IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 1

Towards Adaptive Joint Node and Link Mapping

Algorithms for Embedding Virtual Networks:
A Conciliation Strategy
Khoa Nguyen , and Changcheng Huang , Senior Member, IEEE,

Abstract—Network virtualization (NV) has emerged as a network layer while SPs will aggregate network resources that
momentous facilitator for a notable triumph of future networks are leased from different InPs into VNs providing customised
by allowing a flexibility, cost-efficiency and on-demand services services to end users. In a virtualization environment, a network
through the deployment of heterogeneous network service requests
on a shared physical infrastructure. The most major challenge of service provider (SP) typically converts an application or
NV is to efficiently and effectively map diversified virtual network service into a VN which is then transferred to a corresponding
requests (VNRs), comprising a set of virtual nodes connected by infrastructure provider (InP) as a form of a VNR as illustrated
virtual links, onto a shared substrate network meeting various in Fig. 2. The InP attempts to map such VN onto its substrate
stringent resource constraints. Most of the research papers in infrastructure meeting several rigorous resource constraints by
this field have merely focused on separate virtual node mapping
(VNoM) or virtual link mapping (VLiM) with scalable heuristic adopting an optimization process. In terms of InPs, they attempt
algorithms for simple implementations. Unfortunately, the lack of to approach an efficient resource allocation algorithm that is
a coordination between node and link mapping stages might cause capable to maximize their generated revenues by accepting
low embedding results. In this paper, we present a new approach as many VNRs as possible while keeping the mapping costs
relied upon Genetic Algorithm (GA), that jointly coordinates minimal.
virtual node and link mappings where the link mapping is
based on three different path searching methods. Moreover, A conventional VNR is consisted of a set of virtual nodes that
a novel heuristic conciliation mechanism is proposed to deal are connected by virtual links in order to establish an explicit
with a set of possibly infeasible link mappings while exploring topology. In most of scenarios, it dynamically arrives and
embedding solutions within the operations of GA algorithm. resides in the corresponding network for an arbitrary duration.
Extensive performance results indicate that our proposed GA- Mapping VNRs onto the shared physical infrastructure with
based algorithms outperform state-of-the-art virtual mapping
algorithms in all evaluation metrics we adopt. varied topology and stringent resource demands is commonly
recognized as a virtual network embedding (VNE) problem.
Index Terms—Network virtualization, joint node-link mapping, VNE is widely known as a generalization of second stage of
conciliation strategy, Genetic Algorithm, virtual network embed-
ding.
I. I NTRODUCTION

N ETWORK virtualization (NV) has envisioned as a de

facto paradigm for the foreseen triumph of future net-
works such as virtualized fifth generation [1], and smart
Internet of Things (IoT) networks [2]. The adoption of NV
enables to share physical resources amongst multiple virtual
network requests (VNRs), providing an isolated coexistence
of a number of virtual network (VNs) on a shared underlying
substrate network (SN). Virtualization efficiently enhances
physical network utilization, and simplifies the development Fig. 1: Towards the Future internet Model [3]
and evaluation of new network protocols or architecture designs
without unexpected expansion. network function virtualization resource allocation (NFV-RA),
The fundamental concept of NV is to decouple a typical role namely virtual network function forwarding graph embedding
of an Internet Service Provider (ISP) into two separate business (VNF-FGE) since VNE is in the same problem domain with
components as illustrated in Fig. 1: Infrastructure Providers NFV-RA as the goal of both problems is to efficiently allocate
(InPs) and Service Providers (SPs). InPs will manage the VNRs on the top of SN infrastructure [4]. VNF-FGE is also the
substrate resources, and control Quality of Service (QoS) at the central problem of NFV-RA accompanying with virtual network
functions (VNFs) chain composition (known as service function
Khoa Nguyen and Changcheng Huang are with the Department of Systems chaining) and VNFs scheduling, where the chain composition
and Computer Engineering, Carleton Univeristy, Ottawa, ON, CA e-mail:
{khoatnguyen, huang}@sce.carleton.ca. Some primary results of GA-SEQ stage has been usually neglected [4]. Additionally, VNE is
algorithm in this paper are partly presented at the 2021 IEEE 94th Vehicular assumed to be more complicated than VNF-FGE in some
Technology Conference (VTC2021-Fall), 2021, online. This research was specific cases and topology [5], [6].
supported by the Natural Sciences and Engineering Research Council of
Canada (NSERC) Engage grant (RGPIN-2017-06331). VNE is validated to be N P-Hard either towards VNoM or
Corresponding author: Khoa Nguyen (e-mail: [email protected]) VLiM [4], [7]. Integer Linear Programming (ILP) is usually
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 2

formulated to solve the VNE problem, but it encounters many to solve VNE problem by jointly mapping virtual nodes and
challenges since its model has to confront with scalability, corresponding virtual links. GA algorithm attempts to randomly
intricate deployment as well as time complexity. In fact, Exact find potential solutions for all virtual nodes in a VNR. Then,
solutions are not apt to deal with the online VNE problem. The various path searching methods are conducted to seek for the
majority of VNE studies have concentrated on efficient designs corresponding virtual links sequentially. If no feasible paths
of heuristics or metaheuristics to tackle the aforementioned for some possible link requests are achieved due to network
obstacles of the Exact optimization models. Similar to VNF- congestion for example, corresponding mappings should be
reassessed. Because the node mappings influence the link
mappings accordingly, we should figure out an intelligent
mechanism that considers the minimal number of virtual
nodes and virtual links for reassessments. Thus, we introduce
an efficient conciliation algorithm to address this problem.
Through iterative operations, GA algorithm, driven by an
effective fitness function that considers both network resources
and virtual resource demands, generates the population of
several promising VNE solutions including node and link
mappings. Furthermore, we present a distributed parallel GA-
based operation scheme employing a set of distributed parallel
machines to reduce the operation time. A brief comparison
towards execution time between sequential and parallel manners
is also provided. Splittable mapping may theoretically achieve
better resource utilization, but this solution could generate
abundant overhead in order to maintain consistent network
state and likely results in out-of-order delivery problem with
extra latency that could be unacceptable for sensitive-delay
Fig. 2: The concept of network virtualization [8]
applications [11]. Thus, we only consider unsplittable-enabled
FGE, VNE process consists of two stages: VNoM and VLiM. mapping solutions in this paper.
Up to now, most VNE approaches have been proposed to Major contributions of this paper are summarized as below:
uncoordinately solve VNoM phase following VLiM phase.
• We propose a new joint node and link mapping approach
There are a large number of solutions for VNoM problem
whereas VLiM is usually relied on shortest path methods based on GA algorithm to efficiently and effectively embed
(e.g., Dijkstra’s algorithm), multi-commodity flow (MCF) and heterogeneous VNRs, where the link mapping is relied on
recently distributed parallel GA algorithms [9]. The decoupling three different virtual link mapping algorithms including
can facilitate the complexity of algorithmic deployments, but random path searching, sequential single path searching
this common approach is most likely to sacrifice some degrees and k-shortest path method. We manage node and link
of optimality. Lack of coordination between VNoM and VLim mappings in a coordinated manner so when node mappings
stages would lead to low acceptance ratio and correspondingly are changed, the link mappings are altered accordingly.
• To deal with a possible set of infeasible link mappings
low network revenue as well as high cost. For instance, although
VNoM stage can figure out highly promising substrate nodes caused by improper node mappings, we introduce a
for mapping virtual nodes in a VNR, we may fall into an novel heuristic conciliation mechanism which is aimed at
unexpected situation that there are no paths in the SN that keeping the number of both remapped nodes and links
have sufficient remaining resources to connect one or more minimal.
• An efficient fitness function with various network resource
pairs of those virtual nodes together with respect to node
mappings in previous phase. Then, the corresponding VNR is factors including embedding cost, hop-count, and prop-
obviously rejected. In practice, the most common failures of agation delay is introduced to drive GA algorithm to
embedding VNRs onto the shared underlying SN are usually near-optimal VNE solutions.
• We advise a distributed parallel operation scheme to
derived from the infeasible link mappings [10]. Indeed, those
rejected VNRs can be reassessed after the link mapping stage decrease time complexity of GA algorithm. We also
or be put back into the queue and waiting until the network implement our proposed GA-based algorithms (according
is idle. These approaches are inefficient and unpractical since with three different link mappings) on both sequential
they could violate delay constraints of VNRs, and obviously and parallel manners, and then compare execution time
waste the network resources for processing such failed VNRs between them.
in the past. As a result, it is required VNE algorithms that The rest of this paper is organized as follows. Section II presents
simultaneously consider node and link mappings, and allow the related work while Section III formulates the network model.
an efficient remapping approach if these node mappings are Jointly node-link VNE embedding is described in Section V
inappropriate due to the failure of link mappings. Designing with three variances. Performance evaluation is then introduced
VNE algorithms meeting these demands is still in its infancy. in Section VI. Section VII finally presents our conclusions and
In this paper, we propose metaheuristic-based algorithms future work.
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 3

