Joint Node-Link Embedding Algorithm based on

Genetic Algorithm in Virtualization Environment

Khoa Nguyen , Qiao Lu , and Changcheng Huang
Department of Systems and Computer Engineering
Carleton University, Ottawa, ON K1S 5B6, Canada
Email: {khoatnguyen, qiaolu, huang}@sce.carleton.ca

Abstract—Virtual network embedding (VNE), that efficiently In similarity to NFV-FGE, the VNE process includes two
tackles the mapping problems of heterogeneous virtual networks different stages: VNoM and VLiM. Most of VNE approaches
onto a shared physical infrastructure meeting rigid resource have been solved these mapping stages separately. Specifically,
constraints, is the major challenge in network virtualization
(NV). VNE is widely known as N P-hard due to its intractable a large number of VNE algorithms are proposed to deal
computation. The majority of VNE solutions have concentrated with the VNoM problem, leaving the VLiM phase for the
upon virtual node mapping (VNoM) and virtual link mapping popular shortest path methods (e.g., Dijkstra’s algorithm),
(VLiM) separately. Uncoordinated approaches would facilitate multi-commodity flow (MCF) or recently the distributed
algorithmic implementation, but they lead to low acceptance parallel GA-based algorithm [7]. Decoupling can simplify the
ratio, network revenues and high embedding cost. In this paper,
we propose a new approach relied on Genetic Algorithm (GA), complex implementation of VNE algorithms, but the common
that coordinately joints node and link mappings where the link solution has apparently to sacrifice some degrees of optimality.
Furthermore, a lack of coordination between VNoM and VLim
embedding is based on a fast and efficient sequential path search-
ing method. A novel heuristic conciliation algorithm is presentedphases could result in low acceptance ratio and low network
to deal with a set of infeasible link mappings during producing revenue accordingly. For example, VNoM stage can determine
VNE solutions in GA’s operations. Extensive evaluation results
indicate that our proposed approach outperforms state-of-the- potential substrate nodes for embedding virtual nodes of a
art VNE algorithms in all adopted performance metrics. given VNR, but there might have a situation that no feasible
Index Terms—Network virtualization, virtual network embed- path is found for virtual link that connects a pair of the
ding, joint node-link mapping, conciliation strategy, Genetic already-mapped virtual nodes. In that case, the corresponding
Algorithm. VNR is definitely rejected. It is also observed that the most
I. I NTRODUCTION common failures of VNE problem invariably derive from
NV is widely envisioned as a promising paradigm for the VLiM stage [8]. In this paper, we propose an efficient
the foreseen success of future networks such as virtualized VNE algorithm relied on GA algorithm to address the VNE
fifth generation [1], smart Internet of Things (IoT) networks problem by jointly embedding virtual nodes and corresponding
[2]. Virtualization efficiently enhances network utilization by virtual links. GA tries to seek for potential solutions for
sharing physical resources among several VNRs and provides all virtual nodes of the VNR. After that, the corresponding
an isolated coexistence between VNs on a shared substrate virtual links will be consecutively mapped by a sequential path
network (SN). A service provider (SP) in virtualization searching method. There might have no feasible paths found
environment typically converts a service/application into a for virtual link requests due to network congestion, the current
VN, and then transfers it to the corresponding infrastructure virtual nodes embedding should be reexamined with a goal of
provider (InP) under a form of a VNR. The InP is responsible minimizing node and link mappings revisited. Consequently,
for embedding the VN on SN, satisfying multiple stringent we present a novel conciliation algorithm to handle this issue.
