technology has allowed to
replace a vast amount of paper by bits and bytes. Instead of maintaining paper maps, operators
started to store the geographical data in geographical information systems (GIS). Such a
spatial database system, not only allows an easy visualization and flexible manipulation of the
relevant geographical data, but it also opens the door to using this data for computer aided
planning of the access network.
We developed an access network planning framework [4][5] in which the GIS is
used as a spatial database for storing the relevant geographical information, but also as a
graphical user interface for executing self-developed planning algorithms. These algorithms
are developed in C++, an object oriented programming language which is very well suited for
implementing and executing complex planning algorithms. A software interface between the
planning algorithms and the GIS is provided, such that the planning algorithms can access the
geographical data, and such that the user can call upon these planning algorithms from within
the interactive graphical user interface of the GIS. A schematic overview of the tool
implementation is given in Figure 3.
C++
programs
GIS
street map
node positions
data (cost, …)
planning
algorithms
Output
modelling
displaying
I/O
programming
Figure 3: Access network planning framework
7.3.3 Access network topologies
For the cable topology of the fiber plant between the LEX and the SCs, many
possibilities exist. We have to make a clear distinction between the cable topology and the
fiber topology. Each cable contains a large amount of fibers. These fibers can be routed in the
cable topology, such that the fiber topology is different from the cable topology. E.g., the
cable topology can be a tree structure between the LEX and SCs and the fiber topology can
have a star structure, with a dedicated fiber in the cable topology between the LEX and each
SC. In the remainder, we focus on the planning of the cable topology.
The simplest option is to use a star topology (Figure 4a), in which every SC has a
direct cable connection to the LEX. This however can be a very costly solution. Alternatively,
a tree topology (Figure 4b) is a more cost efficient solution, since it allows to share cable
infrastructure (and possibly optical fiber, e.g. in a passive optical network) among different
SCs. The disadvantage of a tree topology is its vulnerability to failures: a cable cut
disconnects all SCs which were connected with the LEX via a fiber through that cable. A ring
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 213
topology (Figure 4c) between LEX and SCs offers an improved reliability, since it provides
two disjoint paths between each SC and the LEX. However the extra amount of cable
infrastructure needed for the ring topology makes this a rather expensive solution compared to
a tree topology.
LEX
SC SC SC
a. Star topology
LEX
SC SC SC
branching
point
b. Tree topology
LEX
SC SC SC
c. Ring topology
ON
LEX
SC SC SC
ONON
d. Hybrid ring-tree topology
Figure 4: Access network topologies
Cost and reliability are typical antagonists in the telecommunications world, which
makes it difficult for operators to select the appropriate network architecture. We propose a
hybrid ring-tree architecture (Figure 4d) as an intermediate solution between the two
extremes of the tree and ring topology, allowing to combine the best of both worlds. This new
architecture implies the introduction of a new network element between the LEX and the SC,
namely the optical node (ON). Rings are introduced to connect the ONs with the LEX.
Because this first part of the feeder network carries a high amount of traffic, reliability is of
paramount importance. Each street cabinet is in turn connected to one of the optical nodes.
Since in this second part of the feeder network the traffic is typically much more spread out, it
would be too costly to use rings. Therefore, a tree topology is preferred. The placement of the
ONs is crucial, since it influences both the topology of the rings in the first feeder section and
that of the trees in the second feeder section. When a lot of ONs are used close to the SC,
rings will predominate. This means the network cost will be high, but it also implies a high
reliability of the network. In the extreme case, all SCs act as ONs and are on the ring. On the
other hand, a small number of ONs close to the LEX leads to a cheap network with a low
level of reliability. In the other extreme case, only the LEX acts as ON, and all SCs are
interconnected to the LEX with a tree structure. So, by selecting the amount and placement of
the optical nodes, the operator can select the ideal trade-off between cost and reliability for his
specific situation.
From a planning point of view, the star topology is most straightforward to plan.
Each SC is individually connected to the LEX, and can be individually optimized by
CHAPTER 7214
searching the least cost cable path between SC and LEX. For the tree, ring and hybrid ring
tree topologies, the planning is much more difficult and requires more complex algorithms. In
the remainder of this chapter we present a planning methodology for each of these topologies.
7.4 Tree network planning
7.4.1 Problem formulation
For the design problem under study, we assume the following: we know the position
of the LEX and the positions of the SCs on the street map of the city. This street map can be
represented as an undirected weighted graph G(V,E) in which every node v∈V corresponds
with a street corner, or a given position of a SC or the LEX. The edges e∈E represent the
street segments interconnecting these nodes. The set of SCs is represented by N⊂V. The so-
called leaf nodes of the tree we want to determine, are constituted by the SCs together with
the LEX.
Furthermore we know the installation cost of a fiber cable in each street segment.
This cost can be represented as the weight or cost of an edge in the graph G. The cost of
installing a fiber cable in a street segment can be considered as proportional to the length of
the street segment, however if in certain street segments some ducts or dark fiber might
already be available, this cost could be decreased. Costs can also be increased to reflect on the
high cost of installing fiber in a certain part of the network (e.g. when traversing a bridge).
The aim of the topological tree network design problem is to find a minimum cost
sub-graph (i.e., a tree), which interconnects all leaf nodes. Since the LEX is one of the leaf
nodes, this tree will interconnect all SCs to the LEX. The optimal tree topology does not
depend on which of the leaf nodes is the LEX. This problem is also known in graph theory as
the Steiner Tree problem [21]. Whereas several optimal solution methods exist for this
problem, such solution methods typically require extensive calculation times and are not
suitable for large scale problem instances. In access network planning, the graph representing
a street map can easily contain several thousands of nodes and edges and tens or hundreds of
leaf nodes. For such problem instances, heuristic solution methods are a better fit, as they
allow to obtain near-optimal results in short calculation times. In the next paragraph we
present three heuristic techniques to solve the Steiner Tree problem for large-scale access
networks.
7.4.2 Solution methods
Three heuristics were developed to optimize the above-described problem. The first
heuristic (zoom-in technique) starts with a global overview of the problem, i.e. initially only
the positions of the leaf nodes are considered, while the street map is ignored. The second
heuristic (iterative path finding method) also considers some street map independent
information to decide on the sequence of interconnecting leaf nodes. The third heuristic
(artificial minimum spanning tree) is fully based on the real cable installation costs from the
beginning. We describe these heuristics according to the problem abstraction made in the first
phase of the algorithm.
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 215
7.4.2.1 Zoom-in approach
This first heuristic can be described as a zoom-in technique: starting from a global
interconnection scheme of the leaf nodes (without considering the street map), we gradually
consider more and more details of the design problem. The method consists of three phases.
In a first phase, the street map pattern is ignored, i.e. we consider distances from a
bird’s-eye view (Euclidean distances) instead of distances along the street map (real
distances). This way, a tree-like interconnection pattern is built, which provides a connection
between all leaf nodes using straight-line segments, while some branching nodes are
introduced [22].
The second phase maps this straight-line tree on a fiber topology within the street
map. The differences between the Euclidean distances and the real distances (along the street
map, using the shortest path algorithm of Dijkstra [23]) are taken into account by calculating
the ratio between the real distance and the Euclidean distance for each straight-line segment.
We sort the line segments based on this ratio and in this order, we determine cable routes in
the street map for these line segments. In each iteration, the shortest path between the two
corresponding nodes is calculated in the street map. Afterwards, we set the cost of the street
segments used by this path to zero, to allow reuse of street segments with already installed
fiber cables in subsequent iterations. Two implementation orders - increasing or decreasing
ratios - are tried out and the cheapest solution is saved for further calculation.
In the first two phases, we concentrated on creating a good global interconnection
scheme for the leaf nodes, without taking the street map into account at a high level of detail.
The third and final phase studies the obtained solution more locally and tries to improve some
details of the solution. Two types of local optimization methods were implemented. First, we
can try to remove a part of the tree (e.g. a connection between a leaf node and a branching
node) and try to reconnect the separated parts of the tree in a cheaper way. Second we can
attempt to replace a branching node from one street corner to an adjacent street corner if this
leads to a cheaper network.
7.4.2.2 Iterative path finding approach
This second heuristic can be described as an iterative path finding method. In order
to make optimal shared use of cables we will try to introduce the longest cable routes in the
first iterations of the heuristic and try to reuse these cable routes as much as possible towards
the last iterations. We achieve this by initially searching for the shortest path between the two
leaf nodes which are as far as possible apart (from a bird’s-eye view). On this path we will
introduce the first cables. In addition, to force the path finding algorithm to pass along other
leaf nodes we assign a negative "bonus" cost to the other leaf nodes. We indeed prefer a path
which is a little bit longer, but passes several other leaf nodes (which then do not have to be
interconnected later on). Afterwards we try to find the shortest path between the next two
nodes which are as far as possible apart and not interconnected yet, and introduce cables on
this path. In order to reuse the cable that has been introduced in previous iterations as much as
possible, we assign a zero cost to edges corresponding with street segments where cable has
already been introduced in previous iterations. We repeat the above described procedure until
all leaf nodes have been interconnected.
