Efficient Virtual Network Embedding with

Node Ranking and Intelligent Link Mapping

Khoa TD Nguyen† , Qiao Lu† , and Changcheng Huang†
† Department of Systems and Computer Engineering

Carleton University, Ottawa, ON K1S 5B6, Canada

{khoatnguyen, qiaolu, huang}@sce.carleton.ca
Abstract—Network virtualization (NV) is emerged as a key [4]–[6] for node mapping stage, or exploited simplified global network
enabler for the success of the future virtualized networks (e.g. 5G resources for VNoM problems [7], [8]. NTANRC in [9] adopted an
networks and smart Internet of Things (IoT)). Virtual Network efficient node-ranking approach based on five topology attributes
Embedding (VNE) that addresses the embedding problems of (e.g. node degree, node strength, node distance, farness/closeness and
heterogeneous virtual networks (VNs) onto a physical infrastruc- link interference) and global network resources (e.g. node location,
ture is a main challenge in NV. Network topology attributes node capacity, link bandwidth and link propagation delay) to rank all
and network resource-considered (NTANRC) algorithm is a substrate and virtual nodes before embedding each VNR. This node-
virtual node mapping mechanism that considers essential network ranking approach coordinated node and link constraints to provide a
features and global network resources for ranking both substrate better performance of embedding VNRs.
and virtual nodes prior to embedding each given virtual network In fact, a vast number of research papers including [3]–[11] merely
request (VNR). In this paper, we propose NTANRC combined focused on approaching an efficient node mapping, and completely put
with a distributed parallel Genetic Algorithm (GA) for virtual trust in k-shortest path or multicommodity flow (MCF) algorithms for
link mapping, namely NTANRC-GA, to solve online VNE prob- virtual link mapping. Unlike previous papers, [12] primarily centralized
lem. Extensive evaluation results show that our proposed solution VLiM stage utilizing a novel distributed parallel GA-based algorithm,
not only achieves better performance compared to state-of-the-art namely IDPA, with a very simple node mapping based on a greedy
VNE algorithms, but also challenges the rapid speed of shortest method. Through its evaluation, this paper proved that VLiM also
path (SP) method. NTANRC algorithm and the parallel GA-based plays a crucial role in attaining an efficient VNE algorithm along with
algorithm are reverse compliments of each other to achieve an the inevitably important role of VNiM. This intelligent VNE algorithm
efficient VNE solution. outperformed its competitors not only in performance, but also in
Index Terms—Network Virtualization, Virtual Network Em- speed due to a complicated link mapping algorithm implemented in
bedding, 5G-and-beyond networks, smart IoT, Node Ranking, a proper embedding scheme.
Genetic Algorithm. In this paper, we propose the NTANRC-GA algorithm, a dramatic
I. I NTRODUCTION combination between an efficient node-ranking approach and an
intelligent GA-based algorithm in aimed at obtaining an effective VNE
NV is recognized as a de facto paradigm to make provision for algorithm. Our proposed algorithm, that exploits a set of distributed
anticipated success of the future networks such as 5G [1], virtualised parallel machines, enables to embed multiple link mapping requests
IoT networks [2]. NV notably allows to share substrate network at the same time so as to reduce the execution time. We mainly aim
(SN) resources among multiple VNRs, that enables an isolated to increase profit of an InP by maximizing the VNR acceptance ratios
coexistence of several VNs on a single physical network. Virtualization while minimizing the embedding costs. To the best of our knowledge,