resource constraints. VNE, that is one of the most challenging Additionally, we propose a distributed parallel GA-based
tasks in NV, allows to embed multiple VNRs on the shared paradigm exploiting a set of distributed parallel machines
SN with various topology and rigid resource demands. VNE is in order to reduce the operation time. An execution time
commonly acknowledged as a generalization of second stage comparison between sequential and parallel manners is also
of network function virtualization resource allocation (NFV- conducted. In this work, we only consider unsplittable-enabled
RA), or virtual network function forwarding graph embedding mapping solutions for VNE problem.
(NFV-FGE) due to a mutual problem domain. The main goal The remainder of this paper is organized as follows:
of both NFV-RA and VNE is allocating VNRs on the top of the network model is formulated in Section II. Genetic
physical infrastructure efficiently [3]. In addition, NFV-FGE Algorithm approach for jointly node-link VNE mapping is
is the most critical problem of NFV-RA following virtual described in Section III. Performance evaluation is introduced
network functions (VNFs) chain composition (or service in Section IV.Finally, Section V is a conclusion of this paper.
function chaining) and VNFs scheduling, where the chain
composition is normally neglected in most research papers [3]. II. N ETWORK M ODEL AND P ROBLEM D ESCRIPTIONS
In some particular cases and topology, VNE is even revealed A. Virtual Network Assignment
to be more complicated than NFV-FGE [4], [5]. VNE is N P- Similar to [6], our SN is modelled as a weighted undirected
Hard either for VNoM or VLiM [3], [6]. Thus, the majority of graph Gs = (N s , Ls ), in which N s is a set of physical nodes
research papers in this filed have merely focused on designing and Ls is a set of physical links. A substrate node ns ∈ N s
efficient heuristics or meta-heuristic algorithms to address the has its geographical location loc(ns ) and the available CPU
hurdles of the exact optimization models. capacity c (ns ), whereas a substrate link ls ∈ Ls between two
substrate nodes has a b (ls ) bandwidth capacity. Memory and Acceptance ratio: is a ratio between the number of accepted
storage resources are neglected in this paper for simplification. VNRs over the number of arrived VNRs during an interval
Similarly, the ith VNR can be modelled as a weighted time τ is computed as following:
undirected graph denoted as Gvi = (Niv , Lvi ), where Niv and ξ a (τ )
Lvi are the sets of virtual nodes and virtual links respectively. Πτ = (10)
v v v
ξ(τ )
Every virtual node ni ∈ Ni requires a CPU capacity c(ni ),
a
whereas a virtual edge liv (svi , dvi ) ∈ Lvi between a virtual where ξ (τ ) and ξ(τ ) is the number of the successfully
source node svi and a virtual destination node dvi demands a embedded VNRs and the number of VNRs respectively.
bandwidth capacity b (liv ). Each VNR prefers an embedding Node utilization indicates the distribution of workloads on
v v
radius D(ni ) that exposes how far ni is allocated from the corresponding SN. Node utilization defines the amount of
v
loc(ni ). network resources occupied by virtual node requests during a
This paper is aimed at multiple objectives consisting certain time, divided by the total amount of node resources.
of maximizing average acceptance ratio, average revenue, Node utilization can be expressed as follows:
reducing lower cost and improving node and link utilization
X
c(nvi )
meeting several node and link constraints: s
X n v →ns
i
UN (N ) = ( ) ∗ Ti , (11)
c(nvi ) ≤ RN (AN (nvi )) (1) c(ns )
ns ∈N s
D(loc(nvi ), loc(AN (nvi ))) ≤ D(nvi ) (2) where Ti represents the duration of the accepted ith VNR.
Link utilization: Similarly, link utilization can be presented
AN (nvi ) ∈ N s (3)
X as follows:
RN (ns ) = c(ns ) − c(nvi ) (4) X
b(liv )
nv →ns
X lv
i →l
s