II. R ELATED W ORK implementations. On the other hand, our previous work in [9],
[27] presented a novel GA-based approach for VLiM stage,
[3], [12], [13] and [14] have contributed comprehensive which validated the important role of VLiM in VNE problems.
surveys in network virtualization field. The VNE problem, Although these GA-based approaches performed better than
widely known as N P-hard in nature, is intractably solved other metaheuristics and several heuristic algorithms, they
by Integer Programming (IP). Thus, a large number of solved VNE problems uncoordinatedly, leading low embedding
research papers have focused on heuristic algorithms due to efficiency [3].
time complexity of Exact methods. Chowdhury et al. in [7] Recently, several papers have studied the applicability of
introduced a VNE coordination between node and link mapping Reinforcement Learning (RL) techniques on VNE problems.
phases for embedding virtual nodes with respect to minimizing Haeri and Trajkovic [28] modeled VNE problem as Markov De-
the embedding cost. A Mixed Integer Programming (MIP) cision Process (MDP) for solving VNoM stage and determined
was formulated to solve VNoM considering meta nodes on an action policies using Monte Carlo Tree Search (MCTS) that
augmented graph created over a physical network. Integer was driven by a reward function. Yao et al. [29] presented a RL-
constraints of the MIP model were relaxed to reduce its based model utilizing a dynamic attribute matrix representation
time complexity. Authors in [15] proposed VNE Node-Link (RDAM) to address VNE problem. As such, the attributes
Formulation (VNE-NLF), an ILP model, for mapping each of network topology as well as its specific structures under
VNR. Its objective was to minimize the VN embedding cost a matrix representation were considered. Another extension
while supporting load balancing. Huang et al. in [16] extended of [29] was presented in [30] that was inspired by the node
[7] by taking both node splitting and node collocation into ranking method. To train agents, three network attributes were
account for solving VNE efficiently. The authors studied a extracted for each physical node to construct a feature matrix
realistic VNE scenario in which virtual and substrate nodes that was utilized as the input of the policy network in a
were managed as transit nodes or stub nodes that were Convolutional Neural Network (CNN). Simple BFS mechanism
accountable for traffic reply or traffic originator/sinker. Authors was then deployed in VLiM stage. Extending [30] to five
in [17] presented a topology-aware node mapping based on the network features with a security-aware focus was presented
Markov Random Walk model to evaluate network resources. in [31]. In addition, Recurrent Neural Networks (RNNs) were
A node-ranking approach relied on node degree and node used in [32] in order to model the continuity of VNoM
clustering coefficient information to construct the metric of process in time series formulated as a classic seq2seq, aimed
node importance for ranking the underlying physical nodes was at exploiting the historical states of SN. A continuous-decision
proposed in [18]. Moreover, the authors in [19] and [20] studied VNE scheme based on RL relied on policy-gradient mechanism
the attributes of network topology and the global network updated the RNN parameters with respect to average long-
resources in order to rank substrate nodes for VNoM stage. To term revenue-to-cost ratio. However, the aforementioned VNE
improve the possibility of being selected of a substrate node, approaches merely carry on VNs mapping in separate stages,
its physical resources and the relationship with other nodes in including our previous work in [9], [27] where VLiM was the
the network were considered. These node-ranking approaches focal center of interest. This paper is aimed at coordinately
handled network resource fragmentation efficiently. Authors in joining node and link mappings in a single stage based on
[8] tackled the coordination of VNE with the Mod-MaxMatch GA algorithm. Different path searching methods including
algorithm for VNoM aimed at minimizing the substrate links random path searching, sequential path searching mechanisms
reused, and then the path-splitting enabled GA algorithm was and shortest path method for mapping virtual links are proposed.
proposed for VLiM. To handle failed virtual link mappings due to inappropriate node
GA algorithm has been widely applied in many optimization mappings, a novel conciliation mechanism is also introduced.
fields including VNE. Its applications on VNE were first Taxonomy of the VNE approaches is shown in Table I where
investigated in [21]. The authors proposed GA-based node the category classifies each approach into four features Online
ranking methods relied on multiple topology attributes and or oFfline, Centralized or Distributed, Static or Dynamic and
residual network resources. Zhou et al. in [22] advised GA Concise or Redundant, whereas the coordination is involved
algorithm for embedding multiple VNRs simultaneously under in COorDinated or UnCoorDinated approaches [3].
batched arrivals, and designed a resource-aware mechanism
for maximizing long-term revenue. Zhang et al. modified III. N ETWORK M ODEL AND P ROBLEM D ESCRIPTIONS
a conventional GA algorithm by rearranging the mutation A. Virtual Network Assignment
operator before the selection to efficiently explore the search We model the SN as a weighted undirected graph Gs =
space in [23]. Meanwhile, a hybrid solution, that combined (N s , Ls ), in which N s and Ls are associated with the sets
an adaptive GA algorithm with a new crossover operator to of substrate nodes and substrate links, respectively. Each
enhance GA’s convergence and simulated annealing algorithm physical node ms ∈ N s has its geographical location loc(ms )
replacing a traditional mutation operation of GA to avoid and available CPU capacity C (ms ), whereas a substrate
premature VNE solutions, was introduced in [24]. The authors link (ms , ns ) ∈ Ls between two substrate nodes occupies a
in [25] and [26] applied GA algorithm to address VNE B (ms ns ) bandwidth capacity. Similarly, the ith VNR arriving
problem in multiple InP domains. The solution allowed to in the network is modelled as a weighted undirected graph
embed multiple VNRs on different InPs owning several SNs. denoted as Gvi = (Niv , Lvi ), in which Niv is the set of virtual
However, it faced the problems of scalability and complex nodes and Lvi is the set of virtual links. Each virtual node
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 4

TABLE I: Taxonomy of VNE approaches

Category Reference Optimization Coordination Contribution

O/C/S/R [17] Cheng et al. (2011) Heuristic COD, two-stage The topology-aware node ranking approach based on
Markov Random Walk model in VNE.
O/C/S/R [7] Chowdhury et al. (2012) Like-exact COD, two-stage Relaxed MIP model considering a coordinated node and
link approach for VNoM using k-shortest path method
or MCF for VLiM.
O/C/S/C [21] Mi et al. (2012) Metaheuristic UCD, two-stage One of the first GA-based VNE algorithms for VNoM.
O/C/S/R [15] Melo et al. (2013) Exact COD, one-stage VNE Node-Link Formulation (VNE-NLF) based on the
ILP model for mapping VNRs.
O/C/S/C [20] Feng et al. (2014) Heuristic COD, two-stage Seven topological SN characteristics complementary to
each other for node ranking in VNE.
O/C/S/C [18] Zhang et al. (2016) Heuristic UCD, two-stage A VNE node ranking approach based on the node degree
and clustering coefficient information.
O/C/S/R [22] Zhou et al. (2016) Metaheuristic COD, two-stage GA-based approach for embedding multiple VNRs under
batched arrivals in VNE.
O/C/S/C [25] Pathak et al. (2017) Metaheuristic UCD, two-stage GA-based algorithm for mapping multiple VNRs amongst
multiple InPs.
O/C/S/R [16] Huang et al. (2018) Exact COD, two-stage Node splitting and node collocation for VNoM.
O/C/S/C [19] Cao et al. (2018) Heuristic COD, two-stage Node-ranking approach considering five network topology
attributes and global network resources.
O/C/S/R [28] Haeri et al. (2018) RL UCD, two-stage MDP-based RL using the Monte Carlo tree search
algorithm for VNoM.
O/C/S/C [29] Yao et al. (2018) RL UCD, two-stage RL based dynamic attribute matrix representation algo-
rithm for VNoM.
O/C/S/C [23] Zhang et al. (2019) Metaheuristic UCD, two-stage A modified GA-based approach rearranging mutation
operator right after initial population to explore the
searching space efficiently.
O/C/S/C [24] Boyang et al. (2019) Metaheuristic UCD, two-stage A hybrid adaptive GA-based solution combining an
adaptive GA algorithm with new crossover and simulated
annealing to improve convergence, and avoid premature
phenomenon.
F/C/S/R [8] Huang et al. (2019) Heuristic COD, two-stage Mod-MaxMatch algorithm for VNoM and the path-
splitting enabled GA-based algorithm for VLiM.
O/C/S/C [9] Nguyen et al. (2020) Metaheuristic UCD, two-stage Distributed parallel GA-based algorithm for VLiM with
a novel elastic crossover operation and a multi-objective
fitness function.
O/C/S/C [27] Lu et al. (2020) Metaheuristic UCD, two-stage Two distributed parallel GA-based algorithms based on
two versions of crossover and mutation schemes for
VLiM.
O/C/S/C [30] Zhang et al. (2020) RL UCD, two-stage Constructing an agent learning environment based on the
extracted node features of the SN for training the RL
model in CNN.
O/C/S/C [31] Zhang et al. (2020) RL UCD, two-stage A security-aware VNE algorithm based on RL.
O/C/S/C [32] Yao et al. (2020) RL UCD, two-stage A continuous-decision RL-based VNE scheme updating
RNN’s parameters by the policy-gradient mechanism.
O/C/S/C [26] Zhang et al. (2021) Metaheuristic UCD, two-stage A genetic correlation multi-domain VNE algorithm with
a stochastic selection and a feasibility checking.
O/C/S/C This paper Metaheuristic COD, one-stage Joint GA-based node and link mapping approaches, where
VLiM is based on several link mapping algorithms. A
novel conciliation strategy cleverly handles a set of
infeasible link mappings.