technology not only improves network resource utilization of the SN this is the first paper that efficiently deploys a node-ranking approach
efficiently, but also facilitates the deployment and the evaluation incorporated with a complicated link mapping solution to solve VNE
of new architecture designs or network protocols. Subsequently, problems. The results show that our proposed VNE solution improves
NV prevents network infrastructure from an avoidable expansion. the acceptance ratios of VNRs, revenue to cost ratios and link resource
In VN environment, a service provider (SP) typically converts a utilization compared to NTANRC [9], IDPA [12] and three VINE
requested service/application into a VN, and then conveys to an algorithms [3].
infrastructure provider (InP) under a VNR. Thereupon, InP will The remainder of this paper is organized as follows: the network
embed the underlying VN onto its physical infrastructure through an model is formulated in Section II. NTANRC approach for VNE node
optimization process with multiple constraints. Towards InPs, they mapping and intelligent parallel GA-based algorithm are described in
advocate an efficient resource allocation mechanism that increases Section III and Section III-B, respectively. Performance evaluation is
their revenue by serving as many VNRs as possible while keeping the introduced in Section IV while related work is presented in Section V.
embedding costs minimized. A VNR is generally associated with a set Finally, Section VI is a conclusion of this paper.
of virtual nodes connected via virtual links to form a specific topology,
dynamically arriving and residing in the network for a random duration II. N ETWORK M ODEL AND P ROBLEM D ESCRIPTIONS
in most scenarios. Embedding VNRs onto the underlying SN with
manifold topology and rigid resource requirements is known as a A. Virtual Network Assignment
VNE problem. SN is modelled as a weighted undirected graph Gs = (N s , Ls ),
VNE process can be divided into two sub-problems: Virtual Node where N s is the set of substrate nodes and Ls is the set of substrate
Mapping (VNoM) and Virtual Link Mapping (VLiM). VNE has been links. A substrate node ns ∈ N s with a geographical location
proven to be N P-Hard either for VNoM or VLiM [3]. The common loc(ns ) has the available CPU capacity c (ns ), whereas each physical
formulated optimization models (e.g. Integer Linear Programming link ls ∈ Ls between any two physical nodes possesses a b (ls )
(ILP)) are usually recommended to gain optimal VNE solutions, but bandwidth capacity. For simplification, memory and storage resources
they face several issues of scalability, complex implementation, and are excluded in this paper. Let model the ith arriving VNR as a
high time consumption. These solutions cannot be indeed tailored for weighted undirected graph denoted as Gvi = (Niv , Lvi ), in which Niv
online VNE problems. Indeed, most of research papers have adopted is the set of virtual nodes while Lvi is the set of virtual links. Each
light-weighted heuristic algorithms to deal with the aforementioned virtual node nvi ∈ Niv has a requested CPU capacity c(nvi ), whereas
obstacles of the optimization models. Several heuristic algorithms a virtual edge liv (svi , dvi ) ∈ Lvi between a virtual source node svi
[3]–[8] have been proposed over the past decade, embedding VNRs and a virtual destination node dvi possesses a required bandwidth
in separate mapping stages. These algorithms relaxed the integer capacity b (liv ). Each VNR normally prefers a mapping radius D(nvi )
constraints to achieve feasible VNE in polynomial time [3] or revealing how far virtual node nvi can be allocated from loc(nvi ).
considered a single topology attribute and local network resources Mapping the ith VNR Gvi onto the SN Gs can be decomposed into