RL (es ) ≥ b(liv ), ∀es ∈ E s (AL (liv )) (5) UL (Ls ) = ( ) ∗ Ti , (12)

b(ls )
ls ∈Ls
RL (es ) = min RL (ls ) (6)
s sl ∈e Fitness Function (FF) is utilized to evaluate the quality
X of VNE solutions. Fitness values can serve as rewards for
RL (ls ) = b(ls ) − b(liv ) (7)
lv s
guiding the searching process to the optimal solutions. In this
i →l
paper, FF considers multiple factors including embedding cost,
where AN (nvi ) denotes the mapping solution of the virtual hop-count and propagation delay, determining the optimal
node. D(is , j d ) represents the distance between is and j d , and mapping solution for a VNR. These factors reflect total
RN (ns ) is the residual CPU capacity of the physical node. network resources comprising both CPU and bandwidth as
Moreover, E s (AL (liv )) is a set of all available physical paths well as network latency. Our fitness function F(S) can be
from the source AN (svi ) to destination node AN (dvi ). RL (es ) expressed as below:
and RL (ls ) denote available bandwidth of a substrate path 1 1
es ∈ E s and the remaining physical link capacity respectively. F(S) = (
Υ(Gvi )
) ∗ wc + ( P ) ∗ wh
li ∈Li hAL (li )
v v v
An embedding solution is defined as “feasible” if and only (13)
1
if it satisfies all resource constraints (1)-(4) for VNoM and +( P v
) ∗ wp
lv v dP (AL (li ))
(5)-(7) for VLiM. i ∈Li

where, S, h and dP are a feasible solution, hop-count

B. Performance metrics and propagation delay of the link mapping solution of liv
The major objectives of a VNE problem are to maximize respectively. wc , wh , and wp are weight parameters equivalent
the InP revenues by accepting as many VNRs as possible, to cost, hop-count and propagation delay.
and to minimize the mapping cost. InP’s revenue can be III. J OINT NODE - LINK E MBEDDING A LGORITHM
calculated as the sum of total virtual resources mapped on GA is a AI algorithm that is able to solve constrained or un-
the corresponding SN over time, whereas the embedding cost constrained optimization problems. A GA algorithm typically
of the ith VNE Ξ(Gvi ) is defined as the sum of total physical comprises four primary operations: population initialization,
resources allocated to the ith VN [6]. selection, crossover and mutation. Our proposed parallel GA
Revenue of ith VNR Gvi can be formulated as below: scheme is presented in Fig 1.
0
X X
Υ(Gvi ) = wα ∗ b(liv ) + wβ ∗ c(nvi ) (8) A. Sequential Path Searching Algorithm
lv v
i ∈Li nv v
i ∈Ni
Indeed, SN topology is almost static, which means that it
where b(liv ) and c(nvi ) are the requested bandwidth of the is able to determine a number of physical paths for each
virtual link liv and the requested CPU of the virtual node nvi pair of source and destination nodes in advance. In this
while wα and wβ are the unit weights of the bandwidth and paper, a path database for link mappings are extensively
CPU resources respectively. established before VNR arrivals. The shortest path mechanism
Cost of ith VNR Gvi can be formulated as below: (e.g., Dijkstra’s algorithm) based on the hop-count factor is
X X X lv deployed to construct the path database which is recognized
Ξ(Gvi ) = c(nvi ) + flsi (9)
as the initial path pool generation in Fig. 1. Minimizing
nv v
i ∈Ni lv v s
i ∈Li l ∈L
s
hop-count can improve the embedding cost and physical
lv
where flsi defines the bandwidth of substrate link ls that is link utilization reflected by path length metric, leaving
allocated to the virtual link liv . more resources to future VNRs to be accepted. The first