When no local optimization is used, the solution obtained with this method is
dependent on the value of the bonus cost assigned to the leaf nodes. However when we use
CHAPTER 721
6
local optimization procedures afterwards, this dependency diminishes. These local
optimization routines are similar but somewhat less optimal (to improve calculation times) to
the ones described in the previous section and focus on fixing the bad choices that could have
been made during the initial construction phase.
7.4.2.3 Artificial minimum spanning tree approach
A third method is based on the following observation: if it would be possible to go
directly from one leaf node to another (i.e. using a straight line, without having to obey the
street map), a minimum spanning tree in the network induced by the leaf nodes and these
straight lines could be suggested as a reasonable access network topology.
However, we cannot go straight from one leaf node to another, because we can only
lay cables along streets. Nevertheless, the route along the streets, which corresponds best (in
terms of cost) with this direct interconnection between leaf nodes, is the shortest path between
these leaf nodes. Therefore, we propose the following heuristic technique for finding a tree
network layout.
First, we compose an ‘artificial graph’ consisting of only the leaf nodes. With this set
of nodes, we construct a complete graph (i.e. a full mesh between all the leaf nodes). The cost
of each artificial edge is set equal to the length of the shortest path in the original street map
between the corresponding leaf nodes. This can be done by using the shortest path algorithm
of Dijkstra [23] to compute the shortest path lengths between one leaf node and all other leaf
nodes in the street map. Considering these edge costs, we then compute a minimum spanning
tree in the artificial graph. The computation of this tree can be done using the algorithm of
Kruskal or the algorithm of Prim [21].
The set of edges returned by the minimum spanning tree algorithm are then mapped
in the original street map as follows: starting from the shortest artificial edge, the shortest path
along the street map is searched between the corresponding leaf nodes. Every artificial edge
corresponds to an interconnection of a pair of leaf nodes in the original street map. We will
install fiber cables in every street passed by this shortest path. We repeat this procedure until
all leaf nodes are interconnected by cables. Moreover, in order to encourage reuse of ducts,
we set the installation cost of used street segments to zero, each time after we have searched
the best way to connect a pair of leaf nodes.
To decrease the cost of the access network even further, we can subject the resulting
tree network to some local optimisation routines. These routines concentrate on improvement
of details of the network, rather than on producing a global interconnection scheme. Possible
optimisation procedures, based on local alteration of the cable topology, were already
described in section 7.4.2.1.
7.4.3 Results
7.4.3.1 Sample networks
In order to test the performance of the algorithms, different network sizes (number of
nodes, links and leaf nodes) and street map patterns are examined. We consider 5 sample
networks, of which the properties are described in Table 1.
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 21
7
Sample network Number of nodes Number of edges Number of leaf nodes
Small network 41 70 10
Gent 1 843 1281 23
Gent 2 843 1281 45
Grid 25x35 875 1690 50
Grid 100x100 10000 19800 200
Table 1: Sample networks
The first sample network is a small network, which allows us to trace the detailed
evolution and to verify the correctness of the algorithms. The second and third sample
network are based on a part (about 10 km2) of the street map of Gent, considering 23 and 45
leaf nodes respectively. The fourth and fifth examples are both grids. Grid 25x35 has about
the same size as Gent 2, although the street map pattern is totally different. This however does
not imply that the example is unrealistic, since these checkerboard street maps do occur in big
cities as New York. Grid 100x100 is used to investigate the calculation times on huge
problem instances.
In section 7.4.3.2, we present results for these sample networks, and in section
7.4.3.3 these results are discussed in more detail and relevant conclusions are drawn.
7.4.3.2 Numerical results
For each sample network and each algorithm (ZI = zoom-in algorithm, IPF =
iterative path finding algorithm, AMST = artificial minimum spanning tree algorithm), we
measured the cost of the network (symbol c) and the CPU time (symbol t, expressed in
seconds, measured on a standard PC Pentium 166 MHz) needed to calculate that solution,
both before (index 1) and after (index 2) the local optimization procedures. The numerical
results for the small network are shown in Table 2. For this simple problem, the three methods
yield exactly the same solution (after local optimization).
c1t1c2t2
ZI 899 0.051 899 0.101
IPF 957 0.010 899 0.030
AMST 917 0.010 899 0.030
Table 2: Numerical results for small network
Table 3 shows the numerical results for Gent 1. The ZI and the AMST algorithm
yield exactly the same solution, whereas the second method leads to a slightly more expensive
solution (although the difference is only 0.4%).
c1t1c2t2
ZI 11349.42 0.751 10957.90 5.769
IPF 11549.40 0.150 11001.80 2.583
AMST 11167.15 0.261 10957.90 2.214
Table 3: Numerical results for Gent 1
CHAPTER 7218
The results for Gent 2 are shown in Table 4. In this case, three different results are
obtained. The cost difference between the three heuristics is however small (all within the
range of 0.7%) and large parts of the tree networks obtained are exactly the same. The best
result (cost 16436.78) is depicted in Figure 5. The dark gray lines correspond to the street
segments, while the black lines depict the presence of fiber cable (representing our solution).
The small gray dots indicate the positions of the leaf nodes.
c1t1c2t2
ZI 17494.38 3.184 16436.78 20.419
IPF 17177.50 0.290 16452.40 2.704
AMST 16960.35 0.514 16549.64 2.707
Table 4 : Numerical results for Gent 2
Figure 5: Best result for Gent 2
The numerical results for the 25x35 grid are shown in Table 5. The solutions found for
this situation differ more considerably (in the range of 2.5%). The best solution is now found
by the second method.
c1t1c2t2
ZI 870 4.166 811 12.208
IPF 885 0.210 808 4.216
AMST 845 0.586 828 1.738
Table 5 : Numerical results for Grid 25x35
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 219
Table 6 shows the results for the big 100x100 grid. These results are again in the same
neighborhood (in a range of 0.3%).
c1t1c2t2
ZI 4157 291.020 3986 1189.299
IPF 4439 42.481 3999 306.541
AMST 4093 37.214 3996 486.971
Table 6 : Numerical results for Grid 100x100
7.4.3.3 Discussion of results
From the numerical results mentioned above, we can conclude that in general the
quality of the results obtained after local optimization is very similar for the three heuristics.
Because these three independent approaches yield similar results, we can state that each of the
three algorithms produces near-optimal results. On the average, the zoom-in technique
performs best. This is for a large extent due to the local optimization phase, which was
extensively elaborated in this case, which also explains the larger running times. On average,
local optimization allows a cost reduction of about 5-10%. When we compare the results
before local optimization, the artificial minimum spanning tree heuristic is clearly the best. As
mentioned in section 3, the zoom-in technique and (to a lesser extent) the iterative path
finding algorithm make more abstraction of the street map in the beginning of the algorithm,
which leads to solutions that contain more local imperfections before local optimization.
The simulation results are all obtained quickly: even for the grid with 10000 street
corners, all calculations were executed within 20 minutes. When comparing the calculation
times of the different algorithms, significant differences are noticed. The iterative path
method and the artificial minimum spanning tree technique are the fastest methods, while the
zoom-in algorithm is considerable slower.
For the sample problems studied above, we usually assumed that the cost of an edge
is proportional to the length of the corresponding street segment. This allows an intuitive
visualization of the results. If some optical cable infrastructure is already available, this can be
reflected by a reduction of the edge cost in our model. As such, the correlation between the
real costs and the Euclidean distances will decrease. The artificial minimum spanning tree
technique is most robust with respect to this modified problem situation, since this approach
does not rely on Euclidean distances. However, due to the local optimization phase, the other
two techniques turned out to be robust as well.
The grid network is the most difficult structure, because a lot of shortest paths with
equal length exist between two nodes on the grid. This explains why the difference between
the results for the different heuristics is largest for the grid structure.
We can conclude that each of the three approaches has its strengths and weaknesses.
When calculation time is no limiting factor, the zoom-in technique yields - on the average -
the best results. The iterative path finding algorithm combines good results and fast
calculation times. The artificial minimum spanning tree technique has the advantage to
produce reasonable results very quickly (before local optimization) and is most robust with
respect to already available optical cable infrastructure.