978-1-7281-9486-8/20/$31.00 ©2020 IEEE

two main components as determined above: VNoM and VLiM. In Higher remaining bandwidth leads to greater opportunity for accepting
node mapping stage, a virtual node of a VNR can be embedded onto the upcoming VNs.
a substrate node AN : Niv → N s , with nv ∈ Niv subject to: Fitness Function (FF): determine which link solutions will reproduce
and remain in next generation, relevant to the predefined objectives
c(nvi ) ≤ RN (AN (nvi )) (1) to be optimized in our proposed GA-based algorithm in Section III-B.
D(loc(nvi ), loc(AN (nvi ))) ≤ D(nvi ) (2) Consequently, fitness function is used to examine the quality of each
VLiM solution among several feasible ones. Its fitness values are able
AN (nvi )
∈N s
(3) to present a scientific proof of selecting the corresponding solutions in
X GA operators. Similar to [12], we take hop-count and bandwidth into
RN (ns ) = c(ns ) − c(nvi ) (4) consideration in fitness function. Less hop-count solution is preferable
nv →ns since it is substantially associated with less bandwidth consumption,
where nv → ns defines the virtual node nv embedded onto physical leaving more residual network bandwidth that increases the possibility
node ns , and the distance between is and j d is expressed by D(is , j d ) of the upcoming VNRs being accepted.
whereas RN (ns ) denotes the remaining CPU capacity of a substrate Fitness function F(Sz ) is eventually calculated as below:
node. Indeed, each virtual link can be commonly mapped onto the X b(liv ) ∗ wα 1
corresponding physical path including one or more substrate links. F(S) = ( v v
+ ∗ wβ ) (12)
v v b(li ) + RL (AL (li )) hAL (lvi )
The unsplittable link mapping can be expressed as AL : Lvi → Ls li ∈Li

