356 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 10, NO.

4, DECEMBER 2013

Optimal Virtual Network Embedding:

Node-Link Formulation
Márcio Melo, Susana Sargento, Ulrich Killat, Andreas Timm-Giel, and Jorge Carapinha

Abstract—Network Virtualization is claimed to be a key potential reduction of costs and increase of revenues, as an
component of the Future Internet, providing the dynamic support interesting tool from an operational point of view [6], [7].
of different networks with different paradigms and mechanisms Although there is a large interest on virtualized networks
in the same physical infrastructure. A major challenge in the
dynamic provision of virtual networks is the efficient embedding both from the research community and network operators,
of virtual resources into physical ones. Since this problem is several challenges still prevent them from being deployed in
known to be N P-hard, previous research focused on designing real environments [8]. One of the major obstacles lies in the
heuristic-based algorithms; most of them either do not consider efficient embedding1 of a Virtual Network (VN) onto a phys-
a simultaneous embedding of virtual nodes and virtual links, or ical network. Since this process requires the simultaneous op-
apply link-path formulation, leading to non-optimal solutions.
This paper proposes an integer linear programming (ILP) for- timization of virtual nodes and links placement, it is complex
mulation to solve the online virtual network embedding problem in nature, both in formulation and computationally. Several
as a result of an objective function striving for the minimization works, such as [9]–[17], have already proposed solutions to
of resource consumption and load balancing. To this end 3 this problem, mostly based on heuristic approaches; however,
different objective functions are proposed and evaluated. This they do not provide the optimal solution for each VN mapping.
approach applies multi-commodity flow constraint to accomplish
a node-link formulation that optimizes the allocation of physical This paper focuses on the online embedding of VN requests
network resources. in the physical network. An Integer Linear Programming
This proposal is evaluated against state of the art heuristics. (ILP) formulation, the Virtual Network Embedding Node-Link
The performance of the heuristics related to Virtual Network Formulation (VNE-NLF), is used to solve the VN assignment
(VN) request acceptance ratio is, at least, 30% below the one problem on the basis of a minimization of resource consump-
of the Virtual Network Embedding Node-Link Formulation
(VNE-NLF) method. From the three cost functions evaluated, tion and load balancing strategy. The VNE-NLF includes link
the Weighted Shortest Distance Path (WSDP) is the one which delay constraints and supports the specification of the max-
embeds more VNs and also requires, on average, less physical imum distance between virtual nodes. In addition, different
resources per embedding. cost functions are proposed and analyzed, which enforce load
Index Terms—Virtual networks, embedding, mapping, NP- balancing of links and nodes, and shortest distance paths.
hard, optimization, ILP model, heuristics. Simulation experiments show how far the state of the art
heuristics differ from an ILP method. If the VN request
I. I NTRODUCTION acceptance ratio is used as a measurement metric, the solutions
obtained by the state of the art heuristics are, at least, 30%

N ETWORK VIRTUALIZATION has gained an increas-

ing prominence in networking and telecommunications
fields in the last few years. Initially, the interest in network
below the ones of the VNE-NLF (see Fig. 3). From the
cost functions evaluated, the Weighted Shortest Distance Path
(WSDP) is the one which embeds more VNs and also requires,
virtualization was mainly pushed by Future Internet research
on average, less physical resources per embedding. Compared
initiatives [1]–[4], mainly with the objective to find a platform
to our previous work in [18], this paper:
on which novel Internet architectures could be experimented
and evaluated without limitations or constraints, namely those i extends the mathematical formulation to support two new
associated with the traditional IP model. Later on, it became constraints, i.e. link delay and maximum distance between
clear that virtualization could constitute a key component of nodes;
the next-generation Internet architecture itself [5], and not just ii proposes three new cost functions, i.e. Load Balancing
as a mere platform for experimentation. It also became clear plus  Shortest Path (LB+SP), Shortest Distance Path
for network operators that network virtualization could provide (SDP), and WSDP;
a number of short/medium term business advantages, with iii defines a new evaluation metric, i.e. the embedding factor,
which represents the amount of resources that have been
Manuscript received January 2, 2013; revised June 24, 2013. The associate requested over the amount of resources that have been
editor coordinating the review of this paper and approving it for publication leased;
was D. Medhi.
M. Melo and J. Carapinha are with Portugal Telecom Inovação, Aveiro, iv and it provides a performance comparison with 6 state-of-
Portugal, (e-mail: {marcio-d-melo, jorgec}@ptinovacao.pt). the-art heuristics.
S. Sargento is with the Instituto de Telelcomunicaç ões, University of
Aveiro, Portugal (e-mail: [email protected]). The contributions of this paper can be summarized as
U. Killat and A. Timm-Giel are with the Institute of Communication follows:
Networks, Hamburg University of Technology, Germany (e-mail: {killat,
timm-giel}@tuhh.de). 1 The terms embedding, mapping and assignment are used interchangeably
Digital Object Identifier 10.1109/TNSM.2013.092813.130397 in this paper.
1932-4537/13/$31.00 
c 2013 IEEE

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
MELO et al.: OPTIMAL VIRTUAL NETWORK EMBEDDING: NODE-LINK FORMULATION 357

i ILP optimization for virtual network embedding, which network, and enables path splitting (i.e. virtual link composed
is based on node-link formulation; it enables the simul- by different paths) and link migration (i.e. to change the
taneous embedding of virtual nodes and links, which underlying mapping) during the embedding process. However,
optimizes the allocation of physical network resources, i.e. this level of freedom can lead to a level of fragmentation that
Central Processing Units (CPUs) on the physical nodes is unfeasible to manage large scale networks.
and bandwidth on the physical links. In [13] a formal approach is taken to solve the on-line
ii Analysis and evaluation of different objective goals: load VN mapping problem using a mixed integer programming
balancing objective function LB+SP, shortest path ob- formulation. Chowdhury et al. applied a two step approach
jective function SDP, and load balancing combined with to embed VNs on the substrate. In the first step, the virtual
shortest path objective function WSDP; nodes are assigned to physical nodes and in the second step
iii Comparison of the performance of existing VN embedding the virtual links are assigned to physical paths. Compared
solutions, heuristics, with a pure ILP formulation. to the previous state of the art heuristics, i.e. [9], [12], the
iv Evaluation metrics that relate e.g. VN acceptance ratio and formulation proposed by Chowdhury et al. provides a better
link utilization. Moreover, a new metric, the embedding coordination of the two phases, since an “augmented substrate
factor, is proposed. graph construction” is used.
The rest of the paper is organized as follows. After summa- The approach in [13] completely differs from the math-
rizing the related works in section II, section III describes the ematical formulation proposed in this paper, which applies
virtual network embedding problem, the notations and param- a node-link formulation. In our approach, the universe of
eters used, and the embedding process. Section IV describes embedding solutions is considered within the ILP formulation,
the proposed VNE-NLF and the applied constraints, while and the VN embedding problem is solved in a single step using
section V-A presents and discusses the different proposed cost the multi-commodity flow constraint and by considering the
functions. Section VI analyzes the performance of the VNE- notion of direction of the flows.
NLF with different cost functions, and compares it with six Butt et al. [14] proposed a topology aware heuristic for VN
existing heuristics. Finally, section VII concludes the paper mapping, and also suggested algorithms to avoid bottlenecks
and describes the future work. on the physical infrastructure, where they consider virtual node
reallocation and link reassignment for this purpose. Nogueira
et al. [15] proposed a heuristic that takes into account the het-
II. R ELATED W ORK
erogeneity of the VNs and also of the physical infrastructure.
This simultaneous node and link mapping optimization can The heuristic is evaluated by means of simulation and also on
be formulated as an un-splittable flow problem [9], known a small scale testbed, where it achieves mapping times of the
to be N P-hard. In order to solve this problem, several order of tens of milliseconds.
approaches have been suggested, mostly considering the off- Botero et al. [16] proposed an algorithm to solve the VN
line version of the problem where the VN requests are fully mapping problem, which also considers the CPU demand
known in advance. of the hidden hops. Chowdhury et al. [17] extended his
In [10] a backtracking method based on sub-graph isomor- preliminary results [13] and included a generalized window-
phism was proposed; it considers the on-line version of the based VN embedding to evaluate the effect of look ahead on
mapping problem, where the VN requests are not known in the mapping of VNs.
advance, and proposes a single stage approach where nodes Alkmim et al. [20] proposed a mathematical formulation
and links are mapped simultaneously, taking constraints into that aims to: i) map virtual routers and virtual links; ii)
consideration at each step of the mapping. When a bad minimize the bandwidth consumption; and iii) minimize the
mapping decision is detected, a backtrack to the previous valid time required to instantiate a virtual router. In contrast to
mapping decision is made, avoiding a costly re-map. this work, we also aim to optimize link load and CPU load
The work in [11] defined a set of premises about the virtual distribution.
topology, i.e. the backbone nodes are star-connected and the Although all these algorithms provide a solution for the VN
access-nodes connect to a single backbone node. Based on mapping problem, an optimal solution for the embedding task
these premises, an iterative algorithm is run, with different and its efficiency is not provided. Also, some of them fail to
steps for core and access mapping. However, the algorithm solve the assignment problem as a simultaneous optimization
can only work for specific topologies. of the virtual node and link placement, which leads to non-
A distributed algorithm was studied in [19]. It considers optimal solutions.
that the virtual topologies can be decomposed in hub-and- The VNE-NLF applies a node-link formulation to solve
spoke clusters and each cluster can be mapped independently, the VN embedding problem in a single step using the multi-
therefore reducing the complexity of the full VN mapping. commodity flow constraint. This approach provides the opti-
This proposal has lower performance when compared with mal solution for the objective function used, since the universe
centralized approaches. of solutions is considered within the ILP formulation.
Zhu et al. [9] proposed a heuristic based on a centralized
algorithm to deal with VN mapping. The goal of the algorithm III. N ETWORK D ESCRIPTION AND P ROBLEM
is to maintain a low and balanced load of both nodes and links F ORMULATION
of the substrate network. Yu et al. [12] proposed a mapping In this section, we introduce the virtual network embed-
algorithm which considers finite resources in the physical ding problem. In addition, the VN embedding notations used

