An agile all-photonic network

Abstract

This paper presents an overview of recent and current work being conducted in the "Agile All-Photonic Networks", AAPN, Research Network. An AAPN is a wavelength division multiplexed network that consists of several overlaid stars formed by edge nodes that aggregate traffic, interconnected by bufferless optical core nodes that perform fast switching in order to provide bandwidth allocation in sub-wavelength granularities. The architectures, tools and methods being developed for its operation are described, as well as the issues to be solved.

Full text

AN AGILE ALL-PHOTONIC NETWORK
Trevor J. Hall, Sofia A. Paredes, Gregor v. Bochmann
Photonic Network Technology Laboratory, Centre for Research in Photonics
School of Information Technology and Engineering, University of Ottawa, Canada
{thall , sparedes , bochmann}@site.uottawa.ca
ABSTRACT
This paper presents an overview of recent and
current work being conducted in the “Agile All-Photonic
Networks”, AAPN, Research Network. An AAPN is a
wavelength division multiplexed network that consists of
several overlaid stars formed by edge nodes that
aggregate traffic, interconnected by bufferless optical
core nodes that perform fast switching in order to
provide bandwidth allocation in sub-wavelength
granularities. The architectures, tools and methods
being developed for its operation are described, as well
as the issues to be solved.
Keywords: optical networks, WDM, TDM, OBS,
traffic aggregation, bandwidth allocation, scheduling.
1. INTRODUCTION
Most of the existing photonic networks have a mesh
topology, which distributes the traffic load over many
switches but warrants the use of complex routing
algorithms and possibly a large number of O-E-O
conversions. If the switches in the core of a photonic
network have enormous capacity; an overlaid star
topology is possible while maintaining robustness [1].
Complex routing is not necessary in this kind of network
and there is no need to resort to technologies that are not
realizable in the near future, such as optical memory and
optical header recognition.
The Agile All-Photonic Networks, AAPN, research
project [2] proposes such a star topology in which the
photonic core space switches are all-optical and rapidly
reconfigurable. This proposal is based on the observation
that optical switching technologies will mature to a level
where they will be able to introduce “agility” [3]; i.e.,
the ability to perform time domain multiplexing to
dynamically allocate and share the bandwidth of each
wavelength by several information flows to achieve a
higher degree of data flow granularity and be able to
adapt it to traffic flows as the demand varies.
AAPN envisions the use of fast switches in the core
nodes to provide the agility and capacity required; as
opposed to current photonic networks that are relatively
static in their configuration and whose components have
the luxury of relaxed requirements for switching times,
settling times and clock acquisition times.
Ideally, switching would be performed on a
per-packet basis to achieve the finest granularity and
hence the most efficient bandwidth utilization. However,
this would require switching times in the order of
nanoseconds, which is not a viable technology yet. With
this in mind, the constraints on switching speed are
relaxed to some extent by gathering variable sized
packets destined for the same output to form fixed-length
slots. Switching in the core node is then performed for
slots and not for individual packets.
Schemes for provision of subwavelength bandwidth
granularity have been developed, for example in [4],
where a time slotted WDM mesh network is described.
This approach differs from AAPN in the use of ultra-fast
tunable lasers at the edge nodes to avoid fast switching at
the network core. In [5], a meshed WDM network is
partitioned into a number of clusters. Specific nodes that
serve as gateways between clusters undertake the
coordination of frame switching and end-to-end routing.
The core network formed by the gateways is in this case,
as opposed to AAPN, a mesh.
2. ARCHITECTURE
The AAPN consists of a number of hybrid photonic/
electronic edge nodes connected together via a
wavelength stack of bufferless transparent photonic
switches placed at the core nodes (a set of space
switches, one switch for each wavelength), of which
there is a small number.
B
87
A
5
3
2
1
6
Core Node
Edge Node
4
Figure 1. Overlaid star topology that characterises an Agile
All-Photonic Network
Each edge node contains a separate buffer for the
traffic destined to each of the other edge nodes. Traffic
aggregation is performed in these buffers, where packets
are collected together in slots or bursts that are then
transmitted as single units across the network.
The connectivity of a core node is reconfigured in
response to traffic load variations as reported by the edge
nodes. The core node also coordinates the transmission
(Invited)
ICOCN2005, 14-16 December 2005, Bangkok, Thailand
365