 nodes in a given VNR are assigned one by one to feasible
substrate nodes that are selected from the initial node pool in
random. In fact, the random VNoM can avoid the possible
premature convergence problem of GA algorithm.
After the mappings of all virtual nodes in a VNR are
identified, a sequential path searching method described in
III-A is deployed to discover feasible link mappings for the
virtual link requests thanks to the information of already-found
node mappings. These substrate paths must meet all virtual
resource demands (Eq. (5)-(7)) to become the potential link
mapping solutions. When a feasible solution for the whole
VNR has been successfully recognized, the chromosome can
be established. Otherwise, if there is no feasible path (e.g.,
network congestion) found for a virtual link request, the pair
Fig. 1: Parallel operation scheme of corresponding virtual nodes need to be remapped. The
more virtual links are failed to map, the more complicated
path of each pair of source-destination nodes has minimum problem is. Instead of finding appropriate virtual nodes to
hop count. Additionally, by forwarding data to less nodes, remap, VNoM process can be started over again, but this
overheads imposed by sending repetitive data and energy approach is obviously inefficient and might overlook optimal
consumption can be improved. When a virtual link request solutions.
arrives for mapping, our sequential path searching algorithm
consecutively seeks for a feasible path in database based on Algorithm 1 Heuristic Conciliation Algorithm
the node mapping information. If the first path does not meet
1: Input:
the resource requirements of the VNR (e.g., bandwidth), the 2: Initial solutions of virtual node mapping
algorithm will approach the consecutive one until at least 3: A set of failed virtual links after link mapping phase
one feasible path is achieved. As such, the feasible path 4: Output:
possesses minimal hop-count feature. Otherwise, if there is 5: Virtual nodes to be remapped
not any feasible found, the mapping returns failed. Since the 6: procedure H EURISTIC CONCILIATION STRATEGY
7: Step 1: Construct a map Mp of virtual nodes based on their
path database can be generated prior to the arrival of VNRs, presence in a set of failed virtual links
time counted for this process can be neglected. In fact, this 8: Step 2: Create a map Mn of virtual nodes based on their
searching mechanism is simple, fast and efficient. nearest neighbors . These steps can be done in advance prior
this algorithm
B. Conciliation Strategy 9: Initialize an array for remapped nodes
Population in GA contains multiple chromosomes that are 10: for each infeasible virtual link do
11: if svi or dvi NOT in the array then
generated in random manner. A given VNR typically consists 12: if Mp [svi ] > Mp [dvi ] then
of a set of virtual nodes connected by multiple virtual links. 13: add svi into array
Consequently, there are two sets of mapping solutions for 14: else if Mp [svi ] < Mp [dvi ] then
virtual nodes and virtual links. The node mappings apparently 15: add dvi into array
influence the later ones; consequently, if the node mappings 16: else
17: if Mn [svi ] > Mn [dvi ] then
are altered, the link solutions are correspondingly changed. 18: add dvi into array
As a result, our proposed GA algorithm defines the mapping 19: else
solution of a virtual node as a gene. In this paper, virtual 20: add svi into array
21: end if
22: end if
23: end if
24: end for
25: return node array
26: end procedure