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
358 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 10, NO. 4, DECEMBER 2013

sEZĞƋƵĞƐƚƐ

͘͘͘
dŝŵĞсƚ
sEŵďĞĚĚŝŶŐ DĂƉƉŝŶŐZĞƐƵůƚ
ϭϬ ϮϬ ^LJƐƚĞŵ ϭϬ ϮϬ ϮϬ
Ă
ĂϯϬ ϯϬ ď
Ě ϯϬ Ğ ď ϭϱ
Ϯϱ
Đ
WŚLJƐŝĐĂůEĞƚǁŽƌŬ ϭϬ ϱ

sE>ŝĨĞƚŝŵĞсϰϬϬ
sEƌĞƋƵĞƐƚŬ ϱϱ ϴϬ ϵϱ
͘͘͘

ϱϱ ϴϬ ϵϱ
sE
 ϵϱ  ϳϬ   ϵϱ  ϳϬ 
ŵďĞĚsEZĞƋƵĞƐƚ DĂƉƉŝŶŐ
ϳ͘ϱ
ϱϬ ϵϬ ϱϬ ϴϬ DĞƚŚŽĚ
ď
ϭϬ ϭϱ ϱϬ ϵϬ ϱϬ ϴϬ  ϲϬ  ϲϱ &

ϴϱ ϳϱ ϲϱ
Ă Đ sEƐƐŝŐŵĞŶƚ/ŶĨĞĂƐŝďůĞ
ϭϬ ϱ  ϲϬ  ϲϱ &

sE>ŝĨĞƚŝŵĞсϲϱϬ ϴϱ ϳϱ ϲϱ sEƐƐŝŐŵĞŶƚ&ĞĂƐŝďůĞ
sEƌĞƋƵĞƐƚϭ dŝŵĞсϬ
hƉĚĂƚĞWŚLJƐŝĐĂůŶĞƚǁŽƌŬ

Fig. 1. VN Embedding System—topology example. ůůŽĐĂƚĞsE

hƉĚĂƚĞsE>ŝƐƚ
throughout the paper are presented, and the virtual network
ZĞůĞĂƐĞsE
embedding system is explained. Finally, the mapping goals
are introduced to support the mathematical formulation.
zĞƐ sE>ŝĨĞƚŝŵĞ EŽ
A. Network Description džƉŝƌĞĚ͍

We use superscript to distinguish the physical network from

the virtual network, where p and v correspond to physical and
virtual, respectively. Fig. 2. VN request life cycle—activity Diagram.
1) Physical network: A physical network can
be described as a weighted undirected graph
Gp = {N p , Lp , C p , B p , Dp , Disp } composed by a set given lifetime that is, in principle, independent from each
of physical nodes, N p , and a set of physical links, Lp . Each other, and each lifetime could have different time scales,
physical node i is characterized by its processing capacity, since it is strongly dependent on the purpose of the virtual
Cip , commonly referred to as the CPU, and by its physical network request itself. If we consider a VN request for a live
location, which can be defined by x and y coordinates. rock concert, the time scale will be hours, but if we consider
The distance between virtual nodes, Disp , can be obtained a VN for a culinary workshop of one week, the time scale
using expression (1). With respect to the physical links, we will be days.
p
consider that each link ij has a given bandwidth, Bij , and 3) VN Assignment Notations: First, we start with the con-
p
a given link delay, Dij , and we also assume that each link vention used for the index notation: N p represent the set of
is an undirected link. The bottom-right of Fig. 1 illustrates a nodes that belong to the physical network; Lp represent the set
physical network topology example composed of 6 physical of links that belong to the physical network; and Lpi represent
nodes and 8 physical links, and the corresponding capacities a subset of links ij that are directly connected to the node
of the nodes and the links are presented on top of the i. The same type of notation is used to represent the VN
elements. using the letters m and n in the virtual network. The notations
 used throughout this paper for the VN assignment problem
Dispij = (xj − xi )2 + (yj − yi )2 (1) are presented in table I. The table is divided into three parts:
the static parameters of the physical network, the dynamic
2) Virtual Network Request: VN request can parameters of the physical network, and the virtual network
be described as a weighted undirected graph requests with the demanded capacities.
Gv = {N v , Lv , C v , B v , Dv , Disv } composed by a set
of virtual nodes, N v , and a set virtual links, Lv . Each virtual
node m is characterized by the amount of required CPU, Cm v
, B. Unfilled Physical Network Resources
and the virtual links mn are logical connections between The remaining capacity of each physical node at a specific
virtual nodes and characterized by the amount of dedicated time t is given by the difference between the total processing
v
bandwidth, Bmn , and by the maximum link delay permitted, capacity and the capacity consumed by all virtual nodes allo-
v
Dmn . We also assume that each virtual link is an undirected cated on that physical node, and is presented in expression (2),
link. The maximum distance between virtual nodes, Disv , can where u represents the set of all virtual nodes allocated on that
be used to limit the number of intermediate hops between precise physical node and at time t.
virtual nodes. The left part of Fig. 1 represent the example of

two virtual network requests, VN request 1 on the bottom-left ∀i ∈ N p : Cip (t) = Cip (0) − Cuv (t) (2)
and VN request k on the top-left. Each VN request has a u

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
MELO et al.: OPTIMAL VIRTUAL NETWORK EMBEDDING: NODE-LINK FORMULATION 359