CHAPTER 7220
7.5 Ring network planning
7.5.1 Problem formulation
As in the tree network planning problem, the input for the ring network planning
problem consists of the locations of SCs and LEX in a given street map. The aim of the ring
design problem is to find a minimum cost set of rings, such that all SCs are connected to the
LEX within the given street map. Each ring comprises the LEX and each SC is connected to
exactly one ring. In this way, two alternative disjoint paths exist between every SC and the
LEX. The easiest way to solve this problem would be to place all SCs and the LEX on one
ring. Such a 'travelling salesman solution' [24] is however not possible in most cases, because
we have to take into consideration a number of constraints. Typical constraints to be
considered are the maximum number of SCs Nmax, which can be connected on one ring and
the maximum fiber length Lmax of the ring. These constraints can be introduced to reflect the
technical limitations (e.g. available power budget, limitations inherent to the protection
switching protocol on the rings) or to optimize the survivability of the network. The latter can
be explained as follows. A ring offers protection against any single fiber cut. However, when
a second fiber cut occurs in the same ring, before the first fiber cut has been repaired, part of
the ring becomes disconnected. The probability that such an event occurs, increases with the
total length of the ring. When considering a network composed of a large set of short rings,
the probability that a double fiber cut occurs in the same ring is much lower that the
probability that this event would occur in a network composed of a limited set of long rings.
Similar conclusions have also been drawn in Chapter 6 for interconnected WDM rings in the
transport network. Hence, restricting the length of the rings connecting LEX and SCs (thereby
partitioning the network in more rings) offers improved survivability against double fiber cuts
in the local access network.
7.5.2 Solution method
As already mentioned for tree network design, the potentially large size of problem
instances does not favor an exact optimization approach, such as integer linear programming.
Therefore we looked for a heuristic solution method, capable of obtaining good results using
limited computational resources and calculation times. We propose a solution method inspired
by heuristic techniques, which have been successfully applied to solve the vehicle routing
problem [25][26]. Indeed, the problem under study is very related to the vehicle routing
problem, which aims at optimally routing a fleet of trucks from a central depot towards a
number of customers that have to be served. The analogy is obvious: the LEX takes the role
of central depot, the SCs are the customers and the routes of the trucks are the positions of the
rings (see Figure 6).
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 221
Customer
(SC)
Depot
(LEX)
Route
(ring)
Figure 6: Analogy between vehicle routing and ring planning
Solving the problem under study consists of two main parts: determining an optimal
clustering of the SCs (a cluster of SCs represents a set of SCs which are connected to the LEX
on the same ring) and within each cluster determining the optimal ring (or traveling salesman
tour [24]) between the SCs and the LEX. Clarke & Wright [27] have proposed a simple route
building heuristic to solve both problems in an integrated way. The heuristic starts with a
solution in which every customer (SC) is connected to the depot (LEX) with one single route
(ring). This route (ring) is thus dedicated to one single customer (SC), so it consists of two
parallel paths, back and forth to the depot (LEX). The cost of this route (ring) is thus twice the
shortest path cost, between depot (LEX) and the customer (SC). Since the ring in the access
network needs to survive a single cable cut, both parallel routes have to use disjoint cables.
This can be achieved, by installing the cable routes along both sides of the streets.
Afterwards the heuristic iteratively tries to combine any two routes that result in the
best cost savings without violating the constraints. This heuristic performs well in general,
unless for cases where the constraints are tight. In the latter case, the heuristic might end with
a lot of routes (rings) which can not be combined due to the tight constraints.
In contrast to Clarke & Wright, we based the clustering of SCs on a number of
predetermined seed nodes around which rings have to be build. The seed nodes are
determined by taking the constraints into account in advance. The determination of seed
nodes is based on a pull function, which takes into account the geographical distance between
the nodes. Since the fibers of the ring have to be placed along the streets, the distance function
between two nodes is the shortest (or cheapest) path (using the algorithm of Dijkstra [23])
between these two nodes in the graph G, representing the street map. The pull p between two
nodes n and m is expressed as:
2
)],(dist[
1
),( mn
mnp = ,
where dist(n,m) represents the shortest path distance between nodes n and m in graph G.
This pull function is equivalent to an undirected gravitational force between two
objects of unit mass and has also been successfully used for clustering in [28]. We use this
pull function in the following way: the first SC selected to act as seed node for a ring is the
SC with minimum pull (hence maximum distance) towards the LEX. Each additional seed
node is the SC which has the minimum average pull towards the LEX and SCs which have
CHAPTER 7222
been selected to acts as seed nodes so far. We repeat this process until we have found R seed
nodes, determined by:
ù
ê
ê
ê
é
=
max
N
N
R ,
Indeed, since maximum Nmax SCs are allowed per ring, at least R rings will be
required in any feasible solution.
After the SCs which act as seed nodes for a ring have been selected, the other SCs
are iteratively assigned to the ring of the seed node which result in the best cost savings. The
cost savings are calculated using the savings functions as proposed by Clarke and Wright
[27]. This savings function is illustrated in the example of Figure 7 (we omitted the street map
for clarity).
The total cost of the ring connecting SC i and the ring connecting SC j to the LEX is:
c(i,j) = 2.dist(0,i) + 2.dist(0,j) ,
where node 0 represents the LEX, and distances (=costs) are again calculated along the
shortest path in graph G.
The cost of combining both SCs in a single ring is:
c'(i,j) = dist(0,i) + dist(i,j) + dist(0,j).
Thus, the cost savings achieved by combining seed node i and SC j in one ring is:
s(i,j) = c(i,j) - c'(i,j) = dist(0,i) + dist(0,j) - dist(i,j).
0i
j
Seed node SC
SC
LEX
dist(0,i)
dist(i,j)
dist(j,0)
Figure 7: Illustration of the savings function
We can use the same savings function in case more than one SC is already present on
a ring. In this case we see the entire ring as one seed node and try to add the SC between the
LEX and the existing ring. To achieve this, we try to insert the SC between the LEX and the
SC on the ring next to the LEX (the end node of the ring). In this case two possibilities can be
evaluated per ring since two SCs are adjacent to the LEX on each ring.
In order to determine which SC to add to which ring in the next iteration, we
calculate these savings for all SCs (not added to a ring yet) with respect to all rings. We will
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 223
then select to add the SC to the particular ring which results in the best savings, and which
does not violate the constraints. In case no SC can be added to a ring without violating the
constraints, a new ring has to be created by searching for a new seed node. In this case, we
select to use the SC not added to a ring and with maximum (shortest path) distance from the
LEX as new seed node.
After all SCs have been assigned to rings, a final local optimization phase is
performed to optimize the ring within each cluster individually. This is done by performing a
local search on each ring, using a 3-optimal interchange procedure [24]. This procedure
attempts to perform a number of consecutive 3-opt moves that improve the current solution. A
3-opt move, as illustrated in Figure 8, consists of removing three edges from the current ring,
and tries to close the ring again, in the best possible way. Eight possible combinations of
edges exist to close the ring again as shown in Figure 8. The best of these eight combinations
is retained for the 3-opt move. By iterating over all possible 3-opt moves within the ring, we
can determine how to improve the current ring. This process is repeated until no more 3-opt
moves result in a better ring.
ab
c
de
f
g
h
i
Possible new edges:
ab de gh ae dg bh
ab dg eh ae bg dh
ad be gh ag de bh
ad bg eh ag be dh
Figure 8: 3-opt move
7.5.3 Results
To investigate the quality of the results which can be obtained using the heuristic
solution method described in the previous paragraph, we have compared the obtained results
with solutions obtained with an integer linear programming (ILP) solution method. For an
overview of ILP formulations for the vehicle routing problem we refer to [26]. As already
mentioned previously, the ILP approach is not very well suited for large problem instances.
Therefore we have only been able to make comparisons between the heuristic and the ILP
approach for networks with a small number of SCs. We performed 100 simulations for 10
randomly selected SCs on a given street map of 843 nodes and 1281 edges (representing part
of the city of Gent). The other parameters were Nmax = 6 and Lmax==
∞
. On average, the result
of the heuristic solution approach was 2.4% more expensive than the result of the ILP
approach, with a maximum difference of 16% for one particular instance. When roughly
comparing calculation times, solving one ILP instance required about 20 minutes on average,
while the heuristic solution was found within a second. Other simulations yielded comparable
results, which means that on average the heuristic solution is capable of finding good
solutions in moderate calculation times.
In Figure 9, a typical result for the ring network is depicted for the street map of the
city of Gent (843 nodes and 1281 edges) with 50 SCs. We notice that some nodes seem not to
be 'on' the ring, but connected to the ring with only one fiber. In fact this will not be the case
CHAPTER 7224
in practice, but it only seems to be on a street map level. In practice, on the fiber cable level,
the ring is closed by twice placing disjoint fiber cables in this street segment (e.g., on both
sides of the street to ensure reliability). Although placing two disjoint fiber cables in the same
street segment doubles the cost, this might still be more economical than trying to close the
ring in another way. In certain cases, it is not even possible to close the ring in another way
than by passing twice through the same street segment (e.g. for a SC in a dead-end street).