actions in the edge nodes.
Since switch fabrics with large port counts cannot
provide the switching speed required at the core, port
sharing is required to allow a core node to support large
numbers of edge nodes. A selector may therefore be
used between the edges and the core (Figure 2) to
combine the traffic of multiple edges onto a single fibre.
Figure 2. Use of selectors to allow for large numbers of edge
nodes.
Another benefit of using a star topology is that, in the
context of a single star, the routing of packets is trivial
and therefore may be left to higher level protocols, e.g.
MPLS; in which case the whole star may be viewed as a
single MPLS switch and the same labels may be used.
2.1 Performance objectives / assumptions
The topology of AAPNs may include hundreds of edge
nodes operating at 10Gbps (and even up to 100 Gbps in
future upgrades). The optical switching time in the core
node is aimed at 1µs. Real time traffic monitoring will
be implemented for the correct adaptation of the core
nodes connectivity to the most up-to-date traffic needs.
For the slotted mode, the size of the slot is 10µs for links
with rates of 10Gbps and the number of timeslots per
frame is 100 (and up to 1000).
The photonic path lengths will be of up to thousands
of kilometers; therefore all-optical 3R (as well as
dynamic compensation for transmission impairments)
will be required for long-haul. Wavelength conversion
will be avoided in principle; although this may change as
the technology matures. C and L band wavelengths with
50 to 100 GHz channel spacing will be used.
2.2 Issues
2.2.1 Switch speed and Scalability
A major concern in choosing a technology and
architectural option for the switch fabric in the core is its
scalability to large port counts, which is limited by two
main performance degradations: insertion loss and
crosstalk. The former can be compensated all optically,
but preventing the accumulation of crosstalk requires
signal regeneration or O-E-O conversion. It is therefore
necessary to estimate the crosstalk limit that has to be
achieved for the signal to remain in the optical domain.
The major performance impairment in optical
switches, the in-band optical crosstalk, was analysed
using a statistical crosstalk model for three architectural
and technological options for the implementation of
AAPN core switches [6]. The study shows that for DC
switches (wavelength layered), single switch ports up to
64 can be realised with acceptable crosstalk and port
counts can be scaled up using three-stage switches. For
larger number of wavelengths and fibre counts,
wavelength dimensional switches are more efficient.
2.2.2 Synchronization
Given the physical topology it is sufficient for the
propagation delays between each edge node and the core
node to be known and for each edge node to maintain a
local clock locked to the core node clock. The locking of
the clocks and the determination of the propagation
times may be done using a suitable synchronization
signaling protocol. Co-ordination of transmissions at the
core node is achieved by scheduling transmissions at the
edge nodes using the local clock and the known
propagation delays.
2.3 Determining the Optimum Layout
The initial step for deployment of an AAPN is solving
the layout design problem to determine network
parameters such as optimal number, size, and placement
of edge nodes, selectors/multiplexers, core nodes as well
as placement of the DWDM links; the aim being to
minimize network costs while satisfying performance
requirements. The solution to this problem is determined
by demographic and economic factors.
A mixed integer linear programming formulation is
presented in [7] for core node placement and link
connectivity. A number of possible solutions and their
costs are discussed for a wide variety of equipment cost
assumptions for both metropolitan and long-haul
networks with a gravity model for traffic distribution.
The model was solved also for actual population
information and geographical coordinates that were
obtained from a census database of 140 Canadian cites.
3. BANDWIDTH SHARING
3.1 Optical Burst Switching (OBS-AAPN)
Optical Burst Switching (OBS) is a technique where
several packets with the same destination and other
common attributes such as Quality of Service parameters
are assembled into a “burst” (essentially a very large
packet) and forwarded through a bufferless network as
one entity. The header and the associated payload are
sent on two different wavelength channels with the
header being sent ahead in time.
3.1.1 Burst aggregation
Burst aggregation in an OBS network can be timer-based
or threshold-based. In a timer-based approach, a burst is
created and sent into the optical network at periodic time
intervals; which produces variable length bursts and
therefore might yield undesirable burst lengths at
different loads. In a threshold-based approach, a limit is
placed on the length of each burst; which produces
ICOCN2005, 14-16 December 2005, Bangkok, Thailand
366