TABLE I
VN A SSIGNMENT P ROBLEM N OTATION . have, at least, the same amount of bandwidth available as re-
quired by the virtual link which is presented in expression (5).
Gp Physical Network
Np Set of Physical Nodes p
i, j Physical Nodes ∀ij ⊆ M[mn ∈ Lv (k)] : Bmn
v
(k) ≤ Bij (t) (5)
ij Physical Link
Lp Set of Physical Links
Lpi Set of Physical Links directly connected to Physical D. VN Request Life Cycle
Node i
Cip (0) Total CPU of Physical Node i The embedding process begins upon a new VN request
p
Disij Distance Between Physical Nodes ij arrival, which is depicted in Fig. 2. A VN mapping method
p
Bij (0) Total Bandwidth of Physical Link ij
(e.g., heuristic) is used to embed the VN; it takes as inputs
Cip (t) Available CPU at time t on Physical Node i
p
Bij (t) Available Bandwidth at time t on Physical Link ij the current status of the physical network (e.g. available CPU
Gv (k) Virtual Network Request k capacity and existing bandwidth) and the VN request itself.
N v (k) Set of Virtual Nodes of VN Request k If the result of the mapping process is a viable solution, the
Lv (k) Set of Virtual Links of VN Request k
Lvm (k) Set of Virtual Links directly connected to Virtual node
mapping is considered to be feasible; if not, it is considered
m of VN Request k to be unfeasible and the VN embedding process stops.
m, n Virtual Nodes
mn Virtual Link
v (k)
Cm CPU of Virtual Node m of VN Request k E. Mapping Metrics
Disvmn (k) Maximum Distance Between Virtual Nodes mn of VN
Request k In order to assess the performance of an embedding method,
v (k)
Bmn Bandwidth of Virtual Link mn of VN Request k
v (k) and at the same time to compare it with others, different
Dmn Delay of Virtual Link mn of VN Request k
performance metrics were defined.
1) VN Request Acceptance Ratio: The VN request accep-
In parallel, the available bandwidth of each physical link at tance ratio, AVN , is given by expression (6) and defines the
a specific time t is given by the difference between the total overall performance of an embedding method: the number of
bandwidth and the bandwidth consumed by all virtual link VN requests accepted, k  , over the number of all VN requests,
segments allocated on that physical link, and is presented in k.
expression (3), where w represents the set of all virtual link k
segments allocated on that specific physical link and at time AVN = (6)
k
t.
A virtual link can be composed of one or more physical 2) Embedding Factor: The embedding factor, E VN , is given
links, physical path. We consider that each virtual link has a by expression (7) and represents the ratio between the amount
single physical path, and we do not consider link aggregation of virtual resources that were requested for the VN and the
(i.e. virtual link composed by different physical paths). amount of physical resources that were effectively provisioned
One physical link could accommodate one or more virtual to accommodate that VN, i.e. the efficiency on embedding.
link segments belonging to different virtual links. The parameters, α, β, γ and η, are used to weight the different
p p
 types of resources.
∀ij ∈ Lp (k) : Bij (t) = Bij (0) − Bwv
(t) (3)
 v
 v
w
VN α m Cm + β mn Bmn
E =  p  p (7)
γ i Ci + η ij Bij
C. VN Request Embedding Process
The VN request embedding process can be divided into two IV. V IRTUAL N ETWORK E MBEDDING - M ATHEMATICAL
components: the component that ensures the mapping of the F ORMULATION
virtual nodes, and the one that handles the mapping of the
virtual links. This section describes the mathematical formulation devel-
1) Virtual Node Mapping: Each virtual node needs to be oped to solve the online VN embedding problem with the
mapped onto one physical node, this relation is given by defined constraints.
the mapping function M[m ∈ N v (k)] = i, where virtual An Integer Linear Programming (ILP) approach is used to
node m is mapped onto physical node i. Each physical node solve the online VN embedding problem; we propose a node-
candidate needs to have, at least, the same amount of available link formulation, and two assignment variables are applied
CPU as required by the virtual node, which is represented in during the embedding process. The index notation used here
expression (4). is the same as in section III-A3.

∀i, ∀M[m ∈ N v (k) = i] : Cm

v
(k) ≤ Cip (t) (4) A. Assignment Variables
2) Virtual Link Mapping: Each virtual link can be mapped The binary variable x is used in the mapping of the virtual
onto one or more physical links (i.e. physical path), this nodes and is defined in expression (8), where xm
i → N ×N
V P

relation is given by the mapping function M[Lvmn ], where matrix. With respect to the virtual links, the binary variable
mn
the virtual link mn is mapped onto one physical path. Each y is used and it is represented in equation (9), where yij →
V 2 P 2
physical link candidate belonging to the physical path needs to (L ) × (L ) matrix.

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
360 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 10, NO. 4, DECEMBER 2013

1) Virtual Node Assignment: 5) Assignment of virtual links to physical links - multi-

 commodity flow conservation with node-link formulation:
m 1, virtual node m is allocated at physical node i To simultaneously optimize the mapping of virtual links and
xi =
0, else virtual nodes, the multi-commodity flow constraint [21] is
(8) applied with a node-link formulation [22]; moreover, the
2) Virtual Link Assignment: notion of direct flows on the virtual links is used, which is
 represented in Eq. (14), where Lvm represents all the virtual
mn 1, virtual link mn uses physical link ij links that are directly connected with the virtual node m, and
yij = (9) Lpi represent all the physical links that are directly connected
0, else
with the physical node i
∀mn ∈ Lvm , m < n, ∀i :
B. Constraints 
mn mn
(yij − yji ) = xm n
i − xi (14)
To assure the correct mapping of the virtual nodes and of ij∈Lp
i
the virtual links, and also to obey to the conservation law on
the capacities of the physical nodes and physical links, a set 6) Bandwidth conservation: To ensure that the available
of constraints is defined. bandwidth at each physical link is not surpassed, Equation
(15) is defined.
1) Assignment of virtual nodes to physical nodes: Equation
(10) ensures that each virtual node is assigned, and that it is
assigned to just one physical node. ∀ij ∈ Lpi , i < j :

 v
Bmn mn
(yij mn
+ yji p
) ≤ Bij (15)
∀m : xm
i = 1 (10)
mn∈Lv
m ,m<n
i
v
7) Link delay limit: The virtual link delay, Dmn , is a
2) One virtual node per physical node: Equation (11) parameter of the VN embedding problem, and is equal to the
guarantees that each physical node can accommodate in the sum of the delay of all physical links that compose the virtual
maximum one virtual node per VN request, although each link. To ensure that the constraint on the link delay is not
physical node can accommodate other virtual nodes from violated we apply equation (16).
different VNs. This constraint is used to ensure that each
virtual node is assigned to a different physical node per VN
embedding, and can be suitable in application scenarios where ∀mn ∈ Lvm , m < n, ∀i :
 p mn mn v
it is required to have physical node diversity for redundancy Dij (yij + yji ) ≤ Dmn (16)
reasons. ij∈Lp
i ,i<j