Figure 9: Result of ring network for the city of Gent
In section 7.4 we have proposed a set of near-optimal solution methods to solve the
problem of designing a least-cost tree network which connects the LEX with all SCs. As a
result, we are able to compare the cost of ring networks with the cost of tree networks.
In Figure 10, the results of the comparison between ring and tree networks are
presented. These results were obtained for the street map of the city of Gent (843 nodes and
1281 edges) with 50 SCs. The values on the X-axis describe the maximum number of nodes
allowed on a ring. We did not use the maximum length constraint in this comparison because
we did not want to bias the results by this (we did not use such a constraint when planning the
tree network either). In the chart we compare the extra cost of a ring network to the cost of a
tree network. Therefore 100 simulations were performed, and in each simulation the positions
of the 50 SCs were selected randomly. For each simulation, the algorithm for the ring network
and the one for the tree network were used and the relative extra cost of the ring network was
calculated. The points in the chart represent the average extra cost over 100 simulations and
the error bars on these points represent the standard deviation. In Figure 11, we present
similar results for a street map with a regular grid structure (checkerboard pattern).
We notice that for a limited number of rings (or a high number of nodes per ring) the
cost of the ring network is about 25 to 35% higher than that of a tree network. It will depend
upon the operator and his specifications (e.g. frequency of failures) whether he finds the extra
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 225
reliability worth the extra investment. It is however clear that small rings (less than 10 nodes)
are very costly and will not be chosen for.
An interesting observation is that the cost of the ring network is never higher than
twice the cost of the tree network. This can be explained as follows. The ring network will
have the highest cost if we allow only one SC per ring. In this case we can use the tree
topology, but use two diverse fiber cables in each street segment of the tree to realize the ring.
As such an individual ring can be laid towards each SC, using this doubled tree topology, at
twice the cost of the single tree topology.
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
5040302520181614121098765432
N_max
Figure 10: Average extra cost of ring compared to tree for street map of Gent
CHAPTER 722
6
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
50 40 30 20 15 10 5
N_max
Figure 11: Average extra cost of ring compared to tree for grid street map
In a last section of results, we discuss the influence of the maximum length
constraint. On the X-axis of Figure 12, the values of Nmax and Lmax for which we performed
the simulations are represented. Again, for each combination of the two values we performed
100 simulations on the network representing the street map of the city of Gent (843 nodes and
1281 edges). The average length of a street segment in this street map is 100.7 units.
Lowering Lmax below 5,000 results in a lot of infeasible solutions. We notice that the average
extra cost of the ring network compared to the tree network (calculated in terms of percentage
in the chart) is very much dependant on the value of Lmax. For high values of Lmax, the
maximum node constraint is determinative. For Nmax = 50, the average extra cost of a ring
compared to a tree is about 30% for high values of the maximum length. In this case one
single ring can span all SCs. When the Lmax is lowered, the optimal travelling salesman
solution is in most cases not possible anymore, and multiple rings are required, resulting in a
considerable cost increase. As expected, for the case where Nmax is already low, the impact of
the maximum length constraint only becomes apparent for low values of Lmax. E.g. for Nmax =
10, the results are equal for all values of Lmax, except for Lmax = 5,000. When looking at the
results for Lmax = 5,000, the maximum node constraint does no longer matter, since we notice
that the results for Nmax = 50, 20 and 10 are almost equal.
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 22
7
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
50
50,000
20
50,000
10
50,000
50
20,000
20
20,000
10
20,000
50
10,000
20
10,000
10
10,000
50
5,000
20
5,000
10
5,000
N_max
L_max
Figure 12: Influence on the maximum length constraint
In conclusion, a heuristic algorithm based on vehicle routing techniques showed to
be suitable for solving the topological ring design problem. Comparisons with tree networks
showed that ring networks are at least 25% more expensive for the studied networks.
Decisions on whether or not the extra cost of the ring network compared to the tree network is
worth the extra investment depends on the values of Lmax and Nmax and the operator's specific
requirements.
7.6 Hybrid ring-tree network planning
7.6.1 Problem formulation
While planning a tree or ring topology, the main question was to determine the most
economical fiber layout. In contrast, planning of a hybrid ring-tree architecture is much more
difficult, since it requires several other decisions to be taken as well. The problem consists of
determining the optimal number of optical nodes (ON) and their optimal locations.
Furthermore it has to be decided which SCs are to be connected to which ON (clustering of
SCs). Finally, the optimal locations for the fiber cables of the rings (interconnecting ONs to
the LEX), and of the trees (interconnecting SCs to the ONs) have to be determined. The
subproblems mentioned above are all intertwined, leading to a general optimization problem:
minimizing the total network cost while attaining a certain level of reliability [29].
One of the main motivations to adopt the hybrid ring-tree architecture is to increase
the reliability in the first feeder part of the network, where the bulk of the traffic is
transported. Thus, in order to take this reliability aspect into account, we have to know the
CHAPTER 7228
traffic distribution in the network. In section 7.6.1.1, we will describe how to calculate the
network reliability based on the traffic distribution and network topology. In the access
network, all traffic of the SCs is bound towards the LEX. Indeed, the LEX is not only the
physical gateway to the transport network but also the location where the actual switching
takes place, such that also traffic between two SCs in the same geographical access area, will
always go through the LEX. As such the traffic distribution can simply be represented by
associating a value with each SC, that denotes the amount of traffic (in Mb/s) that is bound to
the LEX.
In order to adopt a certain level of network reliability, several restrictions can be
brought forward. A first restriction limits the maximum amount of total traffic allowed per
cluster. Allowing less traffic per cluster, decreases the amount of traffic that can get lost by a
single fiber cut, and thus enhances the network reliability. Besides the amount of traffic, we
can also use an upper and a lower bound on the number of SCs per cluster. Another restriction
limits the distance between two SCs in the same cluster, in order to avoid long cable routes
(lowering the network reliability). Thus, by adapting one or more of these restrictions a higher
or lower reliability can be obtained.
7.6.1.1 Quantifying network reliability
Cable cuts represent the dominant failure type in the hybrid ring-tree networks we
consider. Hence, to measure the reliability of a proposed network design, distinction must be
made between tree topologies and physical ring topologies. In case of a tree topology, a single
cable cut leads to the disconnection of one or more SCs from the LEX. When considering a
physical ring topology, two cable-disjoint connections to the LEX are provided. This implies
that disconnection of some SCs from the LEX can only happen in case of two simultaneous
cable interruptions. Since such double failures are typically far more rare than a single cable
cut in the local loop, we will only consider the disconnection stemming from single cable cuts
in the network.
The part of the network connecting the SCs with the ON uses a tree structure. In
addition, to reduce the cost, certain parts of the ring interconnecting the ONs with the LEX
can pass twice through the same cable using different fibers (see also section 7.6.2), as shown
in Figure 13. This implies that the part of the ring towards one ON can be disconnected from
the LEX by a single cable cut as well.
The probability of disconnection of a SC from the LEX is proportional to the
distance (along the cable) from that SC to the place where two disjoint paths exist towards the
LEX. Hence, the probability of disconnectivity Probdisc, averaged over all network traffic, can
be defined as:
[]
cable
ki
ik
ki
ikk
i
ikikk
disc X
d
dONringLdSCONL
Prob ⋅
ý
ü
î
í
ì⋅+⋅
=
,
,,, ),(),(
Where:
• k indexes the clusters,
• i indexes the street cabinets within each cluster,
• L(ONk , SCk,i) = cable length between street cabinet i and the ON of cluster k,
• dk,i = demand originating from street cabinet i in cluster k,
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 229
• L(ring , ONk) = cable length between the ON of cluster k and the ring,
• Xcable = probability of a cable cut along one unit length of cable.
An example for one cluster is shown in Figure 13.
SC
SC
SC
ON
ring
ring
d=2
d=5
d=4
d=2+4+5=11 L=5
L=2
L=2
L=3
L=1
Figure 13: Example of reliability calculations
The average distance SC – physical ring is:
2.8
254
)254.(52.45.34.3 =
++
+++++
If Xcable = 10-6 this leads to a probability of disconnectivity Probdisc = 8.2 . 10-6.
7.6.1.2 Network cost
When optimizing the total network cost of the hybrid topology, it is no longer
sufficient to look just at fiber cable installation cost. Also the optical nodes which have to be
introduced have a certain cost, partly dependent on the amount of SCs connected to them.
Therefore, different cost elements have to be considered: fiber cable cost, optical node cost
and optical interface cost.
