Content uploaded by Leila Askari
Author content
All content in this area was uploaded by Leila Askari on Nov 04, 2019
Content may be subject to copyright.
A Techno-Economic Evaluation of VNF Placement
Strategies in Optical Metro Networks
Leila Askari, Francesco Musumeci, Massimo Tornatore
Department of Electronics, Information and Bioengineering, Politecnico di Milano, Milan, Italy
E-mail: firstname.lastname@polimi.it
Abstract—Network Function Virtualization (NFV) has
changed the way operators can provision network services.
Decoupling network functions from dedicated hardware and
running them on software, on top of commodity servers
and switches, not only helps operators have more flexible
and easy-to-manage networks, but also reduces their capital
and operational expenditures. This is especially true for
incoming 5G services, characterized by ultra-low latency,
high reliability and bandwidth requirements. To satisfy these
challenging requirements, multi-layer optical networks based
on Optical Transport Network (OTN) over wavelength division
multiplexing (WDM) are being deployed in the metro segment
to support 5G services. In addition, the possibility to equip
metro nodes with computing capabilities, enabled by new
paradigms such as CORD (Central Office Re-architected
as a Datacenter) is being exploited. In this scenario, an
efficient placement of Virtual Network Functions (VNFs) for
Service Chain (SC) provisioning within the metro network
is needed, and different VNF placement strategies can lead
to different costs for network operators. In this paper we
analyze the impact of different VNF placement strategies on the
optical metro network cost, considering specific Service Level
Agreement (SLA) requirements, expressed in terms of service
blocking probability. We provide a cost model which takes
into consideration both capital and operational expenditures.
Through extensive numerical results, we quantify the impact
of using a cost-effective VNF placement strategy in decreasing
network cost while meeting the desired SLA performance.
Index Terms—NFV, VNF placement, Dynamic Service Chain-
ing, cost analysis, metro network
I. INTRODUCTION
In the last decade, two new networking paradigms have
attracted the attention of researchers and practitioner, i.e.,
Network Function Virtualization (NFV) and Software Defined
Networking (SDN), for their ability to lead to more agile
and flexible networks. SDN simplifies network control by
decoupling control plane from data plane [1]. NFV enables
network operators to achieve flexibility and cost saving by
replacing dedicated hardware with software running on top
of commodity server and switches [2]. Traditional network
functions such as, e.g., firewall, Network Address Translation
(NAT), Intrusion Detection System (IDS), etc., can be im-
plemented in software installed on general-purpose hardware,
leading to a simplification of the management of network
function and reduction of costs. These softwarized functions
are called Virtual Network Functions (VNFs). Several of
future-generation (i.e., 5G) network services are characterized
by ultra-low latency, high bandwidth requirements and high
availability and they can be supported through an ordered
sequence of VNFs, which is called Service Chain (SC).
To satisfy the stringent requirements of 5G services, op-
erators are deploying multi-layer optical networks based on
Optical Transport Network (OTN) over Wavelength Division
Multiplexing (WDM). Also, the architecture of the metro
network nodes is evolving towards the concept of Central
Office Re-architected as a Datacenter (CORD) [3], i.e., central
offices are now equipped with processing and storage units. In
this context, achieving a dynamic and flexible (i.e., network-
status-aware) VNF placement for SC provisioning in metro
networks is not a trivial task, as a trade-off arises between:
i) network transport capacity, ii) processing units, i.e., ca-
pacity of network nodes hosting VNFs (called NFV-nodes
in the following), and iii) required Service Level Agreement
(SLA), e.g., expressed in terms of service blocking probability
and/or maximum tolerated service latency. Furthermore, the
SC provisioning should be performed by limiting the overall
network cost, including both Operational Expenditure (OpEx)
and Capital Expenditure (CapEx).
While most of the studies in the literature consider the
impact of placement only on the OpEx (e.g., power consump-
tion) [4], in this paper we evaluate how different placement
strategies can impact both the OpEx and the CapEx, which we
consider as the cost of activating an NFV-node and bandwidth
cost, respectively.
Hence, we perform a techno-economic evaluation of dif-
ferent VNF placement strategies considering a realistic metro
network topology under different physical network settings, in
terms of number of NFV-nodes and number of wavelengths
per link.
The reminder of the paper is organized as follows. In
Section II we provide an overview of related works. Section
III describes the proposed cost model we considered in the
paper. Numerical results obtained comparing different VNF
placement algorithms in different network settings are then
discussed in Section V. Finally, we conclude the paper in
Section VI.