uv ∈ Niv has a CPU demand C(uv ), while a virtual link average revenue over cost ratio, and to improve node and
(uv , k v ) ∈ Lvi from a virtual source node uv to a virtual link utilization and propagation delay, meeting node and link
destination node k v possesses a required bandwidth demand constraints imposed by VNRs.
B (uv k v ). Each VNR generally prefers an embedding radius
D(uv ) revealing the farthest distance the virtual node uv is Node constraints:
allowed to map from its location loc(uv ). D(uv ) can be C(uv ) ≤ RN (AN (uv )) (1)
expressed in reference to physical distance or concerning
permissible delay (e.g., propagation delay) from loc(uv ) similar D(loc(uv ), loc(AN (uv ))) ≤ D(uv ) (2)
to [7]. Without loss of generality, memory and storage are v
AN (u ) ∈ N s
(3)
ignored for simplification. X
RN (ms ) = C(ms ) − C(uv ) (4)
Our objectives are to maximize average acceptance ratio, uv →ms
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 5

where AN (uv ) is the mapping solution of virtual node uv . Revenue to cost ratio (R/C) of ith VNR Gvi is formulated
D(is , j d ) and RN (ms ) denote the distance between is and j d , as follows:
and the residual CPU capacity of a physical node respectively. wb ∗
X
B(uv kv ) + wn ∗
X
C(uv )
Eq. (1), (2) ensure that the residual CPU capacity and mapping Γ(Gvi ) uv kv ∈Lv uv ∈Niv
Υ(Gvi ) =
i
distance of the substrate node meet the requested CPU demand = v v
Ξ(Gvi )
X X X
C(uv ) + u k
fm s ns
and preferred embedding radius of the corresponding virtual
uv ∈Niv uv kv ∈Lv
i
ms ns ∈Ls
node, respectively. Eq. (3) guarantees that the mapped substrate (10)
node belongs to the physical network, whereas Eq. (4) is the
Acceptance ratio: is the fraction between the number of
measurement of available CPU capacity of the substrate node
accepted VNRs over the total number of VNRs arrived in the
ms .
network during a time interval τ is computed as follows:
Link constraints:
ξ a (τ )
RL (es ) ≥ B(uv k v ), ∀es ∈ E s (AL (uv k v )) (5) Aτc = (11)
ξ(τ )
RL (es ) = min RL (ms ns ) (6) where ξ a (τ ) and ξ(τ ) denote the total number of successfully
ms ns ∈es
X mapped VNRs and the total number of VNRs, respectively.
RL (ms ns ) = B(ms ns ) − B(uv k v ) (7)
Node utilization indicates the distribution of network loads
uv kv →ms ns
on the corresponding physical network. It is equal to the total
where E s (AL (uv k v )) is a set of all possible substrate paths network resources that are occupied by virtual node requests
from the source node AN (uv ) to the destination node AN (k v ). during a certain time divided by total resources of substrate
RL (es ) denotes the available bandwidth of a substrate path nodes. Node utilization can be expressed as below:
es ∈ E s , while RL (ms ns ) represents the remaining capacity X
of a physical link. Similarly, Eq. (5) ensures the remaining C(uv )
X uv →ms
bandwidth capacity of the substrate path meeting the requested UN (Gs ) = ( ) ∗ Ti , (12)
bandwidth request of the virtual link, whereas Eq. (6) and (7) C(ms )
ms ∈N s
are available bandwidth capacity of the substrate path es and
where Ti is the time duration of the accepted ith VNR.
the measurement of residual capacity of the substrate link. A
mapping solution is called “feasible” if and only if it meets Link utilization: Similarly, link utilization can be presented
resource constraints (1)-(4) of node mappings and (5)-(7) of as follows:
X
link mappings. A path for a mapped virtual link in the feasible B(uv k v )
solution is therefore called a feasible path. v v s s
( u k →m n s s
X
B. Performance metrics UL (Gs ) = ) ∗ Ti , (13)
The main objectives of VNE problems are associated with B(m n )
ms ns ∈Ls
maximizing the generated revenues of InPs while minimizing
the corresponding mapping cost. In this work, an InP’s revenue Fitness Function (FF) assesses the quality of VNE solutions
v
Γ(Gi ) is defined as the sum of all virtual resources successfully that can be reproduced in next generations. Fitness values
embedded on the underlying SN over time, whereas the serve as rewards to help guide the searching process for the
mapping cost of the ith VNR, denoted as Ξ(Gvi ), is the total optimal solutions. In this paper, FF takes the mapping cost,
network resources required for allocating the ith virtual network hop-count and propagation delay of the mapping solutions
[7]. into consideration and decides which is the best solution for a
Revenue of ith VNR Gvi is formulated as follows: VNR. Due to the coordinated process of our proposed approach,
X X these factors reflect total network resources including both
Γ(Gvi ) = wb ∗ B(uv k v ) + wn ∗ C(uv ) (8) CPU and bandwidth as well as network latency. Additionally,
uv kv ∈Lv
i uv ∈Niv hop-count factor and propagation delay feature allow to
v v
where B(u k ) is the bandwidth demand of virtual link u k v v select neighboring substrate nodes, which possibly reduces the
and C(u ) is the CPU capacity request of virtual node u . resource fragmentation problem, and enhances the possibility
v v

Besides, wb and wn denote the unit weights preferable to of accepting future VNRs. Our fitness function F(Sz ) can be
bandwidth or CPU resources, respectively. expressed as follows:
Cost of ith VNR Gvi can be formulated as below: 1 1
F(S) = ( ) ∗ wc + ( ) ∗ wh
Ξ(Gvi )
X
uv k v HAL (uv kv )
X X X
Ξ(Gvi ) = C(uv ) + fms ns (9)
∀(uv ,kv )∈Lv
uv ∈Niv uv kv ∈Lv s s
i m n ∈L
s i

1
where fm
v
u k v
s ns denotes bandwidth capacity of the physical link
+( X ) ∗ wp
DP (AL (uv k v ))
(m , n ) allocated to the virtual link (uv , k v ).
s s
∀(uv ,kv )∈Lv
The average revenue over cost ratio can be indeed utilized i
(14)
to estimate how efficient a VNE algorithm is. High average
acceptance ratio as well as high average revenue to cost ratio where, S is a feasible solution whereas H and DP define hop-
are highly desired because the results can disclose that substrate count factor and propagation delay of the link mapping of
resources are efficiently allocated. (uv , k v ) respectively. wc , wh , and wp denote the predefined
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 6

TABLE II: List of Notations B. Objectives:

s s s
G (N , L ) Substrate network we propose an objective function as below:
Gvi (Niv , Lvi ) i-th Virtual network v
ms Substrate node, ms ∈ N s
 X
v
X xums 
ms ns Substrate link, (ms , ns ) ∈ Ls Minimize αu C(u ). +
v v s s v v
RN (ms ) + δ
uv Virtual node, uv ∈ Niv u ∈Ni m ∈N ,u ∈Ni
v v
uv k v Virtual link, (uv , kv ) ∈ Lvi  X X u k
ym s ns

loc(ms ) Geographical location of ns βmn B(uv k v ).
RL (ms ns ) + δ
C(m′ ) CPU capacity of m′ , m′ can be ms or uv uv kv ∈Lv
i uv kv ∈Lv s s
i ,m n ∈L
s

B(l′ ) Bandwidth capacity of l′ , l′ can be uv kv or (17)

ms ns C. Constraints
v v
u k
fm s ns bandwidth capacity of ms ns allocated to uv kv X v

D(uv , ms ) Distance between uv and ms ∀uv ∈ Niv , ∀ms ∈ N s , xums .C(uv ) ≤ RN (ms ) (18)
D(uv ) Preferable mapping radius uv
D(uv kv ) Propagation delay or physical distance demand ∀(ms , ns ) ∈ Ls , ms ∈ AN (uv ), ns ∈ AN (k v ), ms ̸= ns ,
Γ(Gvi ) Generated revenue of Gvi
uv kv
X
v v s s
Ξ(Gvi ) Mapping cost of Gvi ym s ns .B(u k ) ≤ RL (m n )
Υ(Gvi ) Revenue to cost ratio of Gvi ∀(uv ,kv )∈Lv
i
Aτc Acceptance ratio during an interval time τ (19)
a
ξ (τ ) Successfully embedded VNRs in time τ ∀(ms , ns ) ∈ Ls , ms ∈ AN (uv ), ns ∈ AN (k v ), ms ̸= ns ,
ξ(τ ) Number of VNRs arrived in time τ X
uv kv s s v v
UN Node utilization of Gs ym s ns .DP (m n ) ≤ D(u k )

UL Link utilization of Gs ∀(uv ,kv )∈Lv

i
AN (uv ) Mapping of the virtual node uv (20)
AL (uv kv ) Mapping of the virtual link uv kv X v
E s (AL (uv kv )) The set of all available paths of AL (uv kv ) ∀uv ∈ Niv : xums = 1 (21)
es A substrate path, es ∈ E s ∀ms ∈N s
RN (ms ) Residual CPU capacity of ms X v
RL (e′ ) Residual bandwidth capacity of e′ , e′ can be ∀ms ∈ N s : xums ≤ 1 (22)
es or (ms , ns ) uv ∈Niv
F(S) Fitness function of a feasible solution S X v v
H Hop-count ∀(uv , k v ) ∈ Lvi : u k
yms ns ≥ 1 (23)
DP Propagation delay ∀(ms ,ns )∈Ls
v
xums Binary node variable
ym
v v
u k
s ns Binary link variable ∀(uv k v ) ∈ Lvi , ∀(ms , ns ) ∈ Ls , ∀ms ∈ N s :
N̄ s (ms ) A set of neighbors of a physical node ms ∈ N s uv kv uv k v uv kv
X
ym s ns − yns ms = xms − xms
(24)
Fiv A set of failed virtual links
Eir Remapping list ∀ns ∈N̄ s (ms )
Q, Q, R Trials in algorithms v
∀uv ∈ N v , ∀ms ∈ N s , xums ∈ {0, 1} (25)
v v
∀(uv , k v ) ∈ Lv , ∀(ms , ns ) ∈ E s (AL (uv k v )), ym
u k
s ns ∈ {0, 1}

weights of cost, hop-count and propagation delay factors, (26)

respectively. List of notations is shown in Table II. Remarks:
• Function (17) is the objective function that tries to
minimize the total cost while balancing the loads. By
IV. M ATHEMATICAL F ORMULATION dividing the remaining resources, the network loads can
be normalized and balanced. It is also aimed at minimizing
This section describes the mathematical formulation to solve
the number of substrate links consumed to the minimum
the online VNE problem with several defined constraints. A
possible. Since the requested demands of the VNR are
virtual node must be embedded on a physical node while a
included in objective function, nodes and links with more
virtual link must be mapped to a substrate path. The following
residual resources that tend to generate less embedding
decision variables indicate node and link mappings.
costs are preferable. δ is a small positive number used to
avoid dividing by zero. αu and βmn are weight factors
A. Variables to determine the significance of the corresponding loads
of either links or nodes, αu > 0 and βmn > 0.
 • Constraints (18), (19) and (20) ensure that the available
1, virtual node uv ∈ Niv is allocated to
v CPU, bandwidth capacity and link delay are not violated
xums = substrate node ms ∈ N s (15)
respectively.
0, otherwise.