fixed-size bursts but does not give any guarantee on the
delay that the packets will experience during the
aggregation process.
A composite burst assembly mechanism that
combines both approaches in AAPN is discussed in [8].
Results with both analytical and simulation models show
that the delay experienced by the packets can be
appropriately bound by choosing a time-out comparable
to the maximum tolerable delay and that only the packets
in the low rate flow suffer this delay. It is also shown
that the hybrid mechanism reserves resources for bursts
for smaller periods of time, therefore reducing blocking
probabilities or reservation delays.
3.1.2 Scheduling
OBS may or may not use a two-way reservation of
resources. The use of acknowledgements eliminates
burst loss but increases overall delays. Simple two-way
reservation approaches have been shown to work with
acceptable delays for short core-edge distances.
Some modifications of retransmission schemes are
shown in [9] for OBS-AAPN, where a framework
without any loss of bursts exploits the entire network
capacity. The results indicate that for high loads the
average burst transmission delay and the buffer capacity
are highly dependent on the network diameter and thus
confirm that these schemes are suitable for the local or
metropolitan cases.
3.2 Time Division Multiplexing (TDM-AAPN)
In a time-slotted mode of operation, each star of the
AAPN may be seen as a distributed three-stage Clos-like
packet switch: the edge nodes can be logically split in
two parts (the source and the destination modules) and
the core node may be explicitely drawn with its
de-multiplexers/multiplexers and its wavelength space
switches in parallel. The connections between the
respective source/destination edge nodes and the core
node are seen now as unidirectional (Figure 3).
Figure 3. An AAPN star viewed as a distributed three- stage
Clos switch.
This architecture is esentially the same as the one
presented in [10] (and others referenced therein) and
hence the same Flexible Bandwidth Provision scheduling
algorithm may be used. Timing considerations must be
made in this case: since all wavelengths and ports in the
core node start a new transmission slot at the same time
and the edge nodes may be located at varying distances
from the core, scheduling in this modality requires
synchronization between the core and the edge nodes to
allow the edge nodes to know the time when a new slot
should be transmitted to the core.
3.2.1 Slot aggregation
The same hybrid mechanism for burst aggregation can
be extended to encapsulate variable length packets into
‘envelopes’ matched to the time slots in a TDM-AAPN.
Emulation results are presented in [8] for this process
using real IP network traffic from a local LAN using
encapsulation methods with and without packet
segmentation. Bandwidth utilization measures confirm
that the model with packet segmentation is more
bandwidth-efficient (even if the processing delay is
slightly larger) and the simplicity of the model suggests
that a low cost software implementation of this process
would be efficient.
3.2.2 Scheduling
Typically, TDM scheduling involves the following steps:
traffic matrix estimation (bandwidth request), service
matrix construction (bandwidth allocation), and
decomposition into configuration matrices (connectivity
for the core switches).
Traffic estimation is mainly done using queue state
information gathered at the edge nodes. If the
propagation delays are large and the traffic information
is considered out-of-date; it is possible to use algorithms
for prediction of traffic loads like the one described in
[11], which uses an approach based on traffic sampling
and distributed expectation-maximization for predicting
the resource requirements of end-to-end flows.
The service matrix defines the bandwidth allocated to
each flow in units of slots within a frame; i.e., the larger
the number of slots, the larger the bandwidth granted to
that flow. The construction of this matrix, can be made
using many mathematical approaches or virtually any
heuristic approach (or a mix of both) [12][13].
The service matrix, must be decomposed into its
correspondent constituents, which will define the
connectivity (schedule) of each space switch for each
time slot. There are also a number of mathematical and
heuristic approaches to solve this problem [10][13].
Ideally, the core switches are scheduled on a per-slot
basis, but the bandwidth requests may be “out-of-date”
because of signaling delays. For long-haul scenarios with
large propagation delays it may therefore be a better
option to take a frame-based approach and reconfigure
the central nodes every frame instead of every slot.
A comparison of various OBS and TDM methods for
bandwidth sharing in AAPN is reported in [14].
4. PROTOTYPE
A demonstrator is currently being built with the aim of
showing that the technologies, architectures and control
protocols can be combined into an operational network.
The AAPN prototype will be the testbed for
synchronization protocols, bandwidth allocation
methods, traffic monitoring, routing protocols and fault
ICOCN2005, 14-16 December 2005, Bangkok, Thailand
367