∀i : xm
i ≤ 1 (11)
m V. V IRTUAL N ETWORK A SSIGNMENT - O BJECTIVE
F UNCTION
3) CPU conservation: Equation (12) assures that the avail-
One of the major challenges when formulating an ILP
able CPU capacity of each physical node is not exceeded.
model for VN assignment resides in the definition of the objec-
 p tive function: the allocation of resources need to be optimized
∀i : xm v
i · Cm ≤ Ci (12)
m
in order to support the efficiency of the corresponding VN
process.
4) Virtual Node distance: Equation (13) assures that the Moreover, the correct specification of the VN mapping
v
maximum distance between virtual nodes, Dmn , is not vi- constraints (see IV) is also a challenge of this approach. In this
olated. The maximum distance between virtual nodes is a section, we describe the main goals that need to be achieved
parameter of the VN embedding problem. The effect of this when formulating an objective function for virtual network
parameter on the VN embedding will be studied on a separate embedding; three different objective functions are proposed
section (see VI-C). to achieve these goals.
This parameter is given in distance units and can be used
to express the maximum radius between virtual nodes (in the A. Objective Goals
simulated scenario the location of the physical nodes is set
A primary goal for the embedding algorithm is to minimize
in a grid). The distance between physical nodes, i.e. Dispij , is
resource consumption in order to have resources available
obtained using formula (1), and K represents a large constant
for forthcoming VN embedding requests. Minimization of
which is used only in situations where the virtual node n is
resource consumption is only possible for the bandwidth
not mapped at physical node i, i.e. xni = 0.
consumption depending on the number of links involved in
an embedding process. The processing power has just to be
installed exactly in the amount required by the VN request on
∀m, n ∈ Lvm , m < n, ∀i :
 some physical nodes.
Dispij · xm v n n
j ≤ Dismn · xi + (1 − xi ) · K (13) Resource minimization consequently means that the VN’s
j should exhibit minimal hop counts on their paths. This in turn

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
MELO et al.: OPTIMAL VIRTUAL NETWORK EMBEDDING: NODE-LINK FORMULATION 361

means that almost every physical node should be available to D. Weighted Shortest Distance Path (WSDP)
host a virtual node. As long as the resources required by VN’s The objective function WSDP, proposed in equation (19), is
are small compared to physical capacities of nodes and links, similar to equation (18), although here the demanded capacity
this availability is guaranteed with high probability by a load by the VN is included in the objective function. This has the
balancing strategy, which results in some spare capacity for effect that high demands are allocated to nodes or links with
each physical node or link. a large amount of free capacity.
Therefore, the dominating aspects in the formulation of an
objective function for the ILP problem are the minimization  
  xm
of bandwidth consumption and load balancing. v i
minimize α Cm (k) p +
m i
Ci (t)
⎛ ⎡ ⎤ ⎞
B. Load Balancing plus  Shortest Path (LB+SP)   yij mn
β⎝ v
Bmn (k) ⎣ p
⎦⎠ (19)
mn ij
B ij (t)
The objective function LB+SP is proposed in expression
(17), and it achieves two goals: the primary goal is to minimize
the maximum load per physical resources; in the case of differ- VI. E VALUATION R ESULTS
ent mapping solutions with the same maximum utilization, the In this section, we describe the simulation scenario, the
second part of the objective function is activated which will evaluation metrics, and depict our major results. We compare
opt for the solution which consumes the lowest bandwidth. the VNE-NLF model in its several versions with six state of
p
Bp
LCmax represents the overall maximum node load; Lmax the art methods.
represents the overall maximum link load. The parameters
Cip (0), Bij
p v
(0), Cm v
(k), Bmn (k) were defined in Table I; the
parameter  represents a small constant, which should be small A. Simulation Parameters
enough to not affect the first objective; and the parameters To evaluate the VNE-NLF model, we have implemented
α and β are used to weight the load cost of each type of a discrete event simulator in Matlab R
, with the proposed
resources. formulation using different objective functions.
The physical network topology is created using the GT-ITM
p p  tool [23], the number of physical nodes is set to 50, which
minimize α · LC B
max + β · Lmax +  ·
mn
yij v
· Bmn (t), is representative of a medium scale infrastructure provider,
mn and the link probability between two physical nodes is set

Cip (t) + m v
m xi · Cm (k) p to 0.5. The node CPU capacity and the link bandwidth are
∀i ∈ N p : p ≤ LC max
Ci (0) real numbers uniformly distributed between 50 and 100. The
p  mn v VNs requests are also representative of either small or medium
p
B ij (t) + mn yij · Bmn (k) p
∀ij ∈ L : p ≤ LBmax (17) scale virtual networks, and are created using the same topology
Bij (0)
generation method; the number of virtual nodes is not fixed,
but follows a uniform distribution, from 2 to 10 virtual nodes
C. Shortest Distance Path (SDP) per VN topology; the virtual link probability is set to 0.5.
The CPU capacity of the virtual nodes and the bandwidth of
The previous objective function (17) works well in sit- the virtual links are also real numbers uniformly distributed
uations where there are abundant resources in the physical between 0 and 20, and between 0 and 50, respectively2. The
network. Then, bandwidth consumption is of no concern and considered values for the bandwidth and for the CPU are nor-
load balancing is beneficial because it gives a high degree of malized, since the objective function aims at simultaneously
flexibility in the resource allocation process. optimizing the allocation of both types of resources.
Nevertheless, in situations where the physical resources are We assume that VN requests arrive according to a Poisson
scarce, it is desirable to reduce the number of physical links process, and that each VN has an associated lifetime measured
consumed to the minimum possible. in time units with an average of 1/μ = 1000, following an
Therefore, the objective function SDP, proposed in expres- exponential distribution. The same assumption was also taken
sion (18), aims to minimize the number of physical links by the authors of [13]. The average number of VN requests
consumed due to the VN embedding, while it prefers physical per time unit, i.e., value of λ, is started with 3 VN requests
links with more available bandwidth, and at the same time per 100 time units, and increases by 1 VN request, up to
chooses physical nodes with more available CPU power, 10 requests. This can give an insight into two opposite case
thereby supporting the load balancing aspect. The parameters scenarios, with a very high and very low acceptance ratio. For
α and β are used to weight the cost of each type of resource. each value of λ, 10 trials are performed. A new set of VN
(Note that the first term in eq. (18) would result in a constant, requests and a new physical network topology are generated
if Cip (t) was missing in the denominator.) for each trial. All simulations are set to run up to 50000 time
units to mitigate the transient phase effect [24] and to obtain
  ⎛ ⎞ the steady-state. A confidence interval of 95% is used for all
  xm   yij mn
results presented below.
minimize α p
i
+β ⎝ p
⎠ (18)
m i
Ci (t) mn ij
Bij (t)
2 These values were also considered by the authors of [9], [12], [13]

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
362 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 10, NO. 4, DECEMBER 2013

TABLE II 1
C OMPARED VN E MBEDDING M ETHODS .

Average Virtual Network Request Acceptance Ratio

G−SP
Notation Method Description 0.9 G−MCF
G-SP [9] Greedy Node Mapping with Shortest Path R−ViNE
D−ViNE
Based Link Mapping. 0.8 D−ViNE−SP
G-MCF [12] Greedy Node Mapping with Splittable Link
D−ViNE−LB
Mapping using MCF. VNE−NLF−LB+εSP
R-ViNE [13] Randomized Node Mapping with Splittable 0.7 VNE−NLF−SDP
Link Mapping using MCF. VNE−NLF−WSDP
D-ViNE [13] Deterministic Node Mapping with Splittable
Link Mapping using MCF. 0.6

D-ViNE-SP [13] Deterministic Node Mapping with Shortest

Path Based Link Mapping. 0.5
D-ViNE-LB [13] Deterministic Node Mapping with Split-
table Link Mapping using MCF, where
αuv = βuv = 1, ∀u, v, w ∈ N S . 0.4
VNE-NLF-LB+SP VN Embedding with node-link Formulation
using overall Load Balancing; in the case
of having more than one solution, it uses 3 4 5 6 7 8 9 10
Shortest Path, where  = 1.0 × 10−11 . Number of Virtual Network Requests per time unit
VNE-NLF-SDP VN Embedding with node-link Formulation
using overall Short Distance Path.
VNE-NLF-WSDP VN Embedding with node-link Formulation Fig. 3. Average VN acceptance ratio as a function of VN request rate.
using overall Weighted Short Distance Path.