We aim at minimizing the number of virtual nodes asso-

ciated with the failed-mapping virtual links that should be
re-embedded. Let’s take an example with a specific VNR with
7 virtual nodes and 6 virtual links as illustrated in Fig. 2.
Fig 2a in the top-left is a complete situation as all virtual
nodes and links are explored their mappings successfully.
Fig. 2: Examples of infeasible virtual link mappings However, next figures demonstrate some “real” problems of
nodes in a given VNR are embedded in order as similar to link mapping failures. For example, there are three failed
NFV-FGE, so their order in all chromosome is also the same. virtual link requests {A - G, A - F and A - D} according to
Towards improving the capability and reducing the execution the node mappings {A → h, G → n, F → m and D → k} as
time, a feasible node pool, in which all feasible substrate shown in Fig. 2b. Unfortunately, no feasible paths are found
nodes satisfying the least CPU capacity requirement of the between {h → n, h → m, and h → k}. The problem is
VNR are collected, is advised to be first created. Then, virtual that if we revisit all already-mapped virtual nodes (definitely
ineffective) or virtual nodes {D, F and G}, remapping at least from the starting point j c + 1 to |N | as depicted in (15).
five virtual links is required. Otherwise, if the virtual node Mutation usually applies a modification on an individual
A, is only re-embedded, the virtual links that need to be parent for delivering new offspring. This operation is aimed
remapped are merely three. In similarity, virtual nodes C/D, at sampling the broad solution space as well as enhancing the
E and G can be revisited in Fig. 2 and virtual nodes C and efficiency of solution exploration. Mutation also avoids poten- m
E should be reconsidered in the last figure by cause of least tial solutions falling into the local optima. Denote j m and grj0
remapping. Motivated by this idea, we propose a heuristic as the random mutation point and new gene substituting the
conciliation algorithm to deal with this issue as described in existing one in C(M+1) , respectively. New solution C 0 (M+1)
m
Algorithm 1. after substitution is defined as C 0 (M+1) = [gs1 · · · grj0 · · · gsN ].
C. Working machines
Population Initialization: As shown in Fig. 1, every work- D. Sorting and Terminations
ing machine running GA algorithm begins with a population The mapping solutions are sorted to select the best one
initialization operation. M is the set of chromosomes where relied on their higher fitness values. It is then delivered to
each chromosome consists of N = |Niv | genes considered synchronization for the global ranking. Waiting for every
as the potential mappings of all virtual nodes of the VNR. worker machine accomplishing its assigned task is trouble-
The initial population P (MxN ) at a working machine is some, without guaranteeing an expected outcome. Hence,
established as elucidated in Section III-B. As soon as a the master node aborts GA algorithms running in working
chromosome is constructed, the sequential path searching machines if no better mapping is obtained within t times,
mechanism (Section III-A) based on the already-achieved where t denotes a termination parameter.
node mappings is implemented for VLiM stage. E. Synchronization and VNR allocation
   1 j
g1 · · · g1 · · · g1N
 In this process, the best mappings of the given VNR is
C1
 C2   1 j
g2 · · · g2 · · · g2 N recognized by globally ranking the solutions received from
  working machines, based on highest fitness values. Finally,

 ..   .. .. .. 
  .. .. 
 .   . . . . . network updates allocated resources.
P= = 1 (14)

 Cf j N 


  gf ··· gf ··· gf 
 .  ..

.. .. .. ..
 IV. P ERFORMANCE E VALUATION
 ..
 
  . . . . . 
CM 1
gM ··· j
gM ··· N
gM
Our proposed GA-based algorithm, namely GASQ, is com-
pared with several state-of-the-art VNE competitors including
Thereupon, a conciliation algorithm addresses possible failures DPGA [7], NTANRC-S [9], R-ViNE and D-ViNE [6] on
of virtual link mappings, minimizing remapping virtual various performance metrics.
node/link mappings. When feasible solutions of all virtual
nodes and links are achieved, the chromosome of a given A. Simulation setup
VNR is eventually formed. It is noted that any change on We have developed a discrete-event simulator to evaluate
the node mappings leads to the associated changes of link the GASQ algorithm with similar parameters set out in [6].
mappings. The popular GT-ITM topology generator [10] is deployed
Evolved generations: Random chromosomes in population to generate SNs and VNs, where SNs are configured with
are selected to become parents for producing their children average 50 nodes and 140 edges adopting Waxman model
or next generations as the results of crossover and mutation (α = 0.5 and β = 0.2). CPU and bandwidth capacity of SNs
operators. The fast and efficient path searching method is are uniformly generated between 50 and 100 units, whereas
deployed to approach link mappings for new generated VNRs arrive dynamically following the Poisson process with
offspring. The population will update these new generations an average rate λ varying from 4 to 8 VNs per 100 time units.
to strengthen the population diversity. Lifetime of VNRs follows an exponential distribution with
Crossover combines parental chromosomes to generate new average µ = 1000 time units, so the workload of VNRs can
offspring. be quantified by µλ Erlangs. Towards VNs, the number of
c c virtual nodes in each VNR is uniformly distributed between 2
g11 ··· g1j g1j +1 ··· g1N
   
C1 and 10, where CPU and bandwidth requirements of VNRs are
..   .. .. .. .. .. 
. . . . . uniformly generated between 0 to 20 and 0 to 50 respectively.