II. RELATED WORK
The problem of VNF placement has attracted the interest of
many researchers in recent years. Some of the existing works
formulate the problem as an optimization model and provide
optimal or near-optimal solutions. For example, authors in
[5] provide a binary integer programming model for optimal
VNF placement with the objective of minimizing expensive
978-1-7281-0875-9/19/$31.00 ©2019 IEEE
Core Network
AMEN
Metro Node
Data Center
MCEN
Metro Node Metro Node
Fixed and Wireless Access
Src VNF1 VNF2 Dst
VNF2VNF1
PU
PU
PU
PU
Fig. 1: Overview of the metro network architecture and VNF mapping into the physical network
optical-electrical-optical conversions. Since the problem of
VNF placement is proven to be NP-hard [6], they also provide
a heuristic algorithm to solve the problem of optimal VNF
placement. Ref. [7] proposes an Integer Linear Programming
(ILP) model for VNF placement that considers latency re-
quirements of SCs with the objective of minimizing resource
utilization. Authors in [8] propose a heuristic algorithm for
VNF placement to consider VNF interference caused by VNF
consolidation. Their algorithm tries to maximize the through-
put of provisioned service requests. Ref. [9] provides two
heuristic algorithms for scalable VNF placement to be able to
accommodate more dynamically arriving user requests. Ref.
[4] provides two algorithms for VNF placement and allocation
problem aiming at minimizing number of VNFs deployed
while satisfying all the data flows in the network.In [10]
authors provide meta-heuristic solution, with the objective
of maximizing the utilization of nodes hosting VNFs and
minimizing the number of nodes that host VNFs in the net-
work. Ref. [11] proposes a Mixed Integer Linear Programming
model for VNF placement that considers QoS requirements of
the VNFs and tries to optimize the resource utilization.
A number of existing works have considered deployment
cost as a constraint in placing VNFs. For example, Ref.
[12] provides two algorithms to perform VNF placement
considering both processing capacity limitation of the nodes
hosting VNFs and budget constraint. In [13] authors provide
an ILP model that satisfies the reliability requirements of the
SC with the objective of minimizing SC orchestration cost.
In addition, there have been some efforts performing cost
analysis of NFV. For example Ref. [14] provides a techno-
economic analysis of a 5G network infrastructure based on
SDN/NFV. Authors in [15] present analysis on performance
and cost of a cloud network system based on NFV. To the best
of our knowledge, none of the existing works has evaluated
the impact of the VNF placement strategy in combination with
the SLA requirements on the network cost.
III. NETWORK MODEL
In this paper we consider a metro network architecture
as defined in the context of Metro-HAUL project [16] and
depicted in Fig. 1. In this optical metro-network there are three
categories of nodes: the Metro Core Edge Nodes (MCENs)
that are gateways towards core network, Access Metro Edge
Nodes (AMENs), that constitute the interfaces between the
metro network and heterogeneous access networks (e.g., fixed
and/or mobile access), and Metro Nodes (MNs), that represent
transport metro nodes and, unlike MCENs and AMENs are
not equipped with processing units (PU).
A. Service chaining model
A SC is composed of different VNFs (virtual nodes)
connected together using virtual links in a specific order. In
order to provision a SC we need to deploy its VNFs on the
NFV-nodes, that we assume are chosen among AMENs and
MCENs.
In such an optical metro network, we focus on a dynamic
traffic environment where SCs are dynamically generated at
forwarding nodes, which constitute the source of the SC.
Based on the SC type, an NFV-node is chosen as the desti-
nation of SC. Moreover, according on the SC type, a specific
end-to-end maximum latency and a total required bandwidth
characterize the SC being provisioned. As depicted in Fig.
1, to provision a SC, its constituting VNFs are mapped (i.e.,
deployed) to NFV-nodes and SC traffic traverses the various
VNFs in a predefined order, while satisfying the latency and
bandwidth requirements of the SC.
Figure 2 shows an example of how two different SC
requests are provisioned in the physical network. In this paper
we consider an Optical Transport Network (OTN) over Wave-
length Division Multiplexing (WDM) architecture, where each
node (either MCEN, AMEN or MN) is constituted by a
Digital Cross Connect (DXC) over an Optical Cross Connect
(OXC). This architecture is being used by China Mobile for its
first deployment of 5G metro aggregation networks [17]. The
figure shows an upper SC layer, where the sequence of VNFS
for the two SCs are highlighted, and the lower OTN/WDM
layer.