Before comparing the different embedding methods and al-

The evaluated embedding methods are Greedy Node Map- gorithms, we should group them into four different categories
ping with Shortest Path based Link Mapping (G-SP) [9], according to the nature of the method itself, i.e. heuristic,
Greedy Node Mapping with Splittable Link Mapping using heuristic combined with mixed integer programming, and
Multi-Commodity Flow Constraint (G-MCF) [12], Random- linear programming:
ized Node Mapping with Splittable Link Mapping using Multi- i Heuristic - the VN embedding problem is solved using a
Commodity Flow Constraint (R-ViNE) [13], Deterministic simple algorithm; this method performs the VN embed-
Node Mapping with Splittable Link Mapping using Multi- ding very f ast and a possibly sub-optimal embedding
Commodity Flow Constraint (D-ViNE) [13], Deterministic solution is obtained. The method G-SP [9] fits into this
Node Mapping with Shortest Path based Link Mapping category;
(D-ViNE-SP) [13], Deterministic Node Mapping with Split- ii Heuristic combined with Mixed Integer Programming
table Link Mapping using Multi-Commodity Flow Constraint (MIP) - the VN embedding problem is solved in two steps:
and Load Balancing based (D-ViNE-LB) [13], and the pro- in the first step a mathematical algorithm is used to map
posed linear programming formulation, i.e. VNE-NLF, with 3 virtual nodes on physical nodes, and in the second step the
different cost functions which were described in the previous MIP is performed to embed the virtual links. The method
section. All these methods are briefly summarized in Table II. G-MCF [12] fits into this category.
The state of the art methods are simulated using an existing iii Heuristic combined with Mixed Integer Programming
implementation [25]; to solve the mixed integer programming (MIP) and a better coordination between mapping phases
on the methods G-MCF, R-ViNE, D-ViNE, D-ViNE-LB, and is added - the same principle is applied, as in the above cat-
D-ViNE-SP, the GLPK [26] solver version 4.20 is used. egory, to solve the VN embedding problem, although a bet-
All the simulations for the different embedding meth- ter coordination between the mapping phases is achieved
ods were performed using an Intel R
Xeon R
CPU using an augmented ”substrategraphconstruction” [13].
[email protected], and the time consumed per VN request The methods R-ViNE, D-ViNE, D-ViNE-SP, D-ViNE-
embedding was registered. LB [13] fit into this category;
The CPLEX R
[27] version 12.2 was used to solve the linear iv Integer Linear Programming (ILP) - the VN embedding
programming problem of the VNE-NLF; a time limit of 600 problem is solved using integer linear programming. This
seconds is defined for each VN mapping, although most of method obtains an optimal solution for a given cost
the VNs are embedded in hundreds of milliseconds; and the function combining resource consumption minimization
CPLEX R
was set to use only one CPU core for comparison with a load balancing strategy. The method VNE-NLF and
purposes with the remaining methods. The evaluation metrics its different objective functions fit into this category.
are the ones defined in section III-E. 1) VN Request Acceptance Ratio: One of the main aspects
of the performance of each embedding method is the VN
request acceptance ratio, which is shown in Fig. 3 and is
B. Impact of the Number of VN Requests given by formula (6). As can be observed, all methods show a
This subsection presents the evaluation results as a function linear behavior with the variation on the VN requests, where
of the VN request rate, for all the previously described metrics. the acceptance ratio decays linearly with the number of VN
To increase the readability of all figures, we have considered requests, and the slope is approximately the same for all
different x values for different strategies, e.g.: 3.4, 4.4, 5.4 for methods. This decay represents the fact that there are no
G-SP; 3.3, 4.3, 5.3 for G-MCF; 3.2, 4.2, 5.2 for R-ViNE. infinite physical resources.

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
MELO et al.: OPTIMAL VIRTUAL NETWORK EMBEDDING: NODE-LINK FORMULATION 363

0.5

Average VN Acceptance Ratio times Average Node Utilization

0.65

0.6 0.45

0.55
G−SP 0.4
0.5 G−MCF
Average Node Utilization

G−SP
R−ViNE 0.35 G−MCF
0.45 D−ViNE R−ViNE
D−ViNE−SP D−ViNE
0.4 0.3
D−ViNE−LB D−ViNE−SP
VNE−NLF−LB+εSP D−ViNE−LB
0.35 VNE−NLF−SDP 0.25
VNE−NLF−LB+εSP
VNE−NLF−WSDP VNE−NLF−SDP
0.3
0.2 VNE−NLF−WSDP
0.25
0.15
0.2

0.1
0.15

3 4 5 6 7 8 9 10 3 4 5 6 7 8 9 10
Number of Virtual Network Requests per time unit Number of Virtual Network Requests per time unit

Fig. 5. Average VN acceptance ratio times average node utilization as a

Fig. 4. Average node utilization as a function of VN request rate. function of VN request rate.

The method VNE-NLF, with its different objective func- 0.65

tions, achieves the highest performance, and it clearly out-
0.6
performs the other approaches. This is expected since integer
linear programming is applied to solve the VN embedding 0.55
Average Link Utilization

problem, and the optimal solution, according to the objective

function considered, is obtained per VN embedding. 0.5

The reason for these results, not only resides in the usage 0.45
of an integer linear programming approach, but also in the uti- G−SP
G−MCF
lization of the node-link formulation by the VNE-NLF, which 0.4
R−ViNE
considers the universe of all possible embedding solutions, D−ViNE
0.35
instead of a few solutions. If we take, for example, the first D−ViNE−SP
D−ViNE−LB
case with only 3 VN requests per 100 time units, the VNE- 0.3 VNE−NLF−LB+εSP
NLF is able to accept nearly all requests, while the remaining VNE−NLF−SDP
0.25 VNE−NLF−WSDP
methods are able to accept only 70% of the requests. The
3 4 5 6 7 8 9 10
embedding method that has the lowest acceptance ratio is Number of Virtual Network Requests per time unit
the D-ViNE-SP, and the method with the highest VN request
acceptance ratio is the VNE-NLF-WSDP. Fig. 6. Average link utilization as a function of VN request rate.
It is expected that the VNE-NLF method will perform
better in all cases. For instance, if the embedding problem
is feasible, i.e., possible solutions exist, the VNE-NLF will sources can be loaded and what kind of embedding methods
find out the optimal solution according to the cost function. tends to saturate them firstly.
Using a heuristic approach or even a combined approach, this The node utilization shows a dependency on the VN ac-
is not always the case: frequently only a feasible solution will ceptance ratio, as it can be perceived from Fig. 3 and Fig. 4.
be presented. To provide a better understanding on this issue, we plot the
2) Node Utilization: The average node utilization as a acceptance ratio metric times the node utilization, which is
function of the number of VN requests is depicted in Fig. shown in Fig. 5. We observe that the methods that make use of
4. With a small number of VN requests, i.e., 3 VN requests, heuristics, e.g. G-SP, or heuristics combined with MIP, e.g. G-
the node utilization does not go beyond 20% and 35% for the MCF and D-ViNE-LB, show the same behavior for all the VN
heuristic group (i.e. groups i, ii, and iii) and for the VNE- requests considered, i.e. the VN acceptance metric multiplied
NLF, respectively. The VNE-NLF group is consuming more by the node utilization metric is nearly constant. The same
resources of the physical nodes than the heuristics, which is does not apply to the VNE-NLF, since it increases per VN
expected according to the acceptance ratio. When the number request considered, until 6 VN requests per 100 time units,
of VN requests is increased, the node utilization also increases, and beyond the 6 VN requests per 100 time units, it shows the
since we are trying to accommodate more VNs on the infra- same behavior as its counterparts. This means that, although
structure, but with the same amount of available physical the VN request rate is increasing, the VNE-NLF approach
resources. is able to keep with this increase until the VN embedding
An efficient embedding method in situations of high VN problem moves from an optimization problem (i.e. there are
demand would be able to load the nodes to their full capacity. sufficient physical resources for the demand), to a feasibility
The important aspect to retain here is how much node re- problem (i.e. there are no sufficient resources for the demand).