The cable cost (Ccab) comprises the cost for digging the ducts, the cost for installing
the cable in the duct, and the cost of the cable and the fibers in it. We assume this cost to be
proportional with the length of the street segment in which the cable has to be installed. One
cable typically contains a large number of optical fibers. We assume the amount of fibers in a
cable to be sufficiently large, and the cost of these fibers is included in cable cost. The cost of
using one of these fibers then equals the cost of installing terminal equipment at the end-
points of the fiber. This is referred to as optical interface cost.
The optical node cost consists of a fixed part and a capacity-dependent part. The
fixed cost CON,fix includes the cost of housing, powering and maintaining the optical node. The
variable cost depends on the amount of SDH interfaces required. For each SC connected to
the ON, one interface is required. For I interfaces a cost CON,var must be counted, where I
represents the number of ports on a tributary card. If all ports are occupied, an additional card
has to be installed in the rack.
CHAPTER 7230
The cost of an optical interface depends on the bit rate of the interface. We only
consider STM-1 (155 Mb/s) and STM-4 (622 Mb/s) interfaces, with costs CSTM-1 and CSTM-4.
The relative values of these cost factors are summarized in Table 7. For the amount
of ports on a tributary card, we assume I = 4.
Ccab CON,fix CON,var CSTM-1 CSTM-4
0.025/m 6 1.5 1 1.5
Table 7: Typical cost values (normalized)
7.6.2 Solution method
7.6.2.1 General optimization scheme
Our solution method for the hybrid ring-tree planning problem is based on an
iterative approach that consists of 4 phases as depicted in Figure 14. In a first phase, the SCs
are divided in groups (clustering) and in each cluster the SCs are connected by a tree
topology. In a second phase the costs of the used street segments for the tree structures are
adapted and passed through to the next phase. As such we can stimulate to reuse the cable
sections of the tree by the ring. In the third phase a least cost ring network is constructed,
connecting the ONs to the LEX, taking into account the adapted costs. The fourth phase feeds
back the position of cables of the ring to the cluster optimization phase (= 1st phase) and
adapts the costs of the cable sections used by the ring. As such, the position of the ring can be
anticipated in the cluster optimization phase in the next iteration. This process, alternating
between cluster and ring optimization, is repeated a few times and the cheapest solution is
selected as final solution.
Cluster
optimization
Cost
adaptation
Cost
feedback
Ring
optimization
Figure 14: Solution method
7.6.2.2 Cluster optimization
Clustering of the SCs can be based on their geographic proximity and traffic
requirements, taking into account the constraints to tune the network reliability. We use an
iterative method, that tries to merge clusters of SCs, based on these constraints. First, we
compose an ‘artificial graph’ consisting of only the SCs. With this set of nodes, we construct
a complete graph (i.e. fully meshed between all the SCs). The cost of each artificial edge is set
equal to the length of the shortest path (calculated using the algorithm of Dijkstra [23]) in the
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 231
original street map between the corresponding SCs. Considering these edge costs, we start
computing a minimum spanning tree for the artificial graph, using the algorithm of Kruskal
[21]. This algorithm starts from a situation where each cluster contains only one SC and
iteratively two clusters are merged. In each iteration, the two closest clusters are merged.
However, in contrast with the original Kruskal algorithm, these two clusters are only merged
if the constraints regarding reliability are respected. When no more clusters can be merged,
the algorithm is stopped. As a result, we obtain a set of clusters and a tree structure within
each cluster.
The set of edges returned by the clustering algorithm are then mapped on the original
street map using the artificial minimum spanning tree algorithm as described in section
7.4.2.3. To further improve the performance of the above described clustering algorithm,
several optimization routines were implemented, as described below.
In some cases, the algorithm described above can lead to a situation where a remote
cluster with only one SC can not be merged with other clusters anymore without violating the
constraints. Indeed, this happens if all clusters in the neighborhood of this remote SC have
reached the maximum number of SCs per cluster and all the other clusters are too far away
(due to the limitation on distance between two SCs in the same cluster). To avoid the
occurrence of remote clusters with only one SC - probably leading to a high ring cost - the
distance of these remote SCs towards other SCs will be artificially lowered and the clustering
algorithm will be repeated. This will encourage the remote SCs to be merged with other
clusters sooner in the clustering process, such that they do not end up as a separate cluster in
the end.
To decrease the cost of the access network even further, we can subject each tree
network to some local optimization routines similar to the ones described in section 7.4.2.
These routines concentrate on improvement of details of a tree network, rather than on
producing a global interconnection scheme. Several local optimization methods were
implemented, based on iteratively trying to replace a part of a tree (e.g., a connection between
a SC and a branching point of the tree) by a cheaper cable connection.
A typical result of the clustering and the design of the tree networks is depicted in
Figure 15. The small dots on the picture represent the SCs, the large dot is the LEX. The
street map is shown with thin lines, the tree networks with full lines.
Figure 15: Result after cluster optimization
CHAPTER 7232
7.6.2.3 Ring optimization
After the clusters and tree structures have been created, the interconnecting ring has
to be designed. First, an ON must be assigned on the tree network in each cluster. Initially, the
street joint (i.e., node) closest to the LEX will be chosen as the ON position, in order to
minimize the ring length. Later on, different locations for the ON can be evaluated.
For finding the cheapest fiber cable ring(s) connecting all the ONs to the LEX, we
can use the algorithm described in section 7.5.2. Alternatively, we can also use a simpler but
very fast heuristic, called the sweep algorithm. In this algorithm the order in which the optical
nodes are connected by the ring is determined by a virtual line, drawn from the LEX. This line
is rotated around the LEX, touching the ONs one by one. This rotating (or sweeping) order
will be used as the order of the ONs along the ring, as illustrated in Figure 16. The LEX is
inserted between two consecutive nodes, ONi and ONi+1, in such a way that the total length of
the ring is as low as possible. This can be easily determined by calculating the following
metric for each two consecutive ONs obtain from the sweep algorithm:
dist(ONi , LEX) + dist(ONi+1 , LEX) - dist(ONi , ONi+1) ,
with the distances calculated along the shortest path in the original street map. This metric can
be looked upon as the extra cost incurred by adding the LEX to the ring of ONs. The LEX
will thus be inserted between the two consecutive ONs that minimize the above value.
Since the sweep algorithm does not result in the optimal ring topology, a second ring
optimization is performed by examining adaptations in the order in which the ONs are visited
by the ring (i.e., the sweeping order is no longer completely maintained). For this re-ordering
process we use the 3-opt exchange procedure as explained in section 7.5.2. If such an
adaptation leads to a cheaper ring design, we adopt this the new order. This process is also
repeated until no cheaper solution can be found among the set of considered order
adaptations.
LEX
ON i
ON
ON
ON
ON i+1
Figure 16: Illustration of sweep algorithm
The fiber ring structure between LEX and ONs is designed by calculating the
shortest paths between the consecutive nodes of the sweeping order. With respect to the cost
of the edges of the path connecting two consecutive nodes ONi-1 and ONi, we can make the
following remark. It might be possible to reduce the cost of the ring, by allowing certain tree
structures in the ring. This implies that the last part of fiber that runs to ONi, and the first part
of fiber that departs from ONi can be in the same cable, thereby reducing the ring cost, but
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 233
decreasing the reliability. This trade-off between low cost and high reliability can be
controlled by multiplying the cost of the paths between previous ONs of the ring by a cost
adaptation factor fadap. A high value of this factor (fadap >> 1) does not allow any cable reuse
and leads to a perfect ring structure. In case of a low value (fadap < 1), cable reuse is
encouraged and some parts of the cable ring will be reduced to tree structures. To avoid
situations where several ONs are connected to the ring via a same tree structure (resulting in a
very unreliable network topology), we only multiply the cost of the edges of path i-1 between
ONi-2 and ONi-1 by fadap, and set the cost of all edges of path i-2, i-3, … to infinity before
calculating path i. Hence, path i can use some edges of path i-1 but no edges of path i-2 for
instance. Figure 17 illustrates the reuse of cable. The last part of the ring towards ONi-1 and
the first part of the ring departing from ONi-1 are reduced to a tree structure. This way, we
obtain tree structures that connect only one ON to the ring structure, such that maximum one
ON can be disconnected in case of a cable break in the ring.
path i-2
cost =∞
ONi-1
ONi
LE
X
path i-1
cost = 0
path i
Figure 17: Example of cable reuse
For each cluster, the initially selected position of the ON is the street joint closest to
the LEX. However, these ON positions do not always result in a minimal cable length of the
ring. Hence, a first optimization considers repositioning of the ONs one by one in each
cluster. Therefore, each street joint of the tree network is checked as ON position and based
on this new ON selection, the ring is redesigned by the sweep algorithm. The position leading
to the lowest ring cost is maintained. This process of successive repositioning of one ON is
repeated until no further improvements can be found.