Assuming the aforementioned OTN/WDM physical archi-
tecture, at each transit node the transported traffic is con-
verted from the optical to the electronic domain (OE), it is
electronically processed by a DXC and, if necessary, it is
groomed (respectively, degroomed) with traffic inserted (resp.
OXC
SC layer
OTN/WDM layer
Src 1 VNF1 Dst1
VNF2Src 2 Dst2
OXCOXCOXCOXC
OXCOXC
OXCOXC OXCOXC
OXCOXC
OXCOXC
OXC
Physical
Link
Digital
Cross Connect
Optical
Cross Connect
OXCOXC
1
2
3
4
56
7
8
VNF3VNF3VNF3
SC1 Traffic
Flow
SC2 Traffic
Flow
Fig. 2: Node architecture
dropped) locally. Then, it is converted back into the optical
domain (EO) and switched towards the next node by an OXC1.
Moreover, all the nodes are assumed as having full wavelength
conversion capability.
In the example of Fig. 2, the two SCs have different
source/destination end nodes (i.e., nodes 1-8 for SC1 and
nodes 2-7 for SC2), but share the processing units for one
of their VNFs, namely, VNF3. As shown in the figure, after
the insertion of traffic at the two source nodes, traffic is OE
and EO converted in all the transit nodes, also including the
ones where no grooming is performed, i.e., nodes 4 and 3,
where the traffic of SC1 is also processed at the SC layer by
VNF1. Moreover, note that at node 5 the two traffic flows are
groomed as they share the physical link between node 5 and
6, and additionally the traffic of SC2 is sent towards the SC
layer to perform traffic processing at VNF2. Finally, when
arriving at node 6, traffic flows are firs OE converted and
processed by the shared VNF3, then they are EO converted
and inserted into two different lightpaths to be sent towards
the two different destinations.
B. Cost model
In this section we provide a model representing the cost of
VNF placement for service chaining. We consider two main
contributions to the total cost, i.e.:
•Active NFV-nodes: deploying a VNF instance to an NFV-
node requires a certain amount of virtual machines to be
used. Each of these virtual machines is equipped with
limited processing capacity and the VNF deployment
leads to operational costs [18], e.g., due to energy con-
sumption and/or software license usage.
•Bandwidth: in order to provision a SC, traffic needs to be
routed through the network, hence, a certain amount of
transport capacity (i.e., a certain number of wavelengths
per link) is required for SC provisioning [19].
Therefore the cost related to the provisioning of one SC
and the corresponding VNFs placement can be formulated as
follows:
1Note that traffic is OE and EO converted at any node also in case no
traffic needs to be added/dropped at that node
CSC =Cnodes +Cwl (1)
where Cnode represents the cost related to the utilization
of NFV-nodes (e.g., due to power consumption, software
licences, etc., [20]), and Cwl represents the cost of the
wavelengths required for traffic transport, namely, due to the
transponders to be installed at the nodes [21]. Note that, in this
paper, we do not consider the capital expenditures due to the
network and computing equipment (i.e., switches, routers and
servers) as we assume it is not affected by the VNF placement
strategy being adopted. Conversely, as we will detail in the
following, the VNF placement strategy has a strong impact
on the amount of active NFV-nodes and utilized bandwidth.
IV. DYNAMIC VNF PLACEMENT STRATEGY
In this section we briefly describe the VNF placement
strategy introduced in our previous work [22] and used for our
cost evaluation. We assume dynamically arriving SC requests,
where each SC is characterized by its source/destination
nodes, the SC type (i.e., the ordered sequence of VNFs to
be used, its bandwidth requirement and maximum latency)
and its duration. Upon the arrival of a given SC requests, the
Dynamic VNF Placement (DVNFP) algorithm will consider
current network state (i.e., current set of deployed SCs with
the corresponding provisioned lightpaths and VNFs) and try to
deploy the incoming request at a minimum cost, while meeting
its bandwidth and latency constraints.
The DVNFP algorithm consists of two phases, as described
in the following (the reader is referred to [22] for further
details). A high level flowchart of the DVNFP algorithm is
depicted in Fig. 3.
1) VNF placement. In the first phase, the algorithm
performs the VNF placement for each of the VNFs in the
SC, which are analysed in order. For each VNF, the algorithm
first tries to reuse an already-active VNF instance on an NFV-
node which is the closest NFV-node both to the source and
destination of the SC request. If such VNF instance is not
found, the algorithm tries to activate a VNF instance on one of
the NFV-nodes located along the shortest path between source
and destination nodes of the SC request (namely, SPs,d ).