• Constraints (21) and (22) guarantee that each virtual
v v v

1, virtual link (u , k ) ∈ Li is assigned node is only assigned to a substrate node and vice versa.
uv kv to a substrate link (m , ns ) ∈ Ls .
s Constraint (23) ensures that each virtual link is assigned
yms ns = (16)

0, otherwise. to a subset of substrate links or a path.
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 7

• Constraint (24) is considered as a flow-conservation, mapping of all virtual links of a VNR where a gene is a
guaranteeing that virtual links are mapped to physical feasible mapping of a single virtual link. Towards the idea
links, where N̄ s (ms ) denotes a set of neighbors of a of joint node and link mappings based on GA algorithm in
physical node ms ∈ N s . this paper, when an online VNR arrives, GA will search for
• Finally, constraints (25) and (26) denote binary domain an optimal embedding solution including both node and link
v
uv kv
constraints on variables xums and ym s ns , respectively. mappings. Population initialization randomly generates node
chromosomes and then associated link chromosomes through
V. J OINT NODE - LINK M APPING
VNoM and VLiM. During this process, heuristic conciliation
A. Background of metaheuristic algorithms strategy is deployed to remap virtual nodes that cause the
Metaheuristics comprise a set of optimization techniques failures of VLiM. Then, selection selects two random node
aimed at achieving the global optimum by discovering the chromosomes to become parents which are utilized in evolution
search space effectively. Intrinsically, metaheuristics exploit processes. These parents will swap their genes starting from a
diversified variation operations to explore new potentials sys- random crossover point to generate new children. Eventually,
tematically, and then their multi-objective fitness function will one of the children is selected for mutation process and one
drive such achieved potentials to the optimum. Approaching random gene of this child is then replaced by a new random
an effective algorithm satisfying several stringent constraints one to produce a new child. The population will update those
becomes challenging. In fact, metaheuristic methods have been children while maintaining elite chromosomes by substituting
successfully employed in various applications from different worst existing chromosomes based on fitness values. Evolution
fields, including operation research, industrial engineering to is continued with new selection and this procedure will run for
management science. Evolutionary Computation (EC) that several iterations in aimed at obtaining the optimal mapping
imitates the natural evolution process consists of a set of solution.
metaheuristic algorithms for seeking the global optimization. Lately, parallel computing is a promising paradigm to
Genetic Algorithm, a mature metaheuristic motivated by the efficiently deal with the complicated issues with considerable
Darwin evolution principle through natural selection, is one of time saving and low cost guarantees thanks to concurrent-
the most popular population-based metaheuristic methods in enabled deployment. In practise, GA can be fundamentally
EC. It is suitable for solving multi-objective linear or non-linear recognized as a parallel searching mechanism [35] with
programming optimization problems due to its simplicity and non-mutual dependency amongst several exclusively feasible
ease of implementation. GA is fast and more efficient than solutions. To resolve the aforementioned problem of GA
many heuristic methods by maintaining a balance between algorithm, we propose a distributed parallel implementation
exploration and exploitation [21]–[26]. based on GA algorithm in next section.
Furthermore, GA-based approaches bear a resemblance to
reinforcement learning (RL) algorithms in the aspect that they
both interact with environments through an iterative action-
reward process. However, to our best knowledge, modern RL
algorithms are focused on integrating deep learning techniques.
The environments are typically modelled as a Markov Decision
Process with transition probabilities as parameters to be
learned through deep learning techniques. Those parameters
are assumed static so that the learning process can estimate
them and used the model with these parameters to decide future
actions. Our GA-based approach does not need to estimate
those parameters. Fundamentally, upon the arrival of a VNR,
it does a smart search to find the best result it can. The only
assumption it uses is that children are likely better than their
parents. This makes it adaptable to the dynamic situations where
loads keep changing. Due to similarities, GA is also advised as
a scalable alternative to the cutting-edge RL algorithms [33].
Even GA does not always outperform RL algorithms, but GA
is considerably faster than RL since it exposed great scalability
and parallel capabilities [33].
A typical GA algorithm comprises four operations: initializa-
tion, selection, crossover, and mutation [34]. In this paper, we
define a chromosome as a feasible mapping of a VNR. Each Fig. 3: Parallel operation scheme
chromosome includes a node chromosome and an associated
link chromosome. A population is an elite set of up to M B. Distributed parallel GA-based algorithms
chromosomes. A node chromosome is a feasible mapping of Distributed and parallel computing has been practically
all virtual nodes of a VNR where a gene is a feasible mapping proven as an efficient paradigm to deal with enormous
of a single virtual node. A link chromosome is a feasible and complicated problems with guaranteed costs and less
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 8

operational time by concurrency support. Our proposed parallel

GA-based scheme in this paper is presented in Fig 3. Every
working machine in this scheme independently carries out a
GA algorithm to approach as many feasible solutions of a Algorithm 2 Genetic Algorithm with Conciliation Strategy at
VNR as it can with a specific number of iterations. The best- each parallel machine
matching result is determined amongst those distributed and
1: Input:
parallel machines. It is noted that when node mappings are 2: + Online Gvi
changed, the link mappings altered accordingly. Operational 3: + Ops and Ons i
details of our proposed scheme are described in Algorithm 1. 4: Output:
Lines [5-15] express initial path and node pool generations 5: + The mapping solution for Gvi
respectively, while lines [16-20] are the VNE procedures. More 6: procedure G ENETIC A LGORITHM OPERATIONS
▷ Generate initial population
details about GA algorithm can be found in Algorithm 2 where 7: p =0
lines [7-22] are associated with the initial population generation 8: for m = 1 to M do
and lines [23-46] include the evolution processes. ▷ Generate a node chromosome with |Niv | genes
9: for n = 1 to |Niv | do
Algorithm 1 Distributed parallel GA-based algorithm ▷ Try to map virtual node n to a randomly selected
node in Ons i with up to Q trials
1: Input: 10: for q = 1 to Q do
2: + Online Gvi 11: Map virtual node n to a randomly selected node
3: Output: in Ons i
4: + Near-optimal mappings for Gvi 12: if feasible goto 15
13: end for
5: function I NITIAL PATH POOL GENERATION 14: goto 22
6: Construct Ops using k-shortest path algorithm based on 15: end for
hop-count ▷ Generate a link chromosome with |Lvi | genes
7: end function 16: for l = 1 to |Lvi | do
▷ Create an initial node pool Ons i for Gvi 17: Try to map virtual link l to a selected path in Ops
▷ GA-RAN, GA-SEQ, GA-STP are different in path
8: function I NITIAL NODE POOL GENERATION selection
9: nvi m = argmin c(nvi ) 18: end for
10: for each ns ∈ N s do 19: if some link mappings are infeasible call Algorithm 3
11: if RN (ns ) ≥ c(nvi m ) add ns to Ons i . 20: if Algorithm 3 returns false goto 22
12: end for 21: p = p + 1 and add chromosome to population
22: end for
13: if |Ons i | ≥ |Niv | goto VIRTUAL NETWORK EMBED - ▷ Evolution process
DING 23: if p > 1 then
14: else reject Gvi 24: for e = 1 to maxIterations do
15: end function 25: if ranN um ∈ (0, 1) < pc then
16: procedure V IRTUAL NETWORK EMBEDDING 26: Randomly select two parents
27: Conduct crossover operation
▷ Implement genetic algorithm in distributed parallel 28: if both children are feasible then
machines 29: One of children is randomly selected for
17: call Algorithm 2 mutation
▷ Synchronize the achieved incumbents among parallel 30: p = p + 1 and add the one produced better
machines fitness values to population
31: else
18: Select the best feasible solution across all machines 32: if only one child is feasible
based on fitness value 33: p = p + 1 and add the one to population
19: if none exist reject Gvi 34: if ranN um ∈ (0, 1) < pm then
20: else accept Gvi and update substrate network resources 35: if none of children in crossover are feasible then
21: end procedure 36: Randomly select one of parents for mutation
37: end if
38: Conduct mutation operation
39: if the mutated child is feasible then
C. Node Mapping Algorithm
40: p = p + 1 and add the one to population
In our proposed joint node-link mapping algorithms where 41: end for
the GA algorithm is the core, the node mapping of a given VNR 42: if p > M eliminate the ones with lower fitness values
simply works in a heuristic way. Population in GA’s operations 43: else if p = 1 then the current mapping will be the final
44: else
consisting of several chromosomes is generated in random 45: reject Gvi for this machine
manner. A VNR includes a set of virtual nodes connected 46: end if
with several virtual links. Accordingly, there are separate node 47: end procedure
mappings and link mappings. The later is dependent on the
former, so if the node mappings are changed, the link mappings
are altered correspondingly. We define a chromosome Cf as
a mapping solution for all virtual nodes in a given VNR.
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 9