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
364 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 10, NO. 4, DECEMBER 2013

Average VN Acceptance Ratio times Average Link Utilization

0.4 0.9
G−SP
G−MCF
R−ViNE

Average Embedding Factor

0.35 D−ViNE 0.8
D−ViNE−SP G−SP
D−ViNE−LB G−MCF
VNE−NLF−LB+εSP 0.7 R−ViNE
0.3 D−ViNE
VNE−NLF−SDP
VNE−NLF−WSDP D−ViNE−SP
D−ViNE−LB
0.6
0.25 VNE−NLF−LB+εSP
VNE−NLF−SDP
VNE−NLF−WSDP
0.5
0.2

0.4

3 4 5 6 7 8 9 10 3 4 5 6 7 8 9 10
Number of Virtual Network Requests per time unit Number of Virtual Network Requests per time unit

Fig. 7. Average VN request acceptance ratio times average link utilization Fig. 8. Average embedding factor as a function of VN request rate.
as a function of VN request rate.
G−SP
2 G−MCF
3) Link Utilization: The physical link utilization metric 1.8
R−ViNE
D−ViNE
is plotted in Fig. 6. Here we do not have the same regular Average VN Embedding Time (s) D−ViNE−SP
1.6
behavior according to the number of VN requests for all the D−ViNE−LB
VNE−NLF−LB+εSP
methods, as shown before for the node utilization. Also, there 1.4
VNE−NLF−SDP
is no consensus in terms of clearly identifying which group 1.2 VNE−NLF−WSDP
causes the highest utilization on the physical links due to the
1
embedding process. Nevertheless, we can clearly state that, on
average, either the G-MCF or D-ViNE-LB shows the highest 0.8

utilization on the links; in the extreme case scenario (i.e. 0.6

with 10 VN requests) they have an average link utilization
0.4
of 60% and 67%, respectively. With respect to the lowest link
utilization, we observe that the embedding methods R-ViNE, 0.2

D-ViNE, D-ViNE-SP, and VNE-NLF-WSDP are the ones that

3 4 5 6 7 8 9 10
tend to consume less bandwidth, reaching values of 50% of Number of VN Requests per 100 time units.
average link utilization, for the same considered situation.
Having in mind that one virtual link could be mapped in Fig. 9. VN solving time as a function of VN request rate.
several ways, it is reasonable to observe different behaviors
according to the strategy of the method. If the strategy is to
save bandwidth, i.e. SP, the embedding will consume the least physical nodes and links, choosing the solution that consumes
bandwidth possible per VN mapping; if the strategy is load the least bandwidth. Therefore, in the situation of only a few
balancing on the links, i.e. LB, it will tend to balance the VN requests, the method will tend to allocate more resources
utilization among all links in order to distribute the total load. than required due to the nature of the load balancing; with a
The the link utilization also shows a dependency on the VN higher VN request rate, this situation tends to disappear once
acceptance ratio, as can be observed in Fig. 3 and Fig. 6. To the available resources are scarcer. Therefore, the embedding
provide a better understanding, we plotted the acceptance ratio factor will increase with the number of VN requests. We can
metric times the link utilization in Fig. 7. In contrast to the also state that the efficiency of the heuristic group, in general,
node utilization, the dependency factor on the link utilization is very low, lower than 50%. With respect to the VNE-NLF
shows a more complex behavior: it still increases significantly group, it has a good efficiency, being in most of the cases
until reaching 6 VN requests for the case of the VNE-NLF higher than 85%; the efficiency of the VNE-NLF-WSDP is
group, although it starts to decrease after 7 VN requests, in a closer to 100% which means that, on average, this method
not so expressive way. For the other methods the dependency provisions the same amount of resources as requested per VN.
on the VN request rate is less pronounced. 5) VN Embedding Time: An important aspect of all the VN
4) Embedding Factor: Fig. 8 shows the embedding factor embedding methods is the time that they require to embed,
as a function of the VN request rate, where the weight on average, a VN request and how it varies with respect to
parameters, α, β, γ, and η of equation (7) are set to 1. The the different loads on the physical infrastructure, i.e. the VN
embedding factor slightly decreases with the number of VN request rate.
requests, except for the case of the VNE-NLF-LB+SP. The Figure 9 shows the solving time for each method as a
behavior obtained when using the VNE-NLF-LB+SP is as function of the number of VN requests per 100 time units.
expected, since it performs an overall load balancing of the Before analyzing the figure, one must consider five different

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
MELO et al.: OPTIMAL VIRTUAL NETWORK EMBEDDING: NODE-LINK FORMULATION 365

aspects: i) all methods have been simulated using the same 0.95
machine; ii) the time to embed a VN strongly depends on

Average Virtual Network Request Acceptance Ratio

the physical characteristics (e.g. CPU) of that machine; iii) 0.9

the time to embed a VN strongly depends on the nature of 0.85

the embedding method (i.e. a mathematical algorithm will
0.8
take just a few milliseconds, while linear programming is
expected to take hundreds of milliseconds); iv) two different 0.75
G−SP
G−MCF
linear programming tools (GLPK was used to solve the MIP D−ViNE−LB
0.7
of G-MCF, R-ViNE, D-ViNE,D-ViNE-SP, D-ViNE-LB; and VNE−NLF−WSDP

CPLEX R
to solve the ILP of the VNE-NLF); v) methods 0.65
R-ViNE, D-ViNE,D-ViNE-SP, and D-ViNE-LB perform two
0.6
linear programming operations, i.e. one for the mapping of
the virtual nodes, and another for the mapping of the links. 0.55
The fourth aspect, although important for the solving time,
0.5
will not interfere with the curve behavior, e.g. polynomial or 5 10 15 20
Maximum Distance Between Virtual Nodes
exponential, since the same method, i.e. branch and cut, is
applied by both solvers (i.e. GLPK and CPLEX) to solve the
VN embedding problem. Fig. 10. Average VN acceptance ratio as a function of the distance between
virtual nodes.
From the figure, we can observe two types of behaviors: the
method VNE-NLF with three different costs functions shows a
decaying behavior with the VN request rate; for the remaining
methods we observe a nearly constant behavior. 0.4

For the first behavior, i.e. method VNE-NLF with three

different cost functions, one should take into consideration
Average Node Utilization

0.35
that the VN request acceptance ratio is considerably higher,
e.g. 90% until 6 VN requests: more than one mapping solution G−SP
per VN request is expected to exist; therefore, the optimization 0.3 G−MCF
D−ViNE−LB
process takes place and will consume the majority of the VNE−NLF−WSDP
solving time to obtain the optimal solution. 0.25
For the remaining methods, the VN request acceptance ratio
is lower and below 70%: usually there is not more than
one mapping solution per VN request, on average, which 0.2

significantly reduces the solving time. This is the case of the

methods G-MCF, R-ViNE, D-ViNE, D-ViNE-SP and D-ViNE-
5 10 15 20
LB. Maximum Distance Between Virtual Nodes
We can also add that the methods R-ViNE, D-ViNE, D-
ViNE-SP, and D-ViNE-LB take twice the time on average to Fig. 11. Average node utilization as a function of the distance between
embed a VN compared to G-MCF. This is related with the virtual nodes.
number of MIP problems solved per VN embedding. The latter
only considers one MIP problem per VN embedding, while
the former ones consider two MIP problems. maximum allowed distance between virtual nodes represented
The method that performs the embedding in the shortest in expression (13), a new set of simulation experiments
time has the poorest performance (the G-SP), which solves was performed. The VN request arrival rate was fixed to
each VN request embedding problem in an average of 20 4 VN requests per 100 time units; the maximum distance
milliseconds. between virtual nodes was set to vary between 5 and 20
The method that requires the longest time to perform the within intervals of 2.5 distance units; for each considered
embedding for the case of 3 VN requests has the highest value of maximum distance, the same set of VN requests
performance (the VNE-NLF), using the WSDP cost function, was used. The remaining parameters i.e., virtual network size,
which takes less than 2 seconds on average for that case. link probability, and number of nodes were maintained. To
However, if we increase the load, the situation significantly increase the readability of all figures, only the best method of
changes, and the methods R-ViNE, D-ViNE and D-ViNE-LB each group is presented in this section: G-SP (heuristic), G-
take more time to obtain the embedding solution. On average, MCF (mixed integer programming - link-path), D-ViNE-LB
they take 1 second to embed a VN, while the VNE-NLF- (mixed integer programming with better node-link embedding
WSDP consumes less than 200ms. coordination - link-path ), and VNF-NLF-WSDP (integer
linear programming - node-link).
1) VN Request Acceptance Ratio: Fig. 10 depicts the VN
C. Impact of the Maximum Distance Between Virtual Nodes request acceptance ratio as a function of the distance between
To evaluate the impact on the overall performance of the virtual nodes. Two different behaviors can be observed:
different embedding methods due to the restriction on the i The acceptance ratio increases with the distance between