In the case that no NFV-node is available on SPs,d, based
on the SC requirements, an NFV-node is chosen. In other
words, if the SC requires high computational capacity (e.g.,
for Augmented Reality SC) the NFV-nodes closer to the core
nodes are selected as these nodes are more likely to have
large computational capacity. However, if the SC has stringent
latency requirements, the NFV-node closer to the source of SC
will be selected to host the required VNF instance. The above-
mentioned step will be repeated for each VNF of SC until an
appropriate NFV-node is chosen for all the VNFs in the SC.
If it is not possible to find any NFV-node to place VNFs of
the SC, the SC request is blocked.
2) Lightpaths provisioning and VNF adjustment. In
this phase, the end-to-end latency of the path traversing all
the VNFs of the SC in the required order (namely, Le2e)
is calculated, considering propagation and switching latency
contributions as in [22]. If Le2eis higher than the latency
NEW SC
REQUEST
VNF
ALREADY
PLACED?
Yes
FAILED?
Yes
No
SELECT NEXT
VNF
SCALE UP EXISITNG
VNF & UPDATE
NETWORK STATE
PLACE ON ANOTHER NFV-
NODE AND UPDATE NETWORK
STATE
FAILED?
Yes
LAST
VNF?
BLOCK
REQUEST
Yes
No
EVALUATE
LATENCY
LATENCY
OK?
Yes
No
TRY TO GROUP
VNFs IN FEWER
NODES
CAN
CONSOLIDATE
MORE
?
No
Yes
No
No
PROVISION SC AND
RELEASE RESOURCE WHEN
HOLDING TIME EXPIRES
INCREASE THE COUNTER FOR
LATENCY VIOLATED SCs
Fig. 3: DVNFP algorithm flowchart
requirements of the SC (LSC), the algorithm tries to shorten
the end-to-end path by consolidating (i.e., placing in the
same NFV-node) one or more VNFs, until Le2eis lower
than or equal to LSC. When all the VNFs of the SC are
consolidated on one NFV-node algorithm calculates the Le2e
and, if it is still higher than LSC, a counter for latency-
violation is increased while the SC is provisioned. Then the
SC is provisioned, i.e., its resources (bandwidth and used
virtual machines) are allocated. The SC resources are then
released after the SC holding time expires.
V. N UMERICAL RESULTS
A. Case study and simulation settings
To perform our analysis, we developed a discrete-event-
driven simulator in C++. We considered a full OTN network
topology as the one in Fig. 4, including 52 nodes and 72
bidirectional WDM links. As depicted in the figure, three
types of nodes are present in this network, i.e., Metro Core
Backbone (MCB), Metro Core (MC) and Metro Aggregation
(MA) nodes. AMEN nodes are chosen among MA nodes and
MCEN nodes are chosen among MBC nodes. Each NFV-node
is equipped with 64 CPU cores (except the MCB, which is
assumed as an NFV-node with unlimited processing resources)
and each WDM links supports Wwavelengths with 10 Gbit/s
capacity. The number of wavelengths per link Wis a tuned
parameter in this paper, as we vary it to evaluate the bandwidth
cost required to support a certain traffic amount with a
given SLA. More specifically, in all the following simulations
(i.e., for each different VNF placement strategy and in each
different NFV-node scenario) the number of wavelengths per
link (i.e., the value of W) is set as the minimum number able
to support the required SLA.
The types of SCs considered in our evaluation are shown
in Table I, where we detail the required VNFs as well as their
latency and bandwidth requirements. In addition, the computa-
tional requirements of the various VNFs are depicted in Table
II in terms of percentage of CPU cores. The considered VNFs
are Network Address Translation (NAT), Intrusion Detection
TABLE I: SC and corresponding VNFs, bandwidth and latency
characteristics
Service Chain Service Chain VNFs Bandwidth Latency
Augmented Reality NAT-FW-TM-VO-IDS 100 Mbps 1ms
MIoT NAT-FW-IDS 100 Mbps 5ms
Smart Factory NAT-FW 100 Mbps 1ms
TABLE II: Percentage of CPU core usage for various VNFs
VNF Name NAT FW VO TM IDS
CPU Core 0.0184 0.018 0.108 0.266 0.214
System (IDS), Firewall (FW), Video Optimizer (VO) and
Traffic Monitor (TM).
We simulate the dynamic arrival of SC requests, where the
arrival instants are randomly generated considering a Poisson
distribution with mean inter-arrival λ=40SC requests per
second, while the holding time for each SC is generated
according to a negative-exponential distribution with mean
μ=1 second2.