while liv = (svi , dvi ) ∈ Lvi , E s (AL (liv )) denotes a set of all available where, S and h are a feasible solution and hop-count of the link
physical paths from the source AN (svi ) to destination node AL (dvi ). mapping solution of liv respectively. wα and wβ are weight parameters
AL (svi , dvi ) ⊆ E s (AN (svi ), AN (dvi )) (5) equivalent to bandwidth and hop-count factors.

subject to: III. NTANRC-GA A LGORITHM

RL (es ) ≥ b(liv ), ∀es ∈ E s (AL (liv )) (6)
NTANRC-GA implements NTANRC algorithm [9] for addressing
RL (es ) = min
s s
RL (ls ) (7) VNVoM while deploying the distributed parallel GA algorithm [12] for
l ∈e
X link mapping. It means that the output of NTANRC algorithm becomes
s s
RL (l ) = b(l ) − b(liv ) (8) the input for VLiM. NTANRC algorithm efficiently seeks the potential
lv
i →l
s embedding nodes for VNRs. Besides, the novel GA-based algorithm
that is based on a distributed parallel scheme efficiently reduces the
where RL (es ) is the available bandwidth of a substrate path es ∈ E s , embedding cost, execution time, and improves link utilization.
and RL (ls ) is the residual substrate link capacity.
B. Performance metrics A. NTANRC Algorithm for Node Mapping
The major objective of VNE algorithm is to maximize the InP’s NTANRC collects five network topology attributes for its node-
revenues while minimizing the embedding cost. Thus, the ratio ranking adoption including node degree, node strength, distance,
between revenue over cost should be considered to estimate the farness and closeness and finally link interference. As shown in
performance of VNE algorithms, adhering to the acceptance ratio. [9], there are two sub-approaches (S: stable and D: direct), but we
For instance, high acceptance ratio while the average revenue to cost prefer S-approach due to its better performance.
ratio is low is unfavourable since this result reveals that the substrate Node degree: Let T˜d (.) denote the function that counts the total
resources are underutilized. In this article, the InP revenue is computed number of adjacent links of node ns , degree of a substrate node is
as the sum of total virtual resources mapped on the SN over time determined as follows:
while the cost of the ith VNE C(Gvi ) is the sum of total network
resources allocated to the ith VN. Subsequently, the revenue to cost Γ(ns ) = T˜d (ns ) (13)
ratio of ith VNR Gvi can be expressed as below:
Node strength is defined by the function T˜s (.) that counts the
sum of all adjacent link bandwidth of the substrate node ns .
X X
wb ∗ b(liv ) + wn ∗ c(nvi )
v
R(Gi ) v
li ∈Liv v v
ni ∈Ni
Θ(ns ) = T˜s (ns ) (14)
Υ(Gvi ) = = X X lv (9)
C(Gvi )
X v
c(ni ) + f si l s
Farness and Closeness of a node: Farness of a node n is the
nv v
i ∈Ni lv v s
i ∈Li l ∈L
s
sum of its shortest distances to all the other possible nodes whereas
where R and C are the generated revenue and network cost Closeness is reciprocal of the Farness as defined in (15) and (16).
respectively. b(liv ) and c(nvi ) are the requested bandwidth of the X
virtual linkv liv and the requested CPU of the virtual node nvi while wb F (ns ) = D(ns , ms ) (15)
l ms ∈N s
and wn flsi defines the bandwidth of substrate link ls that is allocated
to the virtual link liv . 1
Acceptance ratio: is expressed as the ratio between the number of Z(ns ) = (16)
F (ns )
accepted VNRs over the number of arrived VNRs during the interval s s
time τ is calculated as following: Link Interference: Interference of a physical link l(n s ,ms ) ∈ L
between node ns and node ms describes its contribution to the network
ξ a (τ ) connectivity, relying on the idea that traffic routed in a path can be