A typical result of a ring design on a street map with given clusters is shown in
Figure 18. The small black dots represent the locations of the SCs, while the large white dots
represent the positions of the ONs. The ring design is shown with thick full lines (partially
overlapping with the tree networks).
CHAPTER 7234
Figure 18: Result after ring optimization
7.6.2.4 Interaction between cluster and ring optimization
To reduce the cable cost of the hybrid ring-tree network, we want the tree topologies
and the ring to share as many cables as possible. When designing the tree networks in the very
first iteration, we do not know the ring topology yet, hence it is impossible to take the future
position of the ring into account in the first iteration. The ring optimization however is done
for a given position of the trees (see section 7.6.2.1). To stimulate ring-tree overlap, we lower
the cost of the edges used by the tree networks by multiplying the original edge costs with a
factor fcluster (≤ 1). When designing the ring, part of the cable infrastructure of the clusters can
now be reused, resulting in a lower overall cost of the hybrid network. Unfortunately, the
cable for the clusters is not always placed optimally for the ring to reuse this cable. Therefore,
the original cost of the edges used by the ring will now be multiplied with a factor fring (≤ 1) to
stimulate the cluster optimization phase in the next iteration to reuse some cables that may be
useful for the ring design at a later phase. Although the ring from the previous iteration will
not be the same than the ring in the current iteration, its position is representative to be used in
a feedback procedure to allow some kind of "forward looking". Based on this new design of
the tree networks, obtained with the adapted costs, a ring is designed again, etc. This
alternating process between cluster optimization (based on adapted costs for the edges of the
previous ring solution) and ring optimization (based on adapted costs for the edges of the
previous cluster solutions) is repeated a few times (e.g. until no more improvements are found
during a number of iterations), and the cheapest hybrid ring-tree network is selected as the
final solution.
7.6.3 Results
In this paragraph we present some results, focussing on the impact of the main
parameters on the cost and reliability of the network. First we look at the tree structures and
ring structures individually. Finally, we also look at the interaction between both phases.
7.6.3.1 Reliability versus cost in tree networks
By varying the restrictions on the clusters, different network designs, with varying
cost and unreliability values, can be obtained. In Table 8, an example is shown for a network
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 235
with 32 SCs. As can be expected, designs allowing a larger number of SCs per cluster and a
larger distance between SCs in the cluster, result in a lower overall cost but a higher
unreliability. When less SCs are allowed per cluster, or the distance between SCs in the
clusters is limited, the amount of clusters increases, thereby also increasing the cost and the
reliability.
Lower bound
number of SCs
per cluster
Max. SC-SC
distance in
cluster (m)
Number of
clusters
Probability of
disconnectivity
(10-6)
Cost of
clusters
Cost of
hybrid
network
4 800 8 958.8 314.347 378.927
3 800 8 968.1 314.812 379.032
3 600 9 834.8 311.321 397.820
2 800 10 776.2 304.381 394.101
Table 8: Reliability and cost for different restriction settings
7.6.3.2 Reliability versus cost in ring network
As explained in section 7.6.2.3, the reliability/cost dilemma of the ring design can be tuned by
varying the factor fadap. A high factor results in a high ring cost with a high reliability (when
fadap =
∞
, the ring is 'perfect' and its unreliability is reduced to 0), while a low factor leads to
cheap rings with low reliability (some parts of the ring are reduced to tree networks). In
Figure 19, the relation between the ring cost and the adaptation factor is shown for a ring with
10 optical nodes. The cost increases as fadap increases, with the steepest increase for values of
fadap between 0.1 and 1. In Figure 20 a typical relation is given between the unreliability and
the total cost of a hybrid ring-tree network when varying the factor fadap between 0.1 and 10.
210
230
250
270
290
310
0.01 0.1 1 10
Ring cost
Figure 19: Relation between ring cost and fadap
CHAPTER 723
6
345
365
385
405
425
445
172 177 182 187 192 197 202 207 212
Total cost
Probability of
disconnectivity (10^-6
)
Higher value of fadap
Figure 20: Relation between unreliability and cost of hybrid network
The impact on the ring topology for varying factors of the adaptation factor is shown
in Figure 21, Figure 22 and Figure 23.
Figure 21: Ring topology for fadap = 10
Figure 22: Ring topology for fadap= 0.5
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 23
7
Figure 23: Ring topology for fadap = 0.1
7.6.3.3 Interaction between cluster and ring optimization
To obtain hybrid network designs that are satisfactory with respect to reliability and
cost, a good choice of the values fcluster and fring (see section 7.6.2.4) is of primary importance.
In Table 9, we present some results for a network with 43 SCs, using different values for
fcluster and fring. As can be seen, when no feedback is introduced between the different phases
(i.e. fcluster = 1 and fring = 1), the total cost is highest. The more feedback is introduced (i.e. the
lower both values), the better the total result. In general the interaction between the clustering
phase and the ring phase (tuned by fcluster) is the most important. In Figure 24 we represent the
cost evolution for the different iterations of the optimization method with feedback (using
fcluster = 10-5 and fring = 0.5). In the first iteration, only feedback between the clustering phase
and ring phase (tuned by fcluster) is taken into account. At later iterations also feedback
between the ring phase and clustering phase (tuned by fring) comes into play. The best result in
Figure 24 is obtained in the 4th iteration, and this result is kept as final result.
fcluster fring Total cost
1 1 688.655
1 0.5 688.655
0.5 1 662.551
0.5 0.5 665.452
0.5 10-5 665.452
10-5 0.5 617.382
10-5 10-5 616.406
Table 9: Impact of fcluster and fring
CHAPTER 7238
615
620
625
630
635
1234567
Iterations
Total cost
Figure 24: Evolution of cost in different feedback iterations
7.6.3.4 Conclusion
In this paragraph, we proposed an algorithm for hybrid ring-tree network design,
which is based on an alternating process of cluster optimization and ring optimization. To
improve the performance of the algorithm, several post-optimization routines were developed.
Also feedback between the different steps of the algorithm has shown to improve the results.
The influence of the most important algorithm parameters was tested and the antagonistic
nature of high reliability and low cost was examined.
7.7 Conclusion
In this chapter, we have introduced the problem of topological access network
design, considering planning of the cable rollout for future optical access networks. Different
topologies for fiber in the loop have been proposed, such as tree, ring and hybrid
architectures. We developed a framework for planning these different topologies, based on a
geographical information system and self-developed algorithms.
For the planning of tree networks, we proposed three heuristics: the zoom-in
technique, iterative path finding method and artificial minimum spanning tree algorithm. A
comparison between these different heuristics has shown that each algorithm produces near-
optimal results and is well suited for optimizing large-scale problems. The strengths and
weaknesses of each individual algorithm have been investigated. The zoom-in technique on
average yields the best results. The iterative path finding method combines good results and
fast calculation times. The artificial minimum spanning tree algorithm is most robust with
respect to already available optical cable infrastructure.
For the ring network, we developed a heuristic that is inspired on solution techniques
for the vehicle routing problem. The heuristic combines two phases, clustering and ring
optimization, to determine the best-suited rings that interconnect the SCs with the LEX.
Comparisons with optimal solutions obtained through integer linear programming revealed
the good quality of the results obtained by the heuristic. Comparisons with tree networks
showed that ring networks are at least 25% more expensive for the studied networks, and
depend heavily on the constraints to restrict the ring size.
The hybrid ring-tree architecture was described as an intermediate solution between
the ring and the tree structure. Several parameters allow to tune the trade-off between cost and
reliability. We developed an algorithm, based on an alternating process of cluster optimization
and ring optimization, to design such hybrid networks. Feedback between these two phases
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 239
allows to improve the design even further. The influence of the most important algorithm
parameters was tested and the antagonistic nature of high reliability and low cost was
examined.
CHAPTER 7240
7.8 References
[1] F. Yin, P. Arijs, M. Pickavet, M. Gryseels, P. Demeester, “Three heuristic techniques for topological
access network design”, Proceedings of 2nd European Conference on Networks and Optical
Communications, pp. 157-164, Antwerpen (Belgium), June 1997.
[2] P. Arijs, M. Gryseels, M. Pickavet, F. Yin, P. Demeester, “Design of large-scale tree networks: The
case of urban access networks”, Proceedings of 4th Informs Telecommunications Conference, pp. 78-79,
Boca Raton (FLA), March 1998.
[3] P. Arijs, P. Demeester, “Topological design of rings in the local access network”, Proceedings of 6th
International Conference on Telecommunication Systems: Modeling and Analysis, Nashville (TN),
March 1998.