 .   
  c c

gs1 ··· gsj gsj +1 ··· gsN

 Cs  
 

 Each simulation runs 50, 000 time units, 50 times longer than
.. .. ..

 ..   .. .. 
the average lifetime of a VN, exceptionally generating a large
 .   . . . . .

  number of independent samples. All performance figures are

P= = c
j c +1

Cr gr1 ··· grj ··· grN

  gr 

 ..   .. .. .. .. ..
 plotted with average values with 95% confidence interval.
  . .

. . . .

   
B. Evaluation Results
  jc j c +1


 CM   gM1
··· gM gM ··· N
gM 
CM+1   As illustrated in Fig 3a, GASQ algorithm outperformed all
c c

gs1 gsj grj +1 grN
  ··· ··· 
CM+2 c c
compared algorithms in terms of acceptance ratio, resulting to
gr1 ··· grj gsj +1 ··· gsN
(15) highest average revenue which is the most desired objective
Denote Cs and Cr as parental chromosomes indexed s and r in for any embedding approach. Our proposed algorithm also
the initial population. Their new descendants are represented proved its efficiency by achieving the lowest cost (Fig. 3c).
as C(M+1) and C(M+2) respectively, where j c is the random Particularly, GASQ performed better acceptance ratios than
crossover point between any genes within |N |. In crossover, DPGA for more than 19.02% and 13.62% at 40 and 80
offspring is produced by swapping genes between parents Erlang respectively. GASQ also outperformed NTANRC-S,
 1.0
GASQ 100
DPGA GASQ 220 GASQ
NTANRC-S DPGA DPGA
0.9 R-ViNE 95 NTANRC-S NTANRC-S
D-ViNE R-ViNE R-ViNE
200
Average acceptance ratio

D-ViNE D-ViNE

Average generated revenue

90
0.8

Average cost
85 180
0.7
80
160
0.6
75
140
0.5 70
40 50 60 70 80
Traffic load (Erlang)
40 45 50 55 60 65 70 75 80 40 45 50 55 60 65 70 75 80
Traffic load (Erlang) Traffic load (Erlang)

(a)
(b) (c)
Fig. 3: (a) Average Acceptance Ratio (b) Average generated revenue (c) Average mapping cost
0.50 0.60
GASQ GASQ 3.4
DPGA 0.58 DPGA
0.45 NTANRC-S NTANRC-S 3.2
R-ViNE 0.56 R-ViNE
0.40 D-ViNE D-ViNE 3.0
Average Node Utilization

Average link utilization 0.54

Average path length

0.35 2.8
0.52
2.6
0.30 0.50

0.48 2.4
0.25 GASQ
0.46 2.2 DPGA
0.20 NTANRC-S
0.44 2.0 R-ViNE
D-ViNE
0.15 0.42 1.8
40 45 50 55 60 65 70 75 80 40 45 50 55 60 65 70 75 80 40 45 50 55 60 65 70 75 80
Traffic load (Erlang) Traffic load (Erlang) Traffic load (Erlang)

(a) (b) (c)

Fig. 4: (a) Average node utilization (b) Average link utilization (c) Average path length