Algorithm 3 Heuristic Conciliation Algorithm randomly can prevent the premature convergence problem of
1: Input: GA algorithm.
2: A set of failed virtual links after VLiM Fiv
3: Output: D. Link Mapping Algorithms
4: A feasible chromosome or false
5: procedure H EURISTIC CONCILIATION STRATEGY
When node chromosome is successfully formed, we then
6: Construct a map Mp with each element representing the generate the according link chromosome using different link
number of failed virtual links associated with one virtual mapping algorithms, in which each gene of the link chromo-
node. some is a link mapping of a virtual link in the VNR that
7: Create a map Mn with each element representing the degree is embedded onto a substrate path through VLiM. VLiM
of a virtual node.
▷ Select virtual nodes for remapping can be reduced to unsplittable flow problem which has been
8: Initialize a remapping list Eir widely recognized as N P-hard [3]. To solve VLiM, we build
9: for each failed virtual link (uv , kv ) ∈ Fiv do a path database using the shortest path method (e.g., Dijkstra’s
10: if virtual node uv or kv not in Eir then algorithm) with respect to the hop-count factor, which can be
11: if Mp [uv ] > Mp [kv ] completely constructed prior to the arrival of online VNRs.
12: add uv into Eir
13: else if Mp [uv ] < Mp [kv ] then This database is called initial path pool generation in Fig. 3.
14: add kv into Eir This process is possible since we argue that substrate paths
15: else ▷ Mp [uv ] = Mp [kv ] for each pair of physical nodes in a SN can be determined in
16: if Mn [uv ] > Mn [kv ] then advance due to the fact that topology of a SN is most likely
17: add kv into Eir static. In this paper, we propose three virtual link mappings
18: else
19: add uv into Eir with different searching mechanisms for VLiM stage.
20: end for 1) Random Path Searching Algorithm (GA-RAN): For a
21: return Eir virtual link request in a VNR, the path searching algorithm
▷ Create a new chromosome with at almost R attempts randomly selects a substrate path from the path database based
22: for r = 0 to R do on the information of node mappings. This path must pass a
▷ Generate a new chromosome with Eir
23: for each virtual node nvi ∈ Eir do validity check Eq. (5)-(7) to become a feasible path. If the
▷ Try to map virtual node nvi to a randomly selected path fails to pass the resource checking process, the algorithm
node in for up to Q trials attempts to find another one in the database until at least one
24: for q = 1 to Q do feasible path is achieved. If there is no feasible path found in
25: Map virtual node nvi to a randomly selected node the path database, the algorithm returns failed to the virtual link
in Ons i
26: if feasible break request. In that case, the previous node mappings should be
27: end for reconsidered. Each virtual link in the given VNR is sequentially
28: if none of the trials are feasible goto 36 searching for a feasible substrate path in that way until all
29: end for virtual links are processed.
▷ Generate a link chromosome with associated virtual links 2) Sequential Single Path Searching Algorithm (GA-SEQ):
30: for lr in the set of affected virtual links do
31: Try to map virtual link lr to a selected path in Ops Instead of searching the path in random, the sequential path
▷ GA-RAN, GA-SEQ, GA-STP are different in path searching mechanism searches a potential path for a virtual
selection link request from the first path in the database until the last
32: if any link mapping is infeasible goto 35 one sequentially. Whenever there is a substrate path passing
33: end for the validity check, it becomes feasible, and then the searching
34: return new chromosome goto 37
35: end for algorithm will immediately stop there. In fact, the erected path
36: return false database is relied on the hop-count, so the first path in database
37: end procedure has the lowest hop-count feature. In contrast with GA-RAN,
GA-SEQ goes through the path database by a sequential path
searching. After the whole path database is inspected, if there
is no feasible path achieved, the algorithm returns failed similar
Each gene Φjf is associated with a mapping of a virtual node, to GA-RAN. All virtual links in the VNR are handled in the
where f and j indicate the f th chromosome and j th virtual same way. All link mappings that are specified as “failed” will
node in the VNR respectively. We equally consider all virtual be then reprocessed in the conciliation algorithm. Due to the
node mappings. Therefore, their order in a chromosome can simple and straightforward features of the sequential single
be assigned in an arbitrary manner. Once assigned, all the path searching mechanism, we expect that its execution time of
chromosomes will follow the same order. To enhance the this algorithm is better than those of other proposed algorithms.
capability and reduce execution time, we initially generate a 3) Shortest Path Algorithm (GA-STP): Different from the
feasible node pool where all eligible substrate nodes meeting aforementioned path searching mechanisms, we deploy k-
the least resource demand of the VNR are collected. Each shortest path algorithm as an approximation approach so
virtual node in a VNR is consecutively mapped by a substrate as to optimize link delay and bandwidth consumption of a
node which is randomly selected in the initial node pool, that successfully mapped virtual network. Based on the information
must meet resource requirements (Eq. (1)-(4)) to become a of node mappings, GA-STP maps each virtual link of the VNR
feasible node mapping. We assume that mapping virtual nodes to a single physical path. The shortest path algorithm searches
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 10

k-shortest paths by increasing the value of k until a feasible path links, but h should be chosen instead of k and n because least
that has sufficient bandwidth to embed the corresponding virtual virtual nodes need to be reconsidered. Similarly, h should be
link is found. The time complexity of the algorithm increases selected for remapping in Fig. 4d, even there are three failed
with the value of k. For an efficiency of computation and link mappings including {A - G, A - F and A - D}. If these
network resource usage as well as reducing link delay which all already-mapped virtual nodes or virtual nodes {D, F and
is associated with the distance, the value of k should be kept G} are revisited, we need to remap at least five virtual links.
minimal, so we choose k = 1 in this paper. In fact, we prune Otherwise, when the substrate node h, a node mapping solution
all links that do not have sufficient bandwidth to support the of virtual node A, is reconsidered, the remapped virtual links
virtual link request, and then choose the shortest-distance path could be only three. In Fig. 4e, substrate nodes k/j, l and n
in the remaining graph. In case there is no feasible path found, should be remapped, whereas physical nodes j and i can be
the according virtual link is marked as failed. This link method re-embedded in Fig. 4f due to least remapping.
is able to efficiently balance the network load among the links Inspired by this idea, the heuristic conciliation algorithm
and effectively utilize the network resources. By applying the is applied to handle a set of infeasible link mappings for
conciliation strategy, infeasible link mappings can be revised by generating each chromosome. When we remap virtual nodes
revisiting the current node mappings smartly, which requires to associated with the failed virtual links, we will remap them
automatically explore new link mappings due to the dynamic following an order based on the importance of each virtual
changes of node mappings. It is clearly not subject to the value node. Mp and Mn are designed to evaluate the importance of
of k because the explored new link mappings are associated each virtual node based on the impacts of the virtual node.
with different node mappings rather than the k-shortest paths Mp is a map with each element in the map representing the
associated with the same node mappings. In GA-STP, we set number of failed virtual links associated with one virtual node.
k = 1 to highlight this effect. Mn is the map with each element representing the degree of
We expect that our proposed joint node-link mapping a virtual node. Details of the heuristic conciliation algorithm
assisted by the shortest path mechanism is able to achieve is described in Algorithm 3 where lines [08-21] attempt to
the best performance compared to its counterparts, especially quantify potential nodes for re-mapping.
the average link delay metric. A desirable outcome towards For better understanding, let us take an example in Fig. 4f
the delay metric is beneficial for the virtual network requests for our remapping strategy applied on the SN with the topology
that strictly require sensitive delay requirements (e.g., IoT that is demonstrated in Fig. 2. The VN is shown at right corner
applications). of Fig. 5 while the SN can be abstracted on the left side. We
E. Conciliation Construction have three failed virtual links (F − E, D − C and C − B).
After applying the conciliation strategy, it is merely required
When a node chromosome is determined, various path to remap virtual nodes C and E instead of all virtual nodes in
searching methods as described in V-D are applied to form which old node mappings (C → j, E → l) are now remapped
the according link chromosome. If they successfully find out to new mappings (C → s, E → p). As a result, all virtual links
link mappings for the link chromosome, a feasible solution are mapped successfully.
for the whole VNR has been achieved. However, this is an
ideal scenario. In reality, there might have a situation that link
mapping algorithms cannot find any feasible substrate path for F. Working machine
a virtual link request due to network congestion. Generally, the As illustrated in Fig 3, each paralleled machine is indepen-
pair of virtual nodes that form such virtual link needs to be dently running a GA algorithm which is consisted of four
remapped. In worse cases, several virtual link requests fail to major operations: population initialization, selection, crossover
be embedded, so it is desired to have an efficient mechanism and mutation.
to intelligently remap these requests concerning the minimal
Population Initialization: A single machine starts with a
number of virtual nodes associated with the failed virtual links
population initialization. Denote M as a set of chromosomes
that should be revisited.
forming a population. We have separate node population (27)
Let’s take a real VNR with 7 virtual nodes and 6 virtual links and link population (28). Each node chromosome includes
as demonstrated in Fig. 4 with several practical scenarios for an N = |Niv | genes that represent a potential mapping solution
example. Fig 4a is an accepted solution since all virtual nodes for all virtual node requests in a VNR, where each gene is the
and links have been successfully found their feasible mappings. node mapping of a virtual node. An initial node population
Next figure illustrates a simplest failure with a single failed Pn (MxN size) at each machine is generated as described in
virtual link mapping from A − G. In this case, the substrate Section V-C.
nodes h and n, node mappings of virtual nodes A and G,
Φ1 · · · Φj1 · · · ΦN
 n   1 
respectively, need to be remapped. If n is selected, we would C1 1
remap only one virtual link A − G; otherwise, we have to  C2n  
  Φ12 · · · Φj2 · · · ΦN 
2 
 ..   . . .

remap three virtual links {A − G, A − F, A − D}. Fig. 4c .. .. 
 .    .. . .. . .. 
complicates the same problem a bit with two failed virtual Pn =  n  =  1 N  (27)

j
 
C
 f   f Φ · · · Φ f · · · Φ f 
links {A − G and A − D }. We possibly have three options

 .   . .. . .. .

for remapping including substrate nodes {h, k and n}. Each  ..   .. . .. . .. 
n
selection would give the same number of remapped virtual CM 1 j
ΦM · · · ΦM · · · Φ M N
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 11

virtual node substrate node

G n G n G n
x x
h m l h m l h m l

A F E A F E A F E

k j i k j i x k j i

D C B D C B D C B
(a) (b) (c)

G n G n G n
x x
h m l h m l h m l

A x F E A F x E A F
x E

x k j i k j i k j i
x D
x C
x B
D C B D C B
(d) (e) (f)

Fig. 4: Examples of infeasible virtual link mappings

(31)
n G p 42
o (17) D C (8)
(42) (25)
65 28
(39)
(30)
47 E (5) 15 39

g 35
44 41 46 G A B
b h m l (6)
43 (66) 23 (14)
53 (43) F (33) 40
49 55
(55) 57 24
f F E
23 50 40 Virtual
46 (49)
a k (38) Network (13) (11)
e A D j
(37)
49 C
(45) 51 substrate node
q 52
40 58 36
(37) virtual node
57 r d c s B
Substrate 47 i
(62) (39) (35) (32) (34)
Network
Virtual
nodes