recovery methods. This work is currently in its first
phase, in which the transmission platforms will be built
with off-the-shelf components. Figure 4 shows the
preliminary design of the edge node modules and their
interactions. Subsequent work in this area will involve
the incorporation of newly designed optical devices and
components developed also by the AAPN team [2].
Transmission & synchronization layer
Generic
Edge
Control
Module
Burst/slot
buffering and
transmission
Burst/slot
reception and
packetization
Packet
aggregation
Special
Edge
Control
Module
Global AAPN layer
IP traffic layer
Figure 4. Edge node architecture for the prototype.
5. FURTHER ON
The results presented in [7] correspond to the
deployment of new infrastructure for the realisation of
the AAPN. This is, however, a costly alternative and
therefore it is of great interest to study how an AAPN
can be built by using and adapting the already deployed
WDM technology. Migration strategies is therefore an
important ongoing research topic.
In the context of a TDM-AAPN, if the load is
perfectly balanced, then each space switch in the core
node would have the same schedule. In this case, a single
wavelength-independent crossbar could be used in the
core node to switch the wavelength multiplex as a whole,
thus simplifying the scheduling problem significantly.
This is called Photonic Slot Routing [15] and its
application to AAPN, is currently under investigation.
6. SUMMARY AND REMARKS
The architectures, tools and methods so far designed for
the correct and efficient operation of an Agile
All-Photonic Network have been described. Topology
dimensioning, traffic aggregation, bandwidth allocation,
scheduling methods and prototyping have been
discussed. The proposed overlaid star architecture
provides high bandwidth connectivity in sub-wavelength
granularities by dynamically reconfiguring the photonic
network according to the traffic demands.
It is important to note that a large amount of work in
the development of enabling technologies (optical
switches, transmission and amplification) is being
carried out by several investigators of the AAPN
Research Network but these topics are out of the scope
of this paper and have not been discussed. The authors
refer the reader to the AAPN website [2] for information.
7. ACKNOWLEDGEMENTS
This work was supported by the Natural Sciences and
Engineering Research Council (NSERC) and industrial
and government partners, through the Agile All-Photonic
Networks (AAPN) Research Network. Dr. Trevor Hall
holds a Canada Research Chair in Photonic Network
Technology at the University of Ottawa. He and Sofia
Paredes are grateful to the Canada Research Chairs
Programme for their support of this work.
8. REFERENCES
[1] R. Vickers and M. Beshai, “PetaWeb architecture,” 9th Int
Telecom. Netw. Planning Symp., Toronto, Canada, 2000.
[2] The Agile All-Photonic Networks (AAPN) Research
Network, http://www.aapn.mcgill.ca/, 2003-2005.
[3] G. v. Bochmann, et al. “The Agile All-Photonic Network:
An architectural outline”, in Proc. 22nd Bien. Symp. on
Comm., Kingston, Canada, 2004, pp. 217-218.
[4] I. Saniee and I. Widjaja, “A New Optical Network
Architecture that Exploits Joint Time and Wavelength
Interleaving,” in Proc. OFC, Los Angeles, 2004.
[5] A. Stavdas, H. Leligou, K. Kanonakis, C.
Linardakis, and J.D. Angelopoulos, "Scheme for
performing statistical multiplexing in the optical layer," J.
Opt. Netw., 4(5), 237-247, 2005.
[6] R. Shankar and T.J. Hall, “Core Switch Architectures for
the Agile All-Photonic Network: Performance Evaluation
with Crosstalk”, Indicon 2005, Chennai, India, 2005.
[7] L.G. Mason, A. Vinokurov, N. Zhao and D. Plant,
“Topological design and dimensioning of Agile All-
Photonic Networks”, J. of Comp. Comm., Special issue on
Optical Networking, 2005 (accepted).
[8] S. Parveen, R. Radziwilovicz, S.A. Paredes, T.J. Hall.
"Evaluation of burst aggregation methods in an optical
burst switched agile all-photonic network". SPIE
Photonics North 2005, Toronto, 2005.
[9] A. Agustí-Torra, G. v. Bochmann and C. Cervelló-Pastor,
“Retransmission schemes for Optical Burst Switching
over star networks”, IFIP Int. Conf. Wireless Opt.
Commun. Netw., 2004.
[10] S.A. Paredes and T.J. Hall, “Flexible bandwidth provision
and scheduling in a packet switch with an optical core”, J.
of Opt. Netw., 4(5), 260-270, 2005.
[11] T. Ahmed, N. Saberi, M.J. Coates, “Time-slot reservation
in all-photonic networks based on flow prediction”, 2nd
Workshop on Optim. of Opt. Netw., Montreal, 2005.
[12] N. Saberi and M.J. Coates, “Bandwidth Reservation in
Optical WDM/TDM Star Networks”, in Proc. 22nd Bienn.
Symp. on Comm, Kingston, 2004, pp. 219-221.
[13] C. Peng, S.A. Paredes, G. v. Bochmann and T.J. Hall,
“Bandwidth Provisioning in the Core of an Agile
All-Photonic Network”, submitted to Elsevier’s Optical
Switching and Networking, 2005.
[14] X. Liu, A. Vinokurov and L.G. Mason, “Performance
Comparison of OTDM and OBS Scheduling for Agile
All-Photonic Networks”, Proc. IFIP MAN Conference
HCNC, Vietnam, 2005.
[15] H. Zang, J.P. Jue and B. Mukherjee, “Capacity
Allocation and Contention Resolution in a Photonic Slot
Routing All-Optical WDM Mesh Network,” J. of Lightw.
Tech., 18(12), 1728-1741, 2000.
ICOCN2005, 14-16 December 2005, Bangkok, Thailand
368