For an incoming SC request, the SC type is randomly
selected among these SCs with equal probability, whereas the
source node of the SC requests is randomly selected among
MA nodes. Moreover, based on the SC type, the destination
node can be either among MCB nodes or nodes closer to
the source of the SC request. In other words, if the SC has
stringent latency requirements the destination node is chosen
as close as possible to the source node of the SC request.
All the results are obtained with a confidence level of 95%
with at most 5% confidence interval on blocking probability.
B. Evaluation metrics
In order to perform cost analysis, we consider that SLA
requirements defined for the users require that the blocking
probability (defined as the number of blocked SC request out
of total SC requests) is below a certain threshold, initially
2This traffic intensity has been chosen as, for the considered network
topology and SC characteristics, it can be supporter with a maximum blocking
probability of around 10−3.
3
1
94
5
6
7
2
17
45
28
11
41
12
49
40
52
14
24
15
16
18
19
20
21
22
25
27
32
33 34
37
38
39
41
43
44
46
47
48
50
51
Extension
Link
Core
Link
8
10
36
31
29
35
3042
23
26
13
Metro Core
Backbone
Metro Core
Metro
Aggregation
Fig. 4: Network topology
set to 10−3. We consider three different metrics to evaluate
the performance of our algorithm with respect to other VNF
placement strategies, namely:
i) average number of active NFV-nodes (Navg ), that is calcu-
lated as follows:
Navg =SC∈SCprov
NSC ·tSC
ttot
(2)
Here, SC ∈SCprov represents a generic SC in the set of
provisioned (i.e., non-blocked) SC, NSC is the number of
active NFV-nodes that have at least one running VNF instance
at each time instant, tSC is the time between arrival of two
consecutive SC requests and ttot is the total simulation time.
Hence, the formula calculates the number of active NFV-nodes
weighted by the amount of time each NFV-node is serving;
ii) Number of wavelengths per link,W;iii) latency violation
ratio, calculated as the ratio between provisioned SC requests
with violated latency out of the total number of provisioned
SC requests.
In addition to the above-mentioned metrics, we also consid-
ered total network cost for SC provisioning, which is obtained
based on the following equation:
Ctot =αNavg +W·L(3)
where L=72is the total number of links in the network
and parameter αcaptures the relative costs of bandwidth
and active NFV-nodes, i.e., the higher α, the higher is the
importance of Navg in the overall cost.
C. Benchmark VNF placement strategies
To perform our cost evaluation we also consider two
benchmark VNF placement algorithms [22], i.e.:
•Centralized: this strategy is used to evaluate the case
with the lowest possible number of active NFV-nodes.
Specifically, we assume that the network has only one
NFV-node with unlimited computational capacity and
located at the MCB (node2), so that all the VNFs are
TABLE III: NFV-node selection scenarios
Scenario S1 S2 S3 S4
% of MCENs 100% 75% 50% 25%
% of AMENs 0% 25% 50% 75%
# of AMENs 0 5 11 17
# of MCENs 6 4 3 2
# of Datacenters 2 2 2 2
Total # of NFV-nodes 8 11 16 21
embedded at that node. Therefore, in this case, all the
SC requests are directed towards a single NFV-nodes
which will result in higher bandwidth requirements, due
to typically longer routes, and higher blocking, due to
potential bottlenecks in links closer to the NFV-node.
•Distributed: in this algorithm, the main objective is
to reduce network blocking by deploying VNFs along
the shortest paths between SCs’ source/end nodes. This
comes at the cost of higher number of active NFV-nodes,
as VNF instances are typically activated in all NFV-
nodes.
D. Discussion
We conduct the experiments considering four scenarios, in
which a different number of NFV-nodes are present in the
network for the cases of DVNFP and Distributed algorithms.
In the first scenario (S1) we considered that all the MCENs are
NFV-nodes. The second scenario (S2) indicates that 75% of
MCENs and 25% of AMENs are NFV-nodes. Third (S3) and
forth (S4) scenarios, respectively, represent the case where
half of MCENs and half of AMENs are NFV-nodes and
the case where 25% of MCENs and 75% of AMENs are
NFV-nodes. It is worth mentioning that the MCB nodes
are always considered as NFV-nodes (i.e., they are assumed
as datacenters with unlimited computational capacity). We
summarize these scenarios in Table III.