Aτc = (10)
ξ(τ ) minimized traffic interference in the future.
where ξ a (τ ) and ξ(τ ) is the number of the successfully mapped s F (ns ) F (ms )
VNRs and the number of VNRs respectively. I(l(n s ,ms ) ) =
s
+ (17)
Γ(n ) Γ(ms )
Remaining bandwidth: the residual bandwidth of a SN can be
calculated as following: In node-ranking approach, topology attributes and the global network
X X resources are simultaneously quantified in which the node capacity,
Rm (Ls ) = (b(ls ) − b(liv )) (11) node location, link bandwidth and link propagation delay are consid-
ls ∈Ls lv
i →l
s ered as global resource parameters. Accordingly, interaction between
two nodes ns and ms denoted as Φ(ns ,ms ) is defined as follows:
When new VNRs arrived, InP calculates the remaining network
resources, and then attempts to map such VNRs onto the SN B(ns )B(ms )
depending on the residual network resource information obtained. Φ(ns ,ms ) = I (18)
D(ns , ms )2 dP (ns , ms )2
 X
Φ(ns ) = Φ(ns ,ms ) (19) initialization, selection, crossover and mutation. An intelligent GA-
ns 6=ms ,ms ∈N s based orchestration algorithm for virtual link mapping stage not
only approached an efficient VNE solution, but also accelerated
where I is a constant, and Eq. 19 expresses the interaction of a node the embedding speed due to a distributed parallel paradigm [12].
ns with remaining nodes in the entire substrate network. Moreover, A proposed parallel GA scheme is presented in Fig 1. Each working
B denotes the resource block of a substrate node that can be defined machine independently runs GA algorithm with a pre-defined number
as follows: of iterations to explore feasible solutions for virtual link mapping.
X
B(ns ) = c(ns )Θ(ns ) s
I(l(ns ,ms ) ) (20) The best-matching VLiM outcome is selected among these parallel
ms 6=ns ,ms ∈N s
machines. All virtual link requests (VLRs) within a VNR are embeded
at once. Accordingly, a chromosome Cf that consists of several genes
Normalized B(ns ) and Φ(ns ) towards substrate node ns are gij denotes a feasible link mapping solution for all VLRs. Each gene
formulated as below: gij is associated with a substrate path in which i and j indicate the
B(ns ) current chromosome and virtual link respectively.
B̄(ns ) = qP (21) 1) Initial path pool generation To serve link embedding proce-
ns ∈Gs B(ns )2 dures, a path database for mapping VLRs is deliberately created since
we know that our SN is static. During its path generation, a k-shortest
Φ(ns ) path algorithm e.g. Dijkstra’s algorithm is deployed to find all possible
Φ̄(ns ) = qP (22)
Φ(ns ) 2 k-shortest paths for each pair of source-destination.
ns ∈Gs
2) Working node:
By calculating all percentage values Φ̄(ns ) of the substrate net- Population Initialization: Each working machine normally begins
work Gs , an initial node-ranking vector T0 has been made, where with a population initialization step. There are M chromosomes and
T0 = (Φ̄(n0 ) , Φ̄(n1 ) · · · Φ̄(n|N | ) )T . For each node ns ∈ N s , its node- each has N genes. An initial population P (M xN size) at kth working
ranking value can be calculated as follows: machine is represented as follows:
g1 · · · g1j · · · g1N
X    1 
νns = (1 − ρ)B̄(ns ) + ρ Φ(ns ) νms (23) C1
n6=m,m∈N (ns )
 C2   g21 · · · g2j · · · g2N 
  
 ..   .. .. .. 
  .. ..
where ρ is the damping factor and N (ns ) is the set of all nodes that  .   . . . . . 

P= = 1 j N (25)
have a path with the node ns in the substrate network. As a result, a  Cf   gf · · · g · · · gf 

  f
vector V formed by the node-ranking values of all substrate nodes  .   . .. .. .. .. 