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
366 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 10, NO. 4, DECEMBER 2013

0.54 the link utilization with the maximum distance between

nodes. This is expected, considering the increase on the
0.52
VN request acceptance ratio with the distance.
0.5 4) Embedding Factor: The embedding factor is depicted in
Average Link Utilization

Fig. 13, where three distinct behaviors can be observed:

0.48
i The embedding factor slightly decreases with the maxi-
0.46 mum distance: this is the case of methods G-SP and G-
0.44
MCF, where these demonstrate to loose efficiency with the
considered distance.
0.42 ii The embedding factor does not vary with the maximum
0.4
distance: this is the case of method D-ViNE-LB.
G−SP
G−MCF iii The embedding factor increases with the maximum dis-
0.38 D−ViNE−LB tance: this is the case of the VNE-NLF group. This
VNE−NLF−WSDP
demonstrates that the VNE-NLF group is able to be more
5 10 15 20
Maximum Distance Between Virtual Nodes efficient with the relaxation on the distance constraint
expression (13).
Fig. 12. Average link utilization as a function of the distance between virtual
nodes.
VII. C ONCLUSION

0.8
This paper proposed the VNE-NLF to solve the VN em-
bedding problem. The model applies optimization theory to
0.75 simultaneously embed the virtual nodes and the virtual links.
0.7
Three new cost functions are proposed: the LB+SP which
Average Embedding Factor

aims to minimize the overall load on the network per VN

0.65 embedding; the SDP which aims to minimize the number
G−SP of physical links consumed, and at the same time it chooses
0.6
G−MCF
physical nodes with higher availability of resources; and the
D−ViNE−LB
0.55 VNE−NLF−WSDP WSDP which includes the demanded capacity by the VN in
the objective function.
0.5
Simulation experiments show how far the state of the art
0.45 heuristics are from an ILP based optimization method. The
difference between the performance of the heuristics and the
0.4
VNE-NLF approach is, at least, 30% for the VN request
5 10 15 20 acceptance ratio (see Fig. 3). The node utilization is also
Maximum Distance Between Virtual Nodes
higher when comparing with the existing heuristics, which is
Fig. 13. Average embedding factor as a function of the distance between
expected since we are accommodating more virtual nodes on
virtual nodes. the network. However, the link utilization is similar to the ones
of the heuristics, and in some cases (e.g. G-MCF, D-ViNE-LB)
it is lower, which reflects the good efficiency of the embedding
virtual nodes: this is the case of the VNE-NLF group. This when using the VNE-NLF approach. The embedding factor of
behavior is expected if we consider the cases where VN the VNE-NLF is very high (i.e. it is close to 1). The results
requests were initially not mapped due to the distance con- also show that the maximum allowed distance between virtual
straint; increasing the permitted distance between virtual nodes seems to affect differently the performance of each
nodes will in principle result in more accepted VNs. embedding method: for the case of R-ViNE and D-ViNE with
ii Increasing the distance between virtual nodes decreases its three variants, it does affect negatively the performance
the acceptance ratio: this is the case of the assignment of the VN embedding; for the case of G-SP and G-MCF, it
methods G-SP, G-MCF, D-ViNE-LB. does not seem to cause a direct impact on the embedding; and
2) Node Utilization: Fig. 11 depicts the average physical regarding the VNE-NLF approach, it does affect positively the
node utilization as a function of the distance between virtual VN embedding.
nodes. The behavior observed for each method can be com- The VNE-NLF with all its different cost functions proves
pared to that of the VN acceptance ratio. to be an efficient embedding method. Not only it provides
3) Link Utilization: The average physical link utilization better results, but it also performs the VN embedding faster
is shown in Fig. 12, and different behaviors are observed than the compared heuristics (with exception of G-SP) for
according to the embedding method: a large number of VN requests per time unit. From all the
i G-SP and G-MCF methods reduce slightly the link uti- proposed and simulated cost functions, the WSDP is the one
lization with the distance, and the method D-ViNE-LB that demonstrates the best overall performance.
maintains the link utilization, despite some fluctuations Future work will endorse VN embedding with multi-
may be observed. objective optimization support, and re-configuration of virtual
ii The group of VNE-NLF demonstrates to slightly increase networks. The definition of cost functions which take into

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
MELO et al.: OPTIMAL VIRTUAL NETWORK EMBEDDING: NODE-LINK FORMULATION 367