Figure 5 illustrates the comparison between the different
VNF placement strategies considering a target blocking prob-
ability equal to 10−3. The average number of active NFV-
nodes is plotted in Fig. 5(a) for the three strategies. Note that
for DVNFP the number of active NFV-nodes is up to 22%less
than Distributed. This is due to the fact that DVNFP tries to
reuse the already active NFV-nodes as much as possible and
demonstrates the importance of an effective VNF placement
strategy.
On the other hand, as depicted in Fig. 5(b), in order
to satisfy the SLA requirements (i.e., the maximum service
blocking probability), for all the scenarios S1 to S4, the
DVNFP and Distributed algorithms require the same num-
ber of wavelengths per link W. This means that, for the
considered traffic scenarios, the DVNFP algorithm is able to
reduce the cost of active NFV-nodes without impacting on
the required bandwidth. From the figure it is evident that the
Centralized algorithm has the lowest performance in terms of
bandwidth cost, as it requires 30 wavelengths per link, i.e.,
three times more than DVNFP and Distributed cases. This is
due to the fact that, with the Centralized strategy, only one
NFV-node is used to perform VNF placement, therefore a
higher number of wavelengths is needed to avoid congestion
at the links in its proximity.
S1 S2 S3 S4
0
10
20
NFV-node selection scenario
Avg Num of Active NFV-nodes
Centralized DVNFP Distributed
(a) Average number of active NFV-nodes
S1 S2 S3 S4
0
10
20
30
40
NFV-node selection scenario
Num of wavelengths
Centralized DVNFP Distributed
(b) Number of wavelengths per link
S1 S2 S3 S4
0%
10 %
20 %
30 %
40 %
NFV-node selection scenario
Latency violation ratio
Centralized DVNFP Distributed
(c) Latency violation ratio
Fig. 5: Comparison among different strategies
0
500
1000
1500
2000
2500
020 6080100
Ctot
40
A
DVNFP Distributed Centralized
(a) NFV-nodes S1 scenario
0
500
1000
2000
2500
0 20406080100
Ctoƚ
A
DVNFP Distributed Centralized
1ϱ00
(b) NFV-nodes S2 scenario
0
500
1000
1500
2000
2500
0 20406080100
Ctot
A
DVNFP Distributed Centralized
(c) NFV-nodes S3 scenario
0
500
1000
2000
2500
0 20406080100
Ctoƚ
A
DVNFP Distributed Centralized
1ϱ00
(d) NFV-nodes S4 scenario
Fig. 6: Total cost comparison for different VNF placement strategies and NFV-nodes scenarios
Latency violation ratio is shown in Fig. 5(c) for the various
cases. As shown in the figure, the Distributed strategy achieves
the best performance, due to the fact that SC requests are
provisioned deploying VNFs closer to the source node of
the request. Therefore, the shortest path between source and
destination of the SC request is often used for traffic routing,
which leads to the lower latency violation ratio. Finally,
the violation ratio obtained in the DVNFP case is slightly
higher than the Distributed case (around 1-2% higher), and is
independent from the NFV-node selection scenario.
Now we evaluate the impact of the different strategies on
the total network cost described in section V-B. In Fig. 6
we show how the total network cost, as defined in eq. 3, is
affected by the different VNF placement strategies, tuning the
parameter αfrom 1 to 100 to capture the importance of the
two cost contributions in eq. 3.
It is evident that in most cases the Centralized strategy
provides the highest overall network cost, whereas DVNFP
one is in general the most cost-effective solution. However, for
increasing α, i.e., when the cost of active nodes becomes more
relevant than the wavelengths cost, the difference between
DVNFP/Distributed and Centralized strategies is reduced. In
general (i.e., except for the S1 scenario), this reduction is
quicker for the Distributed case than for the DVNFP one, due
to the difference in the number of active nodes between the
two cases.