 ..    .. . .
can be shown as below: . . 
CM g1 · · · gj · · · gN
M M M
Vns = (1 − ρ)B̄(ns ) + ρMVns (24)
Each gene that is associated with a potential link mapping solution is
where Vns = (νn0 , νn1 , · · · , νn|N | )T , and B̄(ns ) = randomly selected from the initial path pool. Such solution must pass
(b̄(n0 ), b̄(n1 ), · · · , b̄(n|N | ))T . M is the transition matrix (|N | × |N |) a feasibility check to be a potential embedding solution. N genes
where each element is calculated based on (21) and (22). Details on already passed the feasibility check compose a chromosome that is
a stable node-ranking approach are described in Algorithm 1 that acknowledged as a feasible solution for the corresponding VLiM.
is utilized to calculate the node-ranking values. NTANRC requires Selection: A fitness-based proportionate selection scheme which is
that all nodes including substrate and virtual nodes should be ranked based on a cumulative sum of the fitness relative weights (12) is
and sorted following a decreasing order of the node-ranking values. utilized to select parents from the initial population.
Similar to substrate nodes, the node ranking values for virtual nodes
  g 1 · · · g j c g j c +1 · · · g N 
are computed in the same procedures. When a VNR arrives, virtual

C1 1 1 1 1
nodes are calculated node-ranking values and then the virtual node ..   .. .. .. .. .. 
  . . . . .
  
with highest node-ranking value is first mapped to the substrate node  .   1 c c

 Cs   gs · · · gs j j +1 N
· · · gs 

that achieved highest node-ranking value, meeting node location and gs 
  .. .. ..
  
capacity constraints. In a VNR, virtual nodes cannot share the same  .
.. . . . .

   . . . . .

substrate node. The mapping is repeated until all virtual nodes are P= C
  
=  g 1 · · · g j c g j c +1 · · · g N 

successfully mapped to corresponding substrate nodes. The output of r

   r r r r 
..   . .. ..

NTANRC algorithm as node mapping results is then utilized as the

  . .. .. 

.   . . . . . 
input to the parallel GA-based algorithm [12] to deal with virtual 
 C   1 j c c
j +1 N 

link mapping.  M   gM · · · gM gM · · · gM 
 CM +1   1 c c
gs · · · gsj grj +1 · · · grN 
CM +2 1 jc j c +1
Algorithm 1 Stable Node-Ranking Approach gr · · · gr gs · · · gsN
(26)
Input: a network G = (N,L), δ: small positive number Crossover: This procedure combines parental chromosomes to
Output: a ranking vector V corresponding to a given G generate a new offspring for next generation. Cs and Cr denote two
Calculate matrix M and initial vector V0 (T0 ) parental chromosomes with their indexes s and r in initial population,
Define iteration number k and variable w. whereas new descendant chromosomes are described as C(M +1) and
while w ≥ δ do C(M +2) respectively. j c indicates a random crossover point between
Vk+1 = (1- ρ)B̄(ns ) + ρ M Vk ; any genes within N length, which is randomly chosen. Offspring is
w = kVk+1 − Vk k; established by swapping parental genes starting from the crossover
k = k +1; point j c + 1 to the end as illustrated in 26.
end while Mutation: This operator implements a random change to an individual
Vk = Vk+1 parent to produce a new offspring. A mutation point denoted as
j m is randomly generated. At this point, a new gene chosen from
B. Parallel GA-based Algorithm for Link Mapping the path database replaces an existing one within the in-processed
chromosome to form a new child. j m is a random mutation point, and
Distributed and Parallel computing has considered as an efficient jm
gr0 denotes a new gene that replaced the existing one in C(M +1) . The
mechanism to address large and complex problems with lower costs new mutation solution C 0 (M +1) after replacement can be represented
and less time consumption by concurrency support. GA is an appealing m
as C 0 (M +1) = [gs1 · · · grj 0 · · · gsN ].
AI approach for solving constrained or unconstrained optimization
3) Solution Sorting and Terminations Sorting procedure selects
problems. A conventional GA usually includes four main operators:
 Start
based upon average values with 95% confidence interval. The error
Node Mapping bars were very small due to a large number of samples used, which
Original Path Pool Generation
proved that our simulation results were obviously reliable. For better
1
presentation, we plotted figures with different colors and markers.
0 P