A B C D E F G Remapping A B C D E F G
mappings
Node

a i j k l m n a i s k p m n
Virtual
links

A G A F A D F E D C C B
mappings
Link

a h n a k h m a k m p k s s i

Fig. 5: Examples of Conciliation and remapping

Crossover point
Virtual nodes Virtual links

A B C D E F G A G A F A D F E D C C B
Chromosome 1: Chromosome 1' s link mappings, F(S1)=7.33

a i s k p m n a h n a k h m a k m p k s s i

Chromosome 2: Chromosome 2' s link mappings, F(S2)=6.41

h r q b e d c h a c h b a d h b d e b f q q r

Virtual nodes Virtual links

A B C D E F G A G A F A D F E D C C B
After crossover

Child 1: Child 1' s link mappings , F(S1)=9.82

a i s k e d c a c a d a k d e k s s i

Child 2: Child 2' s link mappings, F(S2)=9.03

h r q b p m n h n h m h b m p b f q q r

Fig. 6: Crossover operation

IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 12

Virtual nodes Mutation point Virtual links

A B C D E F G A G A F A D F E D C C B

Node mappings Link mappings, F(S)=11.65

h r f b p m n h n h m h b m p b f f r

After mutation

Fig. 7: Mutation operation

|Ln |
 

C1e
 Ψ11 ··· Ψj1 ··· Ψ1 i between exploration and exploitation. When the iteration of
|Ln |
C2e Ψ12 Ψj2 generations is terminated, we select the best ones among all
 
   ··· ··· Ψ2 i 
.. .. .. .. generations as our final solution. This also guarantees the
   
   .. .. 
 .   . . . . . 
best ones are selected. We have added Fig. 12 to show the
Pe = 
 Cfe
= |Ln |
 (28)
  Ψ1 ··· Ψjf ··· Ψf i 
convergence process.
  .f
   
 . .. .. .. ..
 ..

  . c c
 . . . . . Φ11 ··· Φj1 Φj1 +1 ··· ΦN
  
C1n
 
 1
e
CM Ψ1M ··· ΨjM ···
|Ln
ΨMi
|
..   .. .. .. .. .. 
. . . . .

 .   
c
  j c +1

Φ1s ··· Φjs ··· ΦN

For each node chromosome, we generate a corresponding link Csn   Φs 

  s 
.. .. ..

chromosome representing a potential mapping for all virtual  ..   .. .. 
 .  
  . . . . .


links in a VNR, where each gene is the link mapping of a Pn =  = c c

ΦN Crn Φ1r ··· Φjr Φjr +1 ···

 r   
virtual link request based on the information of node mappings.  ..  
  .. .. .. .. ..


The generation of link chromosomes can be found in Section

 .   . . . . . 
n   c c 

ΦN CM   Φ1M ··· ΦjM ΦjM+1 ··· 
V-D, while infeasible link mappings produced by this process  M n
CM+1   jc c

ΦN Φ1s Φjr +1

are dealt with by the conciliation strategy as described in rn
 ··· Φs ··· 
CM+2
N c c

Section V-E. When reached M chromosomes for both node Φ s Φ1r ··· Φjr Φjs +1 ···
(29)
and link mappings, our population is “officially” established.
Crossover: In this operation, parental chromosomes are
In fact, we cannot always generate M chromosomes when combined in order to create new offspring for next generations.
the network becomes more and more congested. Therefore, Cs and Cr represent the selected chromosomes where their
the search for M chromosomes will be stopped after a preset indexes within the initial population are s and r respectively.
number of iterations and the remaining GA operations will j c is the crossover point which is randomly chosen between any
be processed with the available chromosomes. We need at positions within N length. We denote C
(M+1) and C(M+2)
least two parents to produce children. If there is only one as new generated chromosomes. In crossover, offspring is
chromosome available after the iteration process, the GA typically produced by exchanging genes between the selected
process will be finished with the chromosome returned as parents beginning from the random crossover point j c+1 to the
the only feasible solution. As mentioned, we are focusing end of chromosomes as demonstrated in (29). At this stage,
on node mappings, whereas link mappings will be changed node mappings have been changed, link mappings are revisited
according to any changes of node mappings. For simplification, by the path searching methods accordingly as described in
we therefore concentrate on generations of node mappings in Section V-D. In this paper, a physical node cannot host two
next GA operations. virtual nodes with the same VNR. If the crossover or mutation
New generations In this paper, we randomly select chro- leads to nodes mapped to the same physical node, we will
mosomes to be parents for generating their children. Selected treat the mapping as infeasible and discard it. Suppose we
parents produce new generations as a result of crossover and select two chromosomes from the population as parents. As
mutation operations, which consequently makes the mapping depicted in Fig. 6, the random crossover point equals to 4, the
solutions evolved after the number of iterations. After those node mappings of virtual nodes E, F and G will be swapped
operations, the major problem is seeking feasible physical paths between two parents in order to generate new offspring. After
regarding to changes of new nodes. Similar to Section V-E, new node mappings are determined, we need to examine their
the path searching methods are implemented to explore link corresponding link mappings. We can recognize that new link
mapping solutions for new generations towards new pairs of mappings are much better than the previous ones, which means
substrate source-destination nodes. If no feasible paths found that our children get evolved.
on new generations, they will be discarded. Mutation: This operation typically adopts a modification
The chromosome population we maintain has up to M on a parent to generate a new offspring. Mutation samples the
chromosomes. Whenever new children are added to the broad solution space and improves the searching efficiency,
population, we eliminate the ones with lowest fitness values to preventing solutions from falling into the local optima. In
make sure that the population has at most M chromosomes. this paper, one of the feasible children created in Crossover is
Therefore, the population itself is an elite set. By randomly randomly selected for mutation, and a random gene of this child
selecting from the population, we attempt to maintain a balance is then replaced by a new gene to produce a new offspring. The
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 13

1.0 0.8
GA-STP GA-STP
GA-SEQ GA-SEQ
GA-RAN GA-RAN
0.9 MCTS
0.7 MCTS
DPGA
NTANRC-S
DPGA
Average acceptance ratio

D-ViNE NTANRC-S
0.8 D-ViNE
0.6

R/C ratio
0.7
0.5
0.6

0.4
0.5
40 50 60 70 80 40 45 50 55 60 65 70 75 80
Traffic load (Erlang) Traffic load (Erlang)

(a) (b)
Fig. 8: (a) Average acceptance Ratio (b) Average revenue to cost ratio

240
95 GA-STP GA-STP
GA-SEQ GA-SEQ
GA-RAN 220 GA-RAN
90 MCTS MCTS
DPGA DPGA
Average generated revenue

NTANRC-S 200 NTANRC-S

D-ViNE D-ViNE
85
Average cost

180
80
160
75
140
70
120
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)
Fig. 9: (a) Average generated revenue (b) Average embedding cost

gene used for replacement must explicitly meet the resource now substituted by (b → f ) and (f → r) respectively. Both
constraints. However, if none of children in Crossover are new children are feasible and better than their parents based
feasible, one of parents is selected for mutation in random as upon improved fitness values shown on the top right of each
illustrated in Algorithm 2. Similar to crossover, path searching subfigures of Fig. 6 and Fig. 7, so they will be updated into
mechanisms are implemented to find new link mappings due to the population.
node mappings changed. If feasible, new generation is updated
intom population. Let denote j m as a random mutation point and In fact, maintaining a proper balancing between exploitation
Φjr′ is new gene that replaces the existing one in C(M+1) . and exploration is essential in GA. Crossover rate pc is in a
In this operation, new embedding solution C ′ (M+1) after range of [0-1] that is normally kept in high values to provide
m
replacement is presented as C ′ (M+1) = [Φ1s · · · Φjr′ · · · ΦN a better exploitation on the current population. Mutation rate
s ].
As illustrated in Crossover, we achieved new children which are pm is usually smaller than the crossover rate. Selecting an
better than their parents. Suppose the second child is randomly appropriate value for the mutation rate is challenging since a
chosen for the mutation, where the mutation point is selected high mutation rate would increase the possibility of exploring
in random, equal to 3 as depicted in Fig. 7. Old node mapping more areas in the searching space, but could prevent GA from
(C → q) is randomly replaced by (C → f ), which results converging to any optimal solution. In contrast, small mutation
in new link mappings for virtual links ({D-C} and {C-B}). rate might cause a problem of premature convergence that traps
Specifically, substrate paths (b → f → q) and (q → r) are the population in local optima. By preferring high efficiency
of GA while keeping a trade-off between exploitation and
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 14

0.50 0.600

0.45 0.575

0.40 0.550
Average Node Utilization

Average link utilization

0.35 0.525

0.30 0.500
0.25 0.475
GA-STP GA-STP
0.20 GA-SEQ 0.450 GA-SEQ
GA-RAN GA-RAN
0.15 MCTS 0.425 MCTS
DPGA DPGA
0.10 NTANRC-S 0.400 NTANRC-S
D-ViNE D-ViNE
0.05
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)
Fig. 10: (a) Average node utilization (b) Average link utilization

24

22

20
Average Delay

18

16 GA-STP
GA-SEQ
14 GA-RAN
MCTS
DPGA
12 NTANRC-S
D-ViNE
10
40 45 50 55 60 65 70 75 80
Traffic load (Erlang)

(a) (b)
Fig. 11: (a) Average delay (b) Execution time between our algorithms
exploration, we set pc = 0.9 and pm = 0.2 in this paper. As
depicted in Fig. 12, our GA algorithms converged smoothly
and stably. Our numerical results have also shown better
performances than other algorithms.