Moreover, the difference between the Centralized and other
VNF placement strategies becomes lower as we move from
S1 to S4. In particular, as depicted in Fig. 6(a) for the S1
scenario, i.e., when only eight NFV-nodes are deployed, the
total cost for DVNFP and Distributed strategies is the same,
even for increasing α. This is due to the equal values of Navg
and Lrequired in the S1 scenario for the two strategies, as
we observed in Fig. 5. However, increasing the number of
available NFV-nodes, i.e., changing the network deployment
from scenario S1 towards S4, we can see the difference
between DNVFP and Distributed strategies increases up to
16% for S4 scenario when the relative cost of active nodes is
high (α= 100). This is due to the fact that DVNFP is capable
of satisfying the same SLA (i.e., guarantee the same maximum
blocking probability) as Distributed using the same number
of wavelengths per link but using less active NFV-nodes. It
is noteworthy that the Centralized strategy has almost always
the highest total network cost (up to 66% higher than the
other strategies in the S1 scenario and with α=1), since it
requires almost three times more wavelengths per link with
respect to the DNVFP and Distributed strategies. However,
for S4 and for higher values of α(e.g., for α= 100), since
Distributed activates on average 18 NFV-nodes and uses 8
wavelengths per link, the total network cost for Centralized
(which requires 1 NFV-node and 30 wavelengths per link),
is lower if compared to the Distributed case. It is worth
noting that, for the Centralized strategy, the various NFV-
nodes scenarios provide almost no impact on the total cost,
even for increasing α, as the most relevant cost contribution
in this case is constituted by the number of wavelengths in
the network.
Comparing the various costs obtained for the DVNFP
strategy, the maximum difference between the various NFV-
nodes scenarios and for the lowest value for α(i.e., α=
1) is equal to 18% and occurs between S1 and S4. This
demonstrates that a higher number of available NFV-node, in
general, does not always lower the network cost, although it
may guarantee higher flexibility in VNFs placement and thus
higher efficiency in network capacity utilization, due to the
fact that shorter routes can be used for SC provisioning. The
cost gap between S1 and S4 cases for the DVNFP strategy
increases up to 23% for the highest value of αconsidered
(α= 100). The reason is that, for increasing α, the cost
of activating one NFV-node has higher impact, as in the S4
scenario more NFV-nodes are activated with respect to S1.
We perform a similar analysis considering a more stringent
SLA requirement, namely, a maximum blocking probability
target of 10−5. Results are shown in Fig. 7 for the various
cases. As shown in the figure, a more stringent SLA produce
a fixed increase in the overall cost for all the strategies and
in all the NFV-nodes scenarios, even for increasing values
of alpha. This means that the SLA variation impacts only
the number of wavelengths per link, without impacting the
DVNFP-Pb,target=10-
3
Distributed-Wď,target=10-3
Centralized-Pb,target=10-3
DVNFP-Pb,target=10-5
Distributed-Pb,target=10-5
Centralized-Pb,target=10-5
0
500
1000
1500
2000
2500
0 20 40 60 80 100
Ctot
A
(b) NFV-nodes S1 scenario
0
500
1000
1500
2000
2500
0 20 40 60 80 100
Ctot
A
(c) NFV-nodes S2 scenario
Ϭ
ϱϬϬ
ϭϬϬϬ
ϭϱϬϬ
ϮϬϬϬ
ϮϱϬϬ
Ϭ ϮϬ ϰϬ ϲϬ ϴϬ ϭϬϬ
ƚŽƚ
A
(d) NFV-nodes S3 scenario
0
500
1000
1500
2000
2500
3000
0 20406080100
Ctot
A
(e) NFV-nodes S4 scenario
Fig. 7: Impact of SLA on total network cost
average number of active nodes.
VI. CONCLUSION
In this paper we presented a techno-economic analysis of
different VNF placement strategies for SC provisioning, con-
sidering different NFV-nodes deployments and SLAs under
realistic optical metro network topology and traffic assump-
tion. For the considered traffic, results show that an efficient
placement strategy can reduce the cost of service provisioning
up to 16% or 23%, according to the various NFV-nodes
deployments.
ACKNOWLEDGMENT
The work leading to these results has been supported by
the European Community under grant agreement no.761727
Metro-Haul project.
REFERENCES
[1] J. Matias, J. Garay, N. Toledo, J. Unzilla, and E. Jacob, “Toward
an SDN-enabled NFV architecture,” IEEE Communications Magazine,
vol. 53, no. 4, pp. 187–193, 2015.
[2] Network functions virtualisation. [Online]. Available:
https://www.etsi.org/technologies/nfv
[3] L. Peterson, A. Al-Shabibi, T. Anshutz, S. Baker, A. Bavier, S. Das,
J. Hart, G. Palukar, and W. Snow, “Central office re-architected as a data
center,” IEEE Communications Magazine, vol. 54, no. 10, pp. 96–101,
2016.
[4] Y. Sang, B. Ji, G. R. Gupta, X. Du, and L. Ye, “Provably effi-
cient algorithms for joint placement and allocation of virtual network
functions,” in IEEE INFOCOM 2017-IEEE Conference on Computer
Communications. IEEE, 2017, pp. 1–9.