[4] P. Arijs, F. Yin, P. Demeester, “Geoplan: A tool for geographical planning of local access networks”,
Proceedings of 3th IEEE Symposium on the Planning and Design of Broadband Networks, Mont
Tremblant (Canada), October 1998.
[5] P. Arijs, F. Yin, P. Demeester, "A GIS-based planning environment for optical access networks",
Proceedings of 3th Annual Symposium of the IEEE/LEOS Benelux Chapter, pp. 69-72, Gent
(Belgium), November 1998.
[6] P. Arijs, M. Pickavet, N. Nelissen, S. Vanderdonckt, P. Demeester, "Design of a hybrid ring-tree
architecture for optical access networks", Proceedings of 16th International Teletraffic Congress, pp.
195-204, Edinburgh (UK), June 1999.
[7] P. Lagasse et al., "Photonics networks in Europe", Horizon-Infowin (ACTS thematic issue), Telenor AS
R&D, 1998.
[8] B. Khasnabish, "Broadband to the home: Architectures, access methods, and the appetite for it", IEEE
Network, pp. 58-69, January-February 1997.
[9] K. Krueger, "Guess who's bringing fiber to the curb?", Lightwave, pp. 72-73, December 1998.
[10] R.M. Keating, "Fiber re-emerges as a residential access architecture", Lightwave, pp. 96-101,
September 1999.
[11] O. Kawata, I. Sankawa, K. Okada, "Access network evolution scenario and key technological concepts
for the broadband network", Proceedings of IEEE Globecom'97, pp.1471-1475, Phoenix (AZ),
November 1997.
[12] H. Paul, J. Tindle, "Passive optical network planning in local access networks - an optimisation
approach utilising genetic algorithms", BT Technical Journal, Vol.14, No. 2, pp. 110-115, April 1996.
[13] F. Caviglia, A. Gnazzo, G. Baroncelli, "Splitter position in a PON based architecture: a techno-
economic analysis", Proceedings of SYBEN'98, pp. 40-46, Zurich (Switzerland), May 1998.
[14] B. Gavish, "Topological design of telecommunications networks - local access design methods", Annals
of Operations Research, Vol. 33, pp. 17-71, 1991.
[15] H. Pirkul, V. Nagarajan, "Locating concentrators in centralized computer networks", Annals of
Operations Research, Vol. 36, pp. 242-262, 1992.
[16] "Software tools for the optimization of resources in mobile systems", WP5 Deliverable 3, STORM
project (ACTS), October 1997.
[17] B. Jaumard, "Optimization of cellular networks", Proceedings of International Symposium on
Combinatorial Optimization (CO'98), Brussels (Belgium), April 1998.
[18] A. Balakrishnan, T.L. Magnanti, A. Shulman, R.T. Wong, "Models for planning capacity expansion in
local access telecommunications networks", Annals of Operations Research, Vol. 33, pp. 239-284,
1991.
TOPOLOGICAL PLANNING OF THE ACCESS NETWORK 241
[19] A. Schulman, R. Vachani, "A decomposition algorithm for capacity expansion of local access
networks", IEEE Transactions on Communications, Vol. 41, No. 7, pp. 1063-1073, July 1993.
[20] L.A. Ims, J. Mononen, M. Lähteenoja, L. Budry, M. Salerno, U. Ferrero, D. Myhre, B.T. Olsen, K.C.
Aas, "Key factors influencing investment strategies of broadband access network upgrades",
Proceedings ISSLS ‘98, pp. 22-27, Venice (Italy), March 1998.
[21] T.L. Magnanti, L.A. Wolsey, "Optimal trees", in Network Models, pp. 503-616, Elsevier, Netherlands,
1995.
[22] M. Pickavet, "Use of heuristic techniques for global design and planning of telecommunication
networks", Ph.D. Thesis (in Dutch), Ghent University, June 1999.
[23] E. Dijkstra, "A Note on two problems in connection with graphs", Numeriche Mathematics, vol. 1,
pp.269-271, 1959.
[24] M. Jünger, G. Reinelt, G. Rinaldi, "The traveling salesman problem", in Network Models, pp. 225-323,
Elsevier, Netherlands, 1995.
[25] M. Fisher, "Vehicle routing", in Network Models, pp. 1-33, Elsevier, Netherlands, 1995.
[26] G. Laporte, "The vehicle routing problem: An overview of exact and approximate algorithms",
European Journal of Operational Research, Vol. 59, pp. 345-358, 1992.
[27] G. Clarke and J. Wright, "Scheduling vehicles from a central depot to a number of delivery points",
Operations Research, Vol. 12, pp. 568-581, 1964.
[28] C.H.E. Stacey, T. Eyers and G.J. Anido, "A node clustering approach to the design of hybrid
access/backbone networks of servers", Proceedings of the Australian Telecommunication Networks and
Applications Conference (ATNAC'96), Melbourne (Australia), 1996.
[29] N. Nelissen, S. Vanderdonckt, "Development of algorithms for the optimal design of access networks",
M. Sc. Thesis (in Dutch), Ghent University, 1997-1998.
CHAPTER 7242
CHAPTER 8
Conclusions
8.1 Last call…
After 4 years of exciting research in the field of telecommunications network
planning, this work ties together a number of network planning problems with a special
attention towards ring architectures. In this final chapter we summarize the main conclusions
obtained through our research, regarding the applicability of rings in the different parts of the
network. This is done by giving a chapter by chapter overview of this thesis.
In Chapter 1, a short outline of our work was presented.
In Chapter 2, an overview of the telecommunications sector was given, together with
the different planning problems that come forward. The effects of new services, new
technologies and market liberalization on the different players has been discussed. One of the
main effects is the increased competition, which has proven valuable for end users, but is
putting stringent requirements on operators to optimize their profit margins. In this light,
network planning has shown to be an efficient tool to decrease capital and operational
expenditures and to contribute to increasing profit margins. In addition, network planning can
help in making strategic investment decisions, e.g. when evaluating different network
architectures.
Different planning problems have been identified, depending on the considered time
frame, the geographical or functional area of the network and the requirements and objectives
of the network planner. Besides incumbent or new operators, also equipment vendors see
network planning as an important instrument to help them determine optimal applications and
pricing of their products. Finally, a high level overview of some planning techniques has been
given. This includes both investment criteria to make strategic decisions, as optimization
techniques to solve network dimensioning problems. The most relevant optimization
techniques for this work are (integer) linear programming techniques and problem specific
and generic search heuristics.
In Chapter 3, an in-depth overview of transport network technologies and
architectures is given. After a general introduction on the structure of transport networks, the
different functionalities of a transport network are discussed. Starting from the layering and
partitioning principle, we describe the functions of multiplexing, cross-connecting,
consolidation, segregation and grooming. Also some operational aspects are treated such as
management and monitoring of the network, and maintaining network integrity. Afterwards
two protagonists of transport network technologies are described: the Synchronous Digital
Hierarchy (SDH) and Wavelength Division Multiplexing (WDM). We also shortly describe
packet-switched network technologies and the integration of packet switching in transport
CHAPTER 8244
networks, with the focus on IP-over-WDM networks. In a last part of Chapter 3, the different
recovery techniques for SDH and WDM networks are discussed. We describe both
(interconnected) protection rings as protection and restoration in meshed networks, and
compare the different alternatives.
In Chapter 4, a first planning problem is encountered, dealing with SDH-based
WDM networks. The main goal of this chapter was to make a comparison between SNCP
path protection and interconnected MS-SPRings, using different kinds of equipment
configurations in the nodes of the network, also called node scenarios (NS). One of the main
accomplishments was the translation of these node scenarios in mathematical models that can
calculate the amount of equipment needed, taking into account the detailed characteristics and
limitations of the different equipment types. Using these mathematical models in the network
dimensioning process, allowed us to compare network designs for the different node scenarios
and recovery strategies.
For SNCP based network designs, the node scenario using only lower order ADMs
(NS-1) was found to be most expensive due to the excessive amount of equipment required,
because of the limited internal connectivity of these lower order ADMs. Lower order ADMs
in combination with higher order ADMs (NS-2) result in a much better design. The lowest
cost is achieved when using only higher order ADMs and no DXC (NS-5). The node scenario
using intermediate consolidation of lower order traffic through a lower order DXC (NS-4) has
a relatively high node cost, but becomes more interesting for larger networks or when a
heterogeneous traffic mix with a lot of granularities needs to be transported.