G. Sorting and Terminations

Sorting process selects the best mapping amongst the
achieved feasible solutions determined by their fitness values.
Its outcome is then sent to the synchronization process for a
global ranking. In fact, parallelism is associated with several
concurrent processes where each may finish its assigned task
at different time. Waiting until the last process accomplishing
its job without guaranteeing an expected outcome is indeed
troublesome. A total of 16 working machines were used. For
each working machine, we set M = 10. For each generation
of each working machine, two chromosomes are selected as
Fig. 12: Fitness Function Convergence parents. After the crossover process, two children will be
generated. One of the two will be randomly selected to do the
mutation process. Then the one which is not selected and the
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 15

one which is generated after mutation will be added to the pool TABLE III: Compared Algorithms
with up to M chromosomes. If more than M chromosomes Notation Description
are available, the ones with least fitness values will be deleted GA-STP Joint node-link GA-based algorithm assisted by
to maintain the pool as an elite set. The maximum number of shortest path method
iterations (generations) was set to 10. To keep the trade-off GA-SEQ Joint node-link GA-based algorithm assisted by
sequential single path searching
between operation time and performance, the iteration process GA-RAN Joint node-link GA-based algorithm assisted by
will be terminated earlier if the best fitness value does not random path searching
change for t = 4 consecutive iterations. MCTS Monte Carlo Tree Search with BFS algorithm
for VLiM [28]
H. Synchronization and VN allocation DPGA Greedy node mapping with distributed parallel
In this step, the best VNE solution of the corresponding GA-based algorithm for link mapping [9]
NTANRC-S Network topology attributes and network
VNR is determined by globally ranking the VNE solutions
resource-considered algorithm for node map-
received from worker nodes, based on highest achieved FF ping with the shortest path-based link mapping
values. As a result, if accepted the given VNR is allocated [19]
to the corresponding SN following the information of node D-ViNE Deterministic rounding-based approach obtain-
and link mappings. The last step is updating residual network ing a LP relaxation of MIP for node mapping
resources. with shortest path method for link mapping
[7]
I. Execution time analysis
A. Simulation setup
Due to the lower cost of computing hardware recently, A discrete-event simulator is developed to evaluate the
parallel algorithms can be beneficially exploited to tackle proposed joint node-link mapping solutions with the evaluation
intricate computational tasks. As a result, we propose a parameters following [7]. We extensively generate SNs and
distributed parallel GA framework to deal with the online VNs using the most common GT-ITM topology generator [36].
VNE problem in this paper. We measure the execution time of SNs, adopting Waxman model with α = 0.5 and β = 0.2
the proposed framework on two different modes: sequential and where α and β indicate the maximal edge probability and
parallel. In the former, the time complexity increases linearlyedge length respectively, are consisted of 50 nodes randomly
since the total time equals the sum of the execution time of placed on a 25 × 25 Cartesian plane and 140 edges in average.
each worker machine. In contrast, the latter’s total execution Their capacity including CPU and bandwidth is uniformly
time is counted by the worker machine that accomplishes distributed between 50 and 100 units. Similar to the previous
its assigned task at the latest. Cramer-Chernoff method and research [7], VNRs are configured to follow a Poisson process
Jensen’s inequality can approximate the total execution time dynamically arriving in the network with an average λ rate
in parallel mode as detailed in [27]. Accordingly, the proposedranging from 4 to 8 VNs per 100 time units. Each VNR has
distributed parallel framework can rapidly improve the time an exponentially distributed lifetime with an average value
complexity from linear to logarithmic scale. Interested readersof µ = 1000 time units. The load of VNRs can be indeed
may refer to [27] further theoretical analysis. measured by µλ Erlangs [28]. In addition, each VNR’s node
VI. P ERFORMANCE E VALUATION size is uniformly distributed between 2 and 10. The resource
requirements of VNRs are uniformly distributed between 0
We compare our proposed GA-RAN, GA-SEQ and GA- to 20 for virtual nodes and 0 to 50 for virtual link requests.
STP algorithms with several state-of-the-art VNE competitors: Following [12], we set w = w = 1 in this paper. For
b n
DPGA [9], NTANRC-S [19], D-ViNE [7] and MCTS [28] simulation purposes, we deploy three different SNs and two
on several evaluation metrics comprising average acceptance sets of heterogeneous VNRs, where each includes five subsets
ratio, average revenue and cost, average R/C, average node and of VNRs ranging from 2, 000 up to 4, 000 VNRs. In this paper,
link utilization and finally average delay. These are the most we set out Q = 2 × M, Q = R = 2 × (|F v | + |E r |).
i i
popular and important evaluation criteria for evaluating the The evaluation results were generated by averaging over
efficiency of any VNE algorithm. We select D-ViNE since it is three different SNs and two replications of VNRs with each
the best performance algorithm in [7] which is widely known replication running 50, 000 time units under each traffic load
as one of the most highly-cited papers in VNE. Moreover, and each SN. This setting is the same as the parameters in
NTANRC-S, performed best in [19], is the latest representative [12]. We have run all comparisons in the same setting. All
of the node ranking methods that exploits multiple topology figures are plotted with mean values and 95% confidence
attributes and the global network resources for VNoM in VNE. intervals (CIs). In fact, error bars are too tiny, confirming
Recently, our work in [9] has confirmed that VLiM contributes that our evaluation outputs are extremely reliable. For better
a critical role in approaching an efficient solution for VNE visualization, we plot figures with various colours and different
problem. Finally, MCTS is the reinforcement learning-based markers. Because the 50, 000 time units are 50 times longer
algorithm that deploys Monte Carlo Tree Search to explore than the average lifetime of a VN, it provided roughly 50
the action space [28]. The proposed GA-based algorithms in independent samples per replication per substrate network.
this paper are expected to compete those VNE algorithms in Considering the 2 replications and 3 different SNs, we have
all performance metrics. Our simulation was performed on a about 300 independent samples under each load condition, that
Ubuntu 20.04.2 LTS (Focal Fossa) 64-bit platform with 16 GB is why we can get very good CIs.
memory and Intel® CoreTM i5-6200 CPU @2.30GHzx4.
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 16

B. Discussions on Performance Results metric over the other algorithms thanks to effective path
In this section, we discuss on the simulation results between searching approaches as described in Section V-D. Due to
our proposed VNE algorithms and the compared algorithms. the random searching strategy, GA-RAN could not guarantee
Then, a comprehensive analysis in performance between the the link mapping solutions with short path lengths, so GA-RAN
joint node-link algorithms themselves is also presented. Fig. did not have an advantage on the link utilization as shown in
8 presents the average VNR acceptance ratio and the average Fig. 10b. In addition, GA-STP and GA-SEQ performed better
R/C, while Fig. 9 shows the average generated revenue as average delay than MCTS, the best delay performance among
well as the average embedding cost. They are by some means the compared algorithms, for at least 15.33% up to 29.07% as
sufficient to confirm the efficiency and effectiveness of our depicted in Fig. 11a.
joint node-link VNE solutions. The solid performance of the These dramatic results of our GA-based approaches derive
proposed algorithms is intensified in Fig. 10 where the average from the fact that the proposed algorithms consider the
network utilization including nodes and links is illustrated. Fig. coordination between VNoM and VLiM at the same time. The
11 additionally highlights the advantage of our solutions on novel conciliation mechanism not only handles the remapping
average delay, and a comparison of execution time between the strategy intelligently by minimizing the number of virtual
proposed algorithms running on sequential and parallel modes nodes and virtual links for remapping, but also reduces the
is also demonstrated. possibility of missing out optimal solutions due to remapping
1) Joint node-link algorithms vs State-of-the-art VNE al- all virtual nodes. Moreover, GA algorithm efficiently explored
gorithms: Fig. 8a illustrates the average VNR acceptance the search space to approach optimal VNE solutions driven
ratio which is one of the most important metrics to evaluate by an efficient multi-objective fitness function. Due to less
the embedding capability of different VNE algorithms. It is bandwidth consumed to embed VNRs, abundant residual
observed that the average VNR acceptance ratio of all algorithm bandwidth enables to accept more incoming VNRs verified by
decays with the increasing of network loads due to infinite Fig. 8a. The appealing results of our joint node-link GA-based
substrate resources. Our joint node-link VNE algorithms, algorithms are indeed a real challenge for any VNE algorithm
including GA-STP, GA-SEQ and GA-RAN outperformed we suppose.
all competitors in the average acceptance ratio performance. 2) Amongst Joint node-link algorithms: In this paper,
Specifically, GA-STP, GA-SEQ and GA-RAN achieved better we proposed three joint node-link algorithms according to
average acceptance ratios than MCTS (the best algorithm different path searching methods. They were integrated into
amongst compared algorithms in general) for more than GA algorithm for embedding VNRs using the same fitness
{13.35%, 14.63%, 7.5%} and {16.35%, 14.87%, 8.65%} at 40 function to drive the algorithm to optimal VNE solutions.
and 80 Erlangs respectively. They significantly performed better GA-STP performed best following GA-SEQ and GA-RAN.
than DPGA, NTANRC-S and D-ViNE for more than 7.4% up With low load, GA-SEQ was slightly better than GA-STP
to 24.93% at the highest traffic load. for accepting more VNRs but when the load was increasing,
Furthermore, the proposed GA-based algorithms produced GA-STP became efficient in embedding. This could explain
higher average revenue and lower embedding cost, which leads why GA-SEQ was little better than GA-STP in node utilization
to higher R/C as illustrated in Fig. 9a, 9b and 8b respectively. In in Fig. 10a. Additionally, GA-STP demonstrated its better
details, our best joint node-link VNE solution (GA-STP) gained VLiM performance through improved embedding cost, average
30.23%, 34.98%, 63.23% and 61.56% better average R/Cs than link utilization and average delay by leveraging the shortest
those of aforementioned algorithms at the lowest traffic load. path method. Specifically, GA-STP was better than its direct
Its overwhelming dominance towards this performance metric counterpart, GA-SEQ, in the same aforementioned performance
was retaining over various arrival rates as depicted in Fig 8b. metrics for 5.42%, 3.16% and 9.41%, respectively at the arrival
These results have again confirmed by better average revenue rate of 40. Both tended to be asymptotic when the network
and embedding cost of our approach in Fig. 9a and 9b. became more and more congested.
In fact, higher average acceptance and R/Cs are the most GA-STP, GA-SEQ and GA-RAN are all coordinated ap-
desired target for any VN mapping algorithm, and our proposed proaches. They are different in how the link mapping is done.
approaches have successfully proven their great efficiency Comparing with GA-STP and GA-SEQ, GA-RAN is clearly
and significantly reduced the resource fragmentation problem the worst. GA-RAN selects paths randomly, which likely leads
caused by possibly inefficient embedding. With the increasing to paths with more hops. More hops mean more resource usage
network loads, node and link utilization was expected to at each link that leads to higher link utilization as illustrated in
increase as depicted in Fig. 10a and 10b. Specifically, the node Fig. 10b. This also makes links more congested. GA-RAN tries
utilization of our joint node-link algorithms was remarkably only several randomly selected paths for each link mapping,
higher at least 13.41% up to 77.83% than all rivals over various which lead to longer path and lower acceptance ratio. In our
traffic loads (Fig. 10a). Our algorithms were indeed capable of model, the delay of a path is only dependent on the distance
accepting more VNRs compared to other heuristics, considering of the path, which is the sum of distances of all the links in
multiple crucial embedding factors during mapping. When the the path. Because distances of links are random, in average,
network loads were increasing, joint node-link heuristics still longer paths (or paths with longer delay) also mean paths are
successfully embedded given VNRs by exploiting physical traversing more hops, they are positively correlated.
resource capacity more efficiently. In terms of link utilization, In contrast, GA-STP tries the first shortest path, if it fails,
GA-STP and GA-SEQ exceedingly leaded this performance we will go through the conciliation process, which also tries
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 17