[5] M. Xia, M. Shirazipour, Y. Zhang, H. Green, and A. Takacs, “Network
function placement for NFV chaining in packet/optical datacenters,”
Journal of Lightwave Technology, vol. 33, no. 8, pp. 1565–1570, 2015.
[6] F. Bari, S. R. Chowdhury, R. Ahmed, R. Boutaba, and O. C. M. B.
Duarte, “Orchestrating virtualized network functions,” IEEE Transac-
tions on Network and Service Management, vol. 13, no. 4, pp. 725–739,
2016.
[7] D. B. Oljira, K.-J. Grinnemo, J. Taheri, and A. Brunstrom, “A model
for QoS-aware VNF placement and provisioning,” in 2017 IEEE
Conference on Network Function Virtualization and Software Defined
Networks (NFV-SDN). IEEE, 2017, pp. 1–7.
[8] Q. Zhang, F. Liu, and C. Zeng, “Adaptive interference-aware VNF place-
ment for service-customized 5G network slices,” in IEEE INFOCOM
2019-IEEE Conference on Computer Communications. IEEE, 2019,
pp. 2449–2457.
[9] W. Zhou, Y. Yang, M. Xu, and H. Chen, “Accommodating dynamic
traffic immediately: A VNF placement approach,” in ICC 2019-2019
IEEE International Conference on Communications (ICC). IEEE, 2019,
pp. 1–6.
[10] S. Tavakoli-Someh and M. H. Rezvani, “Multi-objective virtual net-
work function placement using NSGA-II meta-heuristic approach,” The
Journal of Supercomputing, pp. 1–37, 2019.
[11] F. B. Jemaa, G. Pujolle, and M. Pariente, “QoS-aware VNF placement
optimization in edge-central carrier cloud architecture,” in 2016 IEEE
Global Communications Conference (GLOBECOM). IEEE, 2016, pp.
1–7.
[12] G. Sallam and B. Ji, “Joint placement and allocation of virtual network
functions with budget and capacity constraints,” in IEEE INFOCOM
2019-IEEE Conference on Computer Communications. IEEE, 2019,
pp. 523–531.
[13] X. Zhong, Y. Wang, and X. Qiu, “Cost-aware service function chaining
with reliability guarantees in NFV-enabled Inter-DC network,” in 2019
IFIP/IEEE Symposium on Integrated Network and Service Management
(IM). IEEE, 2019, pp. 304–311.
[14] C. Bouras, P. Ntarzanos, and A. Papazois, “Cost modeling for SDN/NFV
based mobile 5G networks,” in 2016 8th International Congress on
Ultra Modern Telecommunications and Control Systems and Workshops
(ICUMT). IEEE, 2016, pp. 56–61.
[15] M. Ananth and R. Sharma, “Cost and performance analysis of network
function virtualization based cloud systems,” in 2017 IEEE 7th Inter-
national Advance Computing Conference (IACC). IEEE, 2017, pp.
70–74.
[16] The Metro-Haul project deliverables. [Online]. Available: https://metro-
haul.eu/deliverables/
[17] C. Zhang. (2018) “Optical Networking in the Cloud and 5G Era,”
Keynote talk at Optical Fiber Communications Conference and Ex-
hibition (OFC). [Online]. Available: https://www.ofcconference.org/en-
us/home/ news-and-press/ofc-video-library
[18] B. Addis, G. Carello, F. De Bettin, and M. Gao, “On a virtual network
function placement and routing problem: properties and formulations,”
2017.
[19] A. Leivadeas, M. Falkner, I. Lambadaris, G. Kesidis, C.-H. Lung, and
M. Ibnkahla, “Considerations for a successful network service chain
deployment,” in 2018 4th IEEE Conference on Network Softwarization
and Workshops (NetSoft). IEEE, 2018, pp. 288–292.
[20] A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel, “The cost of a
cloud: research problems in data center networks,” ACM SIGCOMM
computer communication review, vol. 39, no. 1, pp. 68–73, 2008.
[21] P. Yi, H. Ding, and B. Ramamurthy, “Cost-optimized joint resource
allocation in grids/clouds with multilayer optical network architecture,”
IEEE/OSA Journal of Optical Communications and Networking, vol. 6,
no. 10, pp. 911–924, 2014.
[22] L. Askari, A. Hmaity, F. Musumeci, and M. Tornatore, “Virtual-
network-function placement for dynamic service chaining in metro-area
networks,” in 2018 International Conference on Optical Network Design
and Modeling (ONDM). IEEE, 2018, pp. 136–141.