For MS-SPRing, the node scenario using intermediate consolidation (NS-4) yields
the cheapest overall network design. This option makes use of lower order DXCs, which on
the upside give higher routing flexibility, but on the downside result in a lower reliability,
because failures of the DXCs can not be recovered from. For small networks with a dense
traffic matrix, or when only one granularity of traffic is present, the node scenario using
higher order ADMs without DXC is also cost competitive (NS-5). The node scenarios using
lower order ADMs (both NS-1 and 2) are less suited for MS-SPRing, because such ADMs are
superfluous due to the non-blocking nature of higher order ADMs for MS-SPRing.
When comparing SNCP with MS-SPRing, it seems that in case of end-to-end
grooming SNCP is the cheapest solution, while for intermediate consolidation MS-SPRing
becomes more cost-effective than SNCP. For each node scenario and network topology,
hybrid networks combining SNCP and MS-SPRing, yield a lower cost than both architectures
using only SNCP or MS-SPRing. This hybrid design is achieved by placing rings in strategic
positions, with suitable traffic patterns.
In Chapter 5, we focused in more detail on a single topological WDM ring. The
different planning problems have been identified for such a ring, and suitable algorithms have
been developed to optimize the planning. The results obtained through the different planning
algorithms also allowed us to make some interesting side-conclusions.
A first planning problem that was considered, was the ring loading problem, i.e. the
routing problem on a shared protection WDM ring, such to minimize the amount of required
wavelengths. First of all, we derived some mathematical bounds to accurately estimate the
amount of wavelengths for different traffic patterns. Second, an integer linear program was
formulated, which allowed us to determine the optimal routing for random traffic patterns on
the ring. As such, we were also able to compare the performance of a shared protection ring
with a dedicated protection ring. In most cases the shared protection ring has a better
CONCLUSIONS 245
wavelength usage (which improves as the number of nodes on the ring increases), unless the
traffic pattern on the ring is bound towards a hub node or includes a lot of long connections
between opposite nodes.
A second planning problem considered the wavelength allocation of routed traffic on
a shared protection ring. We used an integer linear program to allocate wavelengths to these
routed connections, such to minimize the amount of wavelengths needed. This also allowed
us to compare the wavelength requirements of shared protection rings with and without
wavelength conversion. It was shown that wavelength conversion has very limited benefit for
rings with static traffic. In addition, our results have shown that the ring loading problem and
the wavelength allocation problem, which are intertwined, can be solved separately without
any significant loss of optimality.
A third planning problem for a single topological ring considered a hybrid
architecture, in which both dedicated and shared protection rings can be used in a stack of
multiple rings. This hybrid design was conceived to combine the best of both ring types. An
integer linear program was developed to optimize the cost of such a hybrid design. This
hybrid model has shown to be particularly advantageous in case the cost of a dedicated
protection ring is lower than the cost of a shared protection ring. In this case, the optimized
hybrid architecture yields a lower cost than an architecture using only shared or dedicated
protection rings. An architecture using only one type of ring is only preferred for special
demand patterns, such as hubbed or long demand patterns (favoring the dedicated protection
ring) or adjacent demand patterns (favoring the shared protection ring). When the dedicated
and shared protection ring have the same cost, we have even found the hybrid architecture to
be more cost-effective than an architecture using only one type of ring, in cases that the
amount of inter-ring traffic is substantial and needs to be protected using drop & continue.
A fourth and final planning problem again considered stacked rings, but this time
SDH rings which are stacked on the wavelengths of a WDM ring. For such an architecture we
have devised a way to minimize the installation cost by elimination of SDH ADMs in nodes
of the ring where no SDH traffic needs to be dropped. Optimizing this ADM elimination
involved routing the traffic within the SDH rings in such a way that the amount of traffic
passed through in the nodes of the SDH rings was maximized. For this purpose we developed
an enhanced integer linear program. For medium sized rings with random demand patterns,
40% and more of the amount of SDH ADMs could be eliminated, which resulted in large cost
savings. The largest savings were possible for demand patterns with a lot of hubbed and long
connections.
In Chapter 6, we no longer considered a single ring topology, but a network based on
multiple interconnected dedicated protected WDM rings (OCh-DPRings). The different tasks
of the interconnected ring network design problem have been outlined, and solution methods
based on both heuristics and optimal methods have been proposed.
For the ring routing problem, an efficient heuristic, based on an iterative application
of shortest path routing has been proposed. Next to the heuristic, an optimized integer linear
program based on the K-shortest paths was developed. We have shown that the heuristic
performs in close range with the ILP algorithm, and optimal solutions can be found with ILP
using small values of K.
For the ring dimensioning problem, we also developed a heuristic and an integer
linear programming algorithm. The heuristic has shown to yield near-optimal results in short
calculation times. By applying these dimensioning algorithms to sample networks with
CHAPTER 824
6
different ring constellations, we have found that a network composed of a relatively large
amount of small rings typically has a lower cost and higher availability than a network
composed of a relatively small amount of larger rings.
For the ring identification process, a three phased solution method has been
developed, consisting of ring generation, ring pre-selection and ring optimization. The last
phase can use exhaustive search or can apply more intelligent search methods such as tabu
search. The tabu search heuristic has shown to yield good results in a short number of
iterations. During ring optimization, different ring combinations have to be evaluated using a
fast dimensioning algorithm. The heuristic dimensioning algorithm we have developed for
this purpose, proved very suitable for this.
Finally, the algorithms for interconnected ring design have also been used to
compare ring-based designs with mesh-based designs. We have shown that interconnected
OCh-DPRings are competitive with mesh protected designs in terms of both cost and
availability. Mesh restorable networks can yield lower cost designs, if economical large-scale
OXCs can be manufactured.
In Chapter 7 we considered the planning problem related to installing fiber in the
feeder part of the access network. We targeted the fiber-to-the-cabinet architecture, and
considered different network topologies, such as tree, ring and hybrid architectures. A
planning framework was developed, based on a geographical information system and network
planning algorithms, which allowed to plan and compare the different topologies.
For the planning of tree networks, we proposed three heuristics. A comparison
between these different heuristics has shown that each algorithm produces near-optimal
results and is well suited for optimizing large-scale problems. The strengths and weaknesses
of each individual algorithm have been investigated. On average, the zoom-in technique
yields the best results. The iterative path finding method combines good results and fast
calculation times, as the artificial minimum spanning tree algorithm, which is most robust
with respect to already available optical cable infrastructure.
For the ring network, we developed a heuristic that is inspired on solution techniques
for the vehicle routing problem. The heuristic combines two phases, clustering and ring
optimization. Comparisons with optimal solutions obtained through integer linear
programming revealed the good quality of the results obtained by the heuristic. Comparisons
with tree networks showed that ring networks are at least 25% more expensive for the studied
networks, and depend heavily on the constraints that restrict the ring size.
The hybrid ring-tree architecture was proposed as an intermediate solution between
the ring and the tree structure. Several parameters allow to tune the trade-off between cost and
reliability. We developed an algorithm, based on an alternating process of cluster optimization
and ring optimization, to design such hybrid networks. Feedback between these two phases
has shown to improve the design even further. The influence of the most important algorithm
parameters was tested and the antagonistic nature of high reliability and low cost was
examined.
8.2 Where do we go from here?
This work considered several planning problems for networks relying on the use of
rings. However, the problems dealt with are only pieces of the puzzle and many more
challenging ring design problems are still open for further investigation. In particular the rapid
CONCLUSIONS 24
7
growth of data traffic puts different requirements on rings and networks in general. In this
thesis we always assumed static traffic as input to our planning problem. As data traffic is of a
much more dynamic nature, it would be interesting to investigate the performance of rings
under such conditions. In addition some pro-active vendors have already offered data-centric
ring solutions, such as DPT [1] or DTM [2]. Although these new architectures promise to be
much more efficient and self-managing, planning and dimensioning of the rings still needs to
be performed in advance in order to build such networks. While some of the planning
problems dealt with in this thesis will still be relevant, also some new design problems will
arise for such rings. In addition, comparisons between the different new architectures and
existing rings could be an interesting area of research. Finally, the newly proposed framework
for IP-based WDM networks [3], in which the WDM network falls under the control of the IP
layer, puts question marks on the adoption of rings, since the IP layer is intrinsically
optimized for meshed architectures. This scenario has been surrounded with a lot of hype, and
some have already claimed the death of rings. In order to provide some more valuable
arguments on this topic, there is definitely some more thorough research required.
8.3 References
[1] "Dynamic packet transport technology and applications overview", Cisco Systems White Paper,
February 1999.
[2] L.H. Ramfelt, "DTM offers value-added IP services to metropolitan area networks", Proceedings of
NFOEC'99, Chicago (IL), September 1999.
[3] J. Luciani, B. Rajagopalan, D. Awduche, B. Cain, B. Jamoussi, "IP over optical networks - A
framework", IETF Internet Draft, April 2000.
CHAPTER 8248