Initialization Initialization Initialization

B. Evaluation Results
Selection Selection Selection Simulation results are preferably shown in Figure 2 and 3.
NTANRC-GA achieved highest acceptance ratio by accepting more
no no no
Crossover Crossover Crossover
VNRs with significantly less costs than all rivals, which resulted
yes yes yes in much higher average revenue to cost ratios as depicted in Fig
......
Mutation
no
Mutation
no
Mutation
no 2. Achieving higher both acceptance and revenue to cost ratios
no yes no
yes no yes simultaneously is desirable for any VNE algorithm. Our proposed
Sorting Sorting Sorting solution proved that it could be feasible. For example, NTANRC-GA
yes yes yes
no
improved the acceptance ratios of IDPA as well as NTANRC, and
no no Termination
Termination Termination
was better than R-ViNE (the best performance algorithm in [3]) up
yes
yes yes
to 2.24%, 8.76% and 9.73% at 80 Erlang (Fig. 2a) respectively. Our
Synchronization proposed VNE solution gained 9.07%, 14.08% and 16.60% better
revenue to cost ratios than those of aforementioned algorithms at
no no no
Termination the same traffic load as illustrated in Fig 2b. In this paper, average
yes remaining bandwidth of all compared algorithms was selected and
Allocation
illustrated in Fig. 3a. With the intelligent link mapping, our VNE
Finish solution consumed less bandwidth to embed given link mapping
requests due to the efficient FF as shown in Fig. 3a. Moreover, we
Fig. 1: Parallel operation scheme compared the execution time for embedding a VNR among algorithms
where those of NTANRC-GA and IDPA algorithms are almost similar
the best embedding solution among the feasible ones based on their because they use the same link mapping mechanism. As depicted in
fitness values, and then it is conveyed to synchronization step for Fig .3b, NTANRC-GA is 36.26% and 40.11% absolute faster than
a global ranking. A parallel operation generally adopts a series of NTANRC and G-SP algorithms, respectively. The reason is that both
concurrent processes, and each accomplishes its particular job in NTANRC-GA and IDPA deploy the distributed parallel GA-based
different time. However, waiting for the last process finishing its algorithm for link mapping stage instead of the SP method used in
assigned task is frustrating, and this attempt might not guarantee a NTANRC and G-SP, which was proved to be considerably faster than
desired outcome. Thus, the master node will terminate GA algorithms SP in [12]. The remarkable performance came from the effective node-
at worker nodes if the current best VLiM solution has not been ranking method, that simultaneously took network topology attributes
consecutively improved for t times, where t denotes a termination and global network resources into account, and the intelligent GA-
parameter. based approach exploring the search space efficiently to achieve more
4) Synchronization and VNR allocation feasible link mapping solutions with lower time consumption due to
This operation is deliberately to determine the final VNE solution a proper parallel operation paradigm.
for the corresponding VLiM request by a global ranking process. This V. R ELATED W ORK
final selection is relied on highest achieved FF values compared with
those of all feasible embedding solutions conveyed from the working Consolidating research endeavors in NV, the authors in [14]
nodes. As a result, the VNR will be accepted and then allocated onto contributes a comprehensive survey to this research field. VNE
SN based on the information of the virtual node and link mapping problem is N P-hard in nature, that is intractable to be solved with
solutions found. The residual network resources will be consequently Integer Programming (LP). Thus, most VNE papers are concentrated
updated in advance. on seeking for efficient heuristic algorithms due to the computational
IV. P ERFORMANCE E VALUATION complexity of extract methods. [3] introduced a coordinated node and
link approach for virtual node embedding by relaxing the intractable
We compare our proposed NTANRC-GA algorithm with several integer constraints, and then using rounding techniques to choose
competitors including IDPA [12], NTANRC-S [9], D-ViNE, R-ViNE, unique node mapping. Huang et al. in [15] mainly extended [3] with
and G-SP [3]. These algorithms are evaluated in various performance a novel node splitting scheme and node collocation. The authors
metrics including acceptance ratio, average revenue to cost ratio, in [4] proposed a topology-aware node embedding deploying the
remaining bandwidth and execution time. Markov Random Walk model to quantify node capacity and its joint
link bandwidth. The paper [5] explored seven different topological
A. Simulation setup attributes for better coordinate node and link mapping and proposed
We deployed a discrete-event simulator to assess our proposed VNE various node-ranking approaches. Zhang et al. [6] took the node degree
solution with parameters similar to those in [3]. Hence, a popular and clustering coefficient information to enhance the metric of node
GT-ITM topology generator [13] has been utilized to generate SN and importance which was adopted to rank the substrate nodes, aiming
VNs. SNs are configured with average 50 nodes, that are randomly to determine the nodes with the highest potential for embedding
placed on a 25 × 25 Cartesian plane. These are randomly connected VNRs. Gong et al. [7], [8] proposed node-ranking approaches that
to average 140 edges adopting Waxman model with α = 0.5 and only considered the simplified global network resources for the node
β = 0.2, where α indicates the maximal edge probability and β mapping stage without essential topology attributes. Inspired from
determines the edge length. CPU and bandwidth capacity of SNs Google PageRank algorithm, [9] proposed an efficient node-ranking
are uniformly generated between 50 and 100 units, whereas VNRs algorithm to rank both virtual and substrate nodes prior to embedding
dynamically arrive following the Poisson process with an average each VNR, which is based on five topology attributes and global
rate λ varying from 4 to 8 VNs per 100 time units. Lifetime of network resources. Nguyen et al. [12] proposed a distributed parallel
VNRs follows an exponential distribution with an average value of GA-based algorithm for the link mapping phase, which confirmed
µ = 1000 time units. Hence, the load of VNRs can be quantified the critical role of VLiM for approaching an efficient VNE solution.
by µλ Erlangs. Additionally, the number of virtual nodes in each Recently, reinforcement learning algorithms have been proposed to
VNR is uniformly distributed between 2 and 10. CPU capacity and solve VNE problems and optimize their performance [10], [11].
bandwidth requirements of VNRs are uniformly distributed between 0 VI. C ONCLUSION
to 20 and 0 to 50 respectively. Similar to [14], we set wb = wn = 1 NV is indisputably a major factor of the anticipated success of
in this paper. Each simulation run for 50, 000 time units, 50 times future network architectures (e.g. 5G, virtualized IoT networks) so
longer than the average lifetime of a VN to exceptionally generate a that an efficient VNE algorithm is highly desirable. In this paper,
large number of independent samples. All performance figures were we propose the node-ranking approach that is based on network
 NTANRC-GA 0.675 NTANRC-GA
0.75 IDPA
NTANRC IDPA
0.650 NTANRC
G-SP
R-ViNE G-SP
0.70 0.625
D-ViNE R-ViNE
Average acceptance ratio