account energy parameters will be also addressed, through the [19] I. Houidi, W. Louati, and D. Zeghlache, “A distributed virtual network
study of the impact on energy saving, i.e. the shutdown of an mapping algorithm,” in Proc. 2008 IEEE ICC, pp. 5634–5640.
[20] G. Alkmim, D. Batista, and N. da Fonseca, “Mapping virtual net-
interface or even an equipment, by applying optimization to works onto substrate networks,” J. Internet Services and Applications,
solve the VN embedding problem. vol. 4, no. 1, pp. 1–15, 2013. Available: http://dx.doi.org/10.1186/
1869-0238-4-3
[21] S. Even, A. Itai, and A. Shamir, “On the complexity of time table
ACKNOWLEDGMENT and multi-commodity flow problems,” in Proc. 1975 Symposium on
The author was partially supported by the Portuguese Foun- Foundations of Computer Science, pp. 184–193. Available: http://dx.
doi.org/10.1109/SFCS.1975.21
dation for Science and Technology (FCT) with a scholarship [22] M. Pióro and D. Medhi, Routing, Flow, and Capacity Design in
No. SFRH/BDE/33751/2009. The research leading to these Communication and Computer Networks. Elsevier/Morgan Kaufmann,
results has also received funding from the European Union 2004.
[23] E. Zegura, K. Calvert, and S. Bhattacharjee, “How to model an inter-
Seventh Framework Programme (FP7) under grant agreement network,” in Proc. 1996 IEEE INFOCOM, vol. 2, pp. 594–602.
number 257448. We would like to thank the anonymous [24] R. Jain, The Art of Computer Systems Performance Analysis. John Wiley
reviewers for their insightful and constructive comments. & Sons, 1991, vol. 182. Available: http://www.cmg.org/proceedings/
1991/91INT145.pdf
[25] “ViNE-Yard,” Sep. 2012. Available: http://mosharaf.com/ViNE-Yard.tar.
R EFERENCES gz
[1] L. Peterson, T. Anderson, et al., “A blueprint for introducing disruptive [26] “GLPK (GNU Linear Programming Kit) ,” Sep. 2012. Available: http:
technology into the Internet,” SIGCOMM Comput. Commun. Rev., //www.gnu.org/software/glpk/
vol. 33, pp. 59–64, Jan. 2003. Available: http://doi.acm.org/10.1145/ [27] “IBM ILOG Optimization Products,” Sep. 2012. Available: www-01.
774763.774772 ibm.com/software/websphere/products/optimization
[2] T. Anderson, L. Peterson, et al., “Overcoming the Internet impasse
through virtualization,” Computer, vol. 38, pp. 34–41, April 2005. Ms.C. Márcio Melo is a Ph.D. student at the
Available: http://portal.acm.org/citation.cfm?id=1058219.1058273 University of Aveiro. He received his Ms.C. degree
[3] N. Feamster, L. Gao, and J. Rexford, “How to lease the Internet in in Electronics and Telecommunications Engineering
your spare time,” SIGCOMM Comput. Commun. Rev., vol. 37, no. 1, from the University of Aveiro in 2008. Since 2009 he
pp. 61–64, 2007. has been with PT Inovação in the division of Tech-
[4] Y. Zhu, R. Zhang-Shen, et al., “Cabernet: connectivity architecture for nological Coordination and Exploratory Innovation.
better network services,” in Proc. 2008 ACM CoNEXT Conference, pp. He has been involved in FP7 European projects,
64:1–64:6, ACM ID: 1544076. such as 4WARD – Future Internet and SAIL (Scal-
[5] J. Touch, Y. S. Wang, et al., “A virtual Internet architecture,” ISI able and Adaptive Internet soLutions), and also
Technical Report ISI-TR-2003-570, 2003. Eurescom research studies. His research interests
[6] J. Carapinha and J. Jiménez, “Network virtualization: a view from the are in the area of Next Generation Networks, more
bottom,” in Proc. 2009 ACM Workshop on Virtualized Infrastructure specifically Network Virtualization and Software Defined-Networking.
Systems and Architectures, pp. 73–80. Available: http://portal.acm.org/
citation.cfm?id=1592648.1592660 Dr. Susana Sargento received her Ph.D. in 2003 in
[7] M. Melo, S. Sargento, and J. Carapinha, “Network virtualisation from an Electrical Engineering. She joined the Department
operator perspective,” Proc 2009 Conf. Sobre Redes de Computadores of Computer Science of the University of Porto
- CRC. Available: http://www.av.it.pt/marciomelo/papers/crc09.pdf in September 2002, and is in the Universidade de
[8] N. Chowdhury and R. Boutaba, “Network virtualization: state of the Aveiro and the Instituto de Telecomunicaç ões since
art and research challenges,” IEEE Commun. Mag., vol. 47, no. 7, pp. February 2004. She was also a Guest Faculty of
20–26, July 2009. the Department of Electrical and Computer Engi-
[9] Y. Zhu and M. Ammar, “Algorithms for assigning substrate network neering from Carnegie Mellon University, USA, in
resources to virtual network components,” in Proc. 2006 IEEE INFO- 2008/2009, in the scope of the Carnegie Mellon
COM, pp. 1–12. Portugal Program.
[10] J. Lischka and H. Karl, “A virtual network mapping algorithm based She has been involved in several national and
on subgraph isomorphism detection,” in Proc. 2009 ACM Workshop European projects, taking leaderships of several activities in the projects, such
on Virtualized Infrastructure Systems and Architectures, pp. 81–88. as the QoS and ad-hoc networks integration activity in the FP6 IST-Daidalos
Available: http://doi.acm.org/10.1145/1592648.1592662 Project. She has been recently involved in several FP7 projects (4WARD,
[11] J. Lu and J. Turner, “Efficient mapping of virtual networks onto a shared Euro-NF, C-Cast, WIP, Daidalos, C-Mobile), national projects, and Carnegie
substrate,” Washington University in St. Louis, Tech. Rep., 2006. Mellon Portugal research projects (DRIVE-IN with the Carnegie Melon
[12] M. Yu, Y. Yi, et al., “Rethinking virtual network embedding: substrate University). She has been TPC-Chair and organizing several conferences, such
support for path splitting and migration,” SIGCOMM Comput. Commun. as MONAMI’11, NGI’09, IEEE ISCC’07, and will be organizing NTMS’12
Rev., vol. 38, no. 2, pp. 17–29, Mar. 2008. Available: http://doi.acm.org/ and IEEE FEDNET (with IEEE NOMS’12). Her main research interests are
10.1145/1355734.1355737 in the areas of Next Generation and Future Networks, more specifically QoS,
[13] N. Chowdhury, M. Rahman, and R. Boutaba, “Virtual network embed- mobility, self- and cognitive networks.
ding with coordinated node and link mapping,” in Proc. 2009 IEEE
INFOCOM, pp. 783–791.
[14] N. Farooq Butt, M. Chowdhury, and R. Boutaba, “Topology-awareness Dr. Ulrich Killat studied physics in Hamburg and
and reoptimization mechanism for virtual network embedding,” in Proc. Heidelberg, Germany. He was then with the research
2010 IFIP TC 6 International Conference on Networking, pp. 27–39. laboratories of Philips in Hamburg and Aachen.
Available: http://dx.doi.org/10.1007/978-3-642-12963-6 3 Since 1991, he has been a professor and head of
[15] J. Nogueira, M. Melo, et al., “Virtual network mapping into het- Department of Communication Networks, Hamburg
erogeneous substrate networks,” in Proc. 2011 IEEE Symposium on University of Technology (TUHH). His research
Computers and Communications, pp. 438–444. interests include traffic theory, network planning,
[16] J. Botero, X. Hesselbach, et al., “Optimal mapping of virtual networks and modelling and simulation of communication
with hidden hops,” Telecommun. Systems, pp. 1–10, 2011. networks. He is a professor emeritus since 2009.
[17] M. Chowdhury, M. Rahman, and R. Boutaba, “Vineyard: virtual net-
work embedding algorithms with coordinated node and link mapping,”
IEEE/ACM Trans. Networking, vol. 20, no. 1, pp. 206–219, Feb. 2012.
[18] M. Melo, J. Carapinha, et al., “Virtual network mapping—an opti-
mization problem,” in Mobile Networks and Management, ser. Lecture
Notes of the Institute for Computer Sciences, Social Informatics and
Telecommunications Engineering, K. Pentikousis, R. Aguiar et al., Eds.
Springer, 2012, vol. 97, pp. 187–200.

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.
368 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 10, NO. 4, DECEMBER 2013

Dr. Andreas Timm-Giel was group leader from Ms.C. Jorge Carapinha is a Senior Research Engi-
1994–1999 at the University of Bremen in the area neer at PT Inovação. He graduated in Electrical and
of mobile and satellite communications and involved Computer Engineering at the University of Coimbra
in several EU funded projects. After receiving his in 1984 and received an MSc in Electronics and
PhD he moved to MediaMobil GmbH and M2SAT Telecommunications from the University of Aveiro
Ltd. as Technical Project Leader and Manager Net- in 1998. Since 1985 he has been with PT Inovação
work Operations. In December 2002 he joined the (formerly CET). He has worked in several fields
Communication Networks group at the University related to operator backbone networks and tech-
of Bremen as senior researcher and lecturer. He was nologies, including MPLS Virtual Private Networks
leading several industrial, national and EC funded and Quality of Service. He has a long record of
research projects and from 2006 he was additionally participation in international collaborative research
directing the interdisciplinary activity “Adaptive Communications” of TZI projects. Presently his activity is mainly focused on Network Virtualization
(Center of Computing and Communication Technologies). In November and Cloud Networking. He has authored/co-authored several papers on
2009 he was appointed full professor at Hamburg University of Technology these topics, published in technical journals or presented in international
(TUHH) and is heading the Institute of Communication Networks. Since conferences, as well as Book Chapters.
2012 he is coordinator of the research cluster SOMSED “Self-organizing
mobile sensor and data networks” at TUHH and since 2013 deputy head
of the TUHH’s school of electrical engineering, computer science and
mathematics. His research interests are mobile and wireless communications,
sensor networks and the Future Internet.

Authorized licensed use limited to: Politecnico di Milano. Downloaded on August 24,2020 at 08:32:37 UTC from IEEE Xplore. Restrictions apply.