R-ViNE and D-ViNE for more than 20% at the highest sequential and parallel operations of our proposed solution is
traffic load. Our joint node and link approach improved consequently provisioned. The extensive evaluation indicates
average revenue and cost up to 11.77% and 27.78% compared that the joint node and link approach based on GA algorithm in
to the aforementioned algorithms at the same traffic load a single embedding stage outperforms state-of-the-art heuristic
respectively. Furthermore, the node utilization of the proposed VNE algorithms in all indispensable performance metrics we
algorithm was much better than all competitors ranging from adopted.
10.15% up to 18.93% at various traffic loads as shown in R EFERENCES
Fig. 4a. Fig. 4b and 4c indicated that GASQ and DPGA [1] A. Hakiri and P. Berthou, “Leveraging SDN for the 5g networks:
had average path lengths shorter than the other algorithms, Trends, prospects and challenges,” CoRR, vol. abs/1506.02876, 2015.
contributing to lower embedding cost and improved average [Online]. Available: http://arxiv.org/abs/1506.02876
[2] I. Ishaq, J. Hoebeke, I. Moerman, and P. Demeester, “Internet of things
link utilization. Thanks to less bandwidth consumption, a virtual networks: Bringing network virtualization to resource-constrained
plethora of remaining bandwidth allowed to accept more devices,” in 2012 IEEE International Conference on Green Computing
arriving VNRs. The remarkable results were of the efficient and Communications, Nov 2012, pp. 293–300.
[3] J. Gil Herrera and J. F. Botero, “Resource allocation in nfv: A
sequential path searching based on hop-count factor in a novel comprehensive survey,” IEEE Transactions on Network and Service
joint node-link approach and the multiple constrained FF Management, vol. 13, no. 3, pp. 518–532, 2016.
driving the GA algorithm to approach the optimal mapping [4] T. A. Q. Pham, J.-M. Sanner, C. Morin, and Y. Hadjadj-Aoul, “Virtual
network function–forwarding graph embedding: A genetic algorithm
solutions. Our proposed algorithm successfully solved the approach,” International Journal of Communication Systems, vol. 33,
lack of coordination between node and link mappings, which no. 10, p. e4098, 2020, e4098 0.1002/dac.4098. [Online]. Available:
leads to an ultimate embedding outcome that could be a https://onlinelibrary.wiley.com/doi/abs/10.1002/dac.4098
[5] B. Addis, G. Carello, and M. Gao, “On a virtual network functions
tough challenge for any future VNE algorithm. Additionally, placement and routing problem: Some properties and a comparison of
we compared execution time of our VNE approach on both two formulations,” Networks, vol. 75, no. 2, pp. 158–182, 2020. [Online].
sequential and parallel schemes in order to quantify time Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/net.21915
[6] M. Chowdhury, M. R. Rahman, and R. Boutaba, “Vineyard: Virtual
reduction of parallel operation. Due to the distributed parallel network embedding algorithms with coordinated node and link mapping,”
paradigm, GASQ needed 1.19s to allocate a VNR whereas IEEE/ACM Transactions on Networking, vol. 20, no. 1, pp. 206–219,
the sequential counter part finished the same task in 9.73s. Feb 2012.
[7] K. T. Nguyen, Q. Lu, and C. Huang, “Rethinking virtual link mapping
Regarding time complexity, interested readers may refer to in network virtualization,” in 2020 IEEE 92nd Vehicular Technology
the paper [7] for further theoretical analysis. Conference (VTC2020-Fall), 2020, pp. 1–5.
[8] Hong-Kun Zheng, J. Li, Y. Gong, W. Chen, Zhiwen Yu, Z. Zhan, and
V. C ONCLUSION Ying Lin, “Link mapping-oriented ant colony system for virtual network
This paper proposes a joint node and link GA-based embedding,” in 2017 IEEE Congress on Evolutionary Computation
(CEC), June 2017, pp. 1223–1230.
embedding algorithm for dealing with the VNE problem [9] H. Cao, L. Yang, and H. Zhu, “Novel node-ranking approach and
in one stage. A novel heuristic conciliation algorithm is multiple topology attributes-based embedding algorithm for single-
utilized to handle multiple infeasible link mappings due to domain virtual network embedding,” IEEE Internet of Things Journal,
vol. 5, no. 1, pp. 108–120, Feb 2018.
the inappropriate node embedding in GA operations. We [10] E. W. Zegura, K. L. Calvert, and S. Bhattacharjee, “How to model an
also advise a distributed parallel operation scheme to reduce internetwork,” in Proceedings of IEEE INFOCOM ’96. Conference on
execution time of GA algorithm. A comparison between Computer Communications, vol. 2, March 1996, pp. 594–602 vol.2.