D-ViNE
0.600
0.65

R/C ratio
0.575
0.60 0.550

0.525
0.55
0.500
0.50
0.475
40 50 60 70 80 40 45 50 55 60 65 70 75 80
Traffic load (Erlang) Traffic load (Erlang)

(a) (b)
Fig. 2: (a) VNR Acceptance Ratio (b) Average revenue to cost ratio

0.600
NTANRC-GA
0.575 IDPA 250
NTANRC
G-SP
0.550
Average remaining bandwidth

R-ViNE
Average execution time (ms)
200
D-ViNE
0.525
150
0.500

0.475 100

0.450
50
0.425

0.400 0
40 45 50 55 60 65 70 75 80 NTANRC-GA IDPA NTANRC G-SP RViNE DViNE
Traffic load (Erlang)

(a) (b)
Fig. 3: (a) Average remaining bandwidth (b) Average CPU execution time
topology attributes and global network resources for virtual node [7] Long Gong, Yonggang Wen, Zuqing Zhu, and T. Lee, “Revenue-driven
mapping stage and the intelligent parallel GA-based algorithm for virtual network embedding based on global resource information,” in
link mapping stage to solve online VNE problem. Our proposed 2013 IEEE Global Communications Conference (GLOBECOM), 2013,
NTANRC-GA algorithm outperforms state-of-the-art algorithms in all pp. 2294–2299.
matrices towards performance and operation time. We can conclude [8] L. Gong, H. Jiang, Y. Wang, and Z. Zhu, “Novel location-constrained
virtual network embedding lc-vne algorithms towards integrated node
that NTANRC and intelligent GA-based algorithms are reverse and link mapping,” IEEE/ACM Transactions on Networking, vol. 24,
compliments of each other to achieve an efficient VNE solution. no. 6, pp. 3648–3661, 2016.
In future work, we will investigate simultaneously embedding nodes [9] H. Cao, L. Yang, and H. Zhu, “Novel node-ranking approach and multiple
and links in one-stage mapping utilizing Genetic Algorithm. topology attributes-based embedding algorithm for single-domain virtual
R EFERENCES network embedding,” IEEE Internet of Things Journal, vol. 5, no. 1, pp.
108–120, Feb 2018.
[1] A. Hakiri and P. Berthou, “Leveraging SDN for the 5g networks: Trends, [10] H. Yao, S. Ma, J. Wang, P. Zhang, C. Jiang, and S. Guo, “A continuous-
prospects and challenges,” CoRR, vol. abs/1506.02876, 2015. [Online]. decision virtual network embedding scheme relying on reinforcement
Available: http://arxiv.org/abs/1506.02876 learning,” IEEE Transactions on Network and Service Management,
[2] I. Ishaq, J. Hoebeke, I. Moerman, and P. Demeester, “Internet of things vol. 17, no. 2, pp. 864–875, 2020.
virtual networks: Bringing network virtualization to resource-constrained [11] Z. Yan, J. Ge, Y. Wu, L. Li, and T. Li, “Automatic virtual network
devices,” in 2012 IEEE International Conference on Green Computing embedding: A deep reinforcement learning approach with graph convo-
and Communications, Nov 2012, pp. 293–300. lutional networks,” IEEE Journal on Selected Areas in Communications,
[3] M. Chowdhury, M. R. Rahman, and R. Boutaba, “Vineyard: Virtual vol. 38, no. 6, pp. 1040–1057, 2020.
network embedding algorithms with coordinated node and link mapping,” [12] K. T. D. Nguyen and C. Huang, “An intelligent parallel algorithm for
IEEE/ACM Transactions on Networking, vol. 20, no. 1, pp. 206–219, online virtual network embedding,” in 2019 International Conference
Feb 2012. on Computer, Information and Telecommunication Systems (CITS), Aug
[4] X. Cheng, S. Su, Z. Zhang, H. Wang, F. Yang, Y. Luo, and J. Wang, 2019, pp. 1–5.
“Virtual network embedding through topology-aware node ranking,” [13] E. W. Zegura, K. L. Calvert, and S. Bhattacharjee, “How to model an
SIGCOMM Comput. Commun. Rev., vol. 41, no. 2, p. 38–47, Apr. 2011. internetwork,” in Proceedings of IEEE INFOCOM ’96. Conference on
[Online]. Available: https://doi.org/10.1145/1971162.1971168 Computer Communications, vol. 2, March 1996, pp. 594–602 vol.2.
[5] M. Feng, J. Liao, J. Wang, S. Qing, and Q. Qi, “Topology-aware virtual [14] N. M. K. Chowdhury and R. Boutaba, “A survey of network
network embedding based on multiple characteristics,” in 2014 IEEE virtualization,” Computer Networks, vol. 54, no. 5, pp. 862 – 876,
International Conference on Communications (ICC), 2014, pp. 2956– 2010. [Online]. Available: http://www.sciencedirect.com/science/article/
2962. pii/S1389128609003387
[6] P. Zhang, H. Yao, and Y. Liu, “Virtual network embedding based on the [15] C. Huang and J. Zhu, “Modeling service applications for optimal parallel
degree and clustering coefficient information,” IEEE Access, vol. 4, pp. embedding,” IEEE Transactions on Cloud Computing, vol. 6, no. 4, pp.
8572–8580, 2016. 1067–1079, Oct 2018.