the first shortest path for each new node mapping. Therefore, heuristic VNE algorithms in entirely indispensable evaluation
it always leads to shortest paths and lowest utilization. In GA- metrics we adopted. In fact, current conciliation strategy simply
SEQ, we search feasible link mapping sequentially starting approaches the random mapping instead of attempting all
with the shortest first. If a feasible mapping is found, we possible combinations, and then selecting the best one among
do not need to go through conciliation process. This will them. Hence, there are some potential to further explore this
likely lead to longer path and higher utilization than GA-STP. remapping strategy and we leave this idea for future work.
Because it provides more alternative mappings than GA-STP,
its acceptance ratio will be higher when load is lower as shown ACKNOWLEDGMENT
in Fig. 8a. When load is higher, longer paths tend to make
The authors would like to thank anonymous reviewers for
congestion worse, which leads to lower acceptance ratio.
their careful reading of our manuscript and many insightful
Furthermore, we conducted the joint node-link VNE mapping
comments and suggestions.
algorithms on different sequential and parallel schemes, and
then compared execution time between them to quantify the
time reduction towards parallel operation. The distributed R EFERENCES
parallel paradigm as shown in Fig.3 was deployed for the [1] A. Hakiri and P. Berthou, “Leveraging SDN for the 5g networks: Trends,
parallel operation. Average execution time of the proposed prospects and challenges,” CoRR, vol. abs/1506.02876, 2015. [Online].
Available: http://arxiv.org/abs/1506.02876
algorithms for processing a VNR is shown in Fig. 11b. As [2] I. Ishaq, J. Hoebeke, I. Moerman, and P. Demeester, “Internet of things
such, GA-SEQ was fastest compared to the other algorithms virtual networks: Bringing network virtualization to resource-constrained
in both experiment schemes. Towards sequential operation, devices,” in 2012 IEEE International Conference on Green Computing
and Communications, Nov 2012, pp. 293–300.
GA-SEQ was faster than GA-STP and GA-RAN for more than [3] A. Fischer, J. F. Botero, M. T. Beck, H. de Meer, and X. Hesselbach,
26% and 28% respectively. In this scheme, GA-RAN that is “Virtual network embedding: A survey,” IEEE Communications Surveys
based on random mechanism needed more time to seek for Tutorials, vol. 15, no. 4, pp. 1888–1906, 2013.
[4] J. Gil Herrera and J. F. Botero, “Resource allocation in nfv: A
feasible solutions while GA-SEQ due to its simplicity took comprehensive survey,” IEEE Transactions on Network and Service
least time to finish processing a VNR. On the other hand, Management, vol. 13, no. 3, pp. 518–532, 2016.
parallel scheme reveals that shortest path method and random [5] T. A. Q. Pham, J.-M. Sanner, C. Morin, and Y. Hadjadj-Aoul, “Virtual
network function–forwarding graph embedding: A genetic algorithm
mechanism most likely achieved the same execution time, approach,” International Journal of Communication Systems, vol. 33,
which means that GA-RAN exploited the distributed machines no. 10, p. e4098, 2020, e4098 0.1002/dac.4098. [Online]. Available:
approaching VNE solutions efficiently with the random manner. https://onlinelibrary.wiley.com/doi/abs/10.1002/dac.4098
[6] B. Addis, G. Carello, and M. Gao, “On a virtual network functions
GA-SEQ is still a winner since it was faster than the other placement and routing problem: Some properties and a comparison of
counterparts for more than 25%. Although, GA-STP performed two formulations,” Networks, vol. 75, no. 2, pp. 158–182, 2020. [Online].
best within the set of proposed joint node-link VNE algorithms Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/net.21915
[7] M. Chowdhury, M. R. Rahman, and R. Boutaba, “Vineyard: Virtual
in terms of performance, if we need a more rapid mechanism, network embedding algorithms with coordinated node and link mapping,”
an alternative choice could be GA-SEQ. On the convergence IEEE/ACM Transactions on Networking, vol. 20, no. 1, pp. 206–219,
of GA, the tendency of fitness values of GA-based algorithms Feb 2012.
in this paper is depicted in Figure 12 with 95% CIs. Due to an [8] C.-W. Huang, C.-A. Shen, C.-Y. Huang, T.-L. Chin, and S.-H. Shen, “An
efficient joint node and link mapping approach based on genetic algorithm
appropriate approach driven by an efficient fitness function, our for network virtualization,” in 2019 IEEE 90th Vehicular Technology
proposed solutions are rapidly converged after few iterations. Conference (VTC2019-Fall), 2019, pp. 1–5.
[9] K. T. Nguyen, Q. Lu, and C. Huang, “Rethinking virtual link mapping
VII. C ONCLUSION in network virtualization,” in 2020 IEEE 92nd Vehicular Technology
Conference (VTC2020-Fall), 2020, pp. 1–5.
Network virtualization is a primary enabler for the antici- [10] Hong-Kun Zheng, J. Li, Y. Gong, W. Chen, Zhiwen Yu, Z. Zhan, and
pated success of future networks (e.g., virtualized 5G, smart Ying Lin, “Link mapping-oriented ant colony system for virtual network
embedding,” in 2017 IEEE Congress on Evolutionary Computation
IoT networks); thus, efficient VNE algorithms are eminently (CEC), June 2017, pp. 1223–1230.
desirable. VNE has been well studied in last decade, but [11] G. S. Paschos, M. A. Abdullah, and S. Vassilaras, “Network slicing
there are very few papers dealing with online VNE problem with splittable flows is hard,” in 2018 IEEE 29th Annual International
Symposium on Personal, Indoor and Mobile Radio Communications
using heuristic or metaheuristic algorithms in a joint manner. (PIMRC), Sep. 2018, pp. 1788–1793.
In this article, we propose joint node and link embedding [12] N. M. K. Chowdhury and R. Boutaba, “A survey of network
approaches relied on GA algorithm for simultaneously solving virtualization,” Computer Networks, vol. 54, no. 5, pp. 862 – 876,
2010. [Online]. Available: http://www.sciencedirect.com/science/article/
virtual node and link mappings. When the node mappings are pii/S1389128609003387
changed, the link mappings are accordingly altered. A heuristic [13] H. Cao, H. Hu, Z. Qu, and L. Yang, “Heuristic solutions of virtual
conciliation mechanism is then used to tackle infeasible link network embedding: A survey,” China Communications, vol. 15, no. 3,
pp. 186–219, March 2018, doi: 10.1109/CC.2018.8332001.
mappings due to inappropriate virtual node embedding in GA [14] H. Cao, S. Wu, Y. Hu, Y. Liu, and L. Yang, “A survey of embedding
operations. Furthermore, we deploy our proposed algorithms algorithm for virtual network embedding,” China Communications,
on a distributed parallel operation scheme in order to reduce vol. 16, no. 12, pp. 1–33, Dec. 2019, doi: 10.23919/JCC.2019.12.001.
[15] M. Melo, S. Sargento, U. Killat, A. Timm-Giel, and J. Carapinha,
time execution and then we make a time comparison of the “Optimal virtual network embedding: Node-link formulation,” IEEE
proposed VNE solutions running on sequential and parallel Transactions on Network and Service Management, vol. 10, no. 4, pp.
operations. The extensive evaluation indicates that our joint 356–368, 2013, doi: 10.1109/TNSM.2013.092813.130397.
[16] C. Huang and J. Zhu, “Modeling service applications for optimal parallel
node-link combination in a single VNE mapping stage based embedding,” IEEE Transactions on Cloud Computing, vol. 6, no. 4, pp.
on GA-based approaches outperforms most of state-of-the-art 1067–1079, Oct 2018.
IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT 18

[17] X. Cheng, S. Su, Z. Zhang, H. Wang, F. Yang, Y. Luo, and J. Wang, Khoa Nguyen received his M.Sc. degree in Telecom-
“Virtual network embedding through topology-aware node ranking,” munications Engineering from the University of
SIGCOMM Comput. Commun. Rev., vol. 41, no. 2, p. 38–47, Apr. 2011. Sunderland, England, in 2013 and the Ph.D. de-
[Online]. Available: https://doi.org/10.1145/1971162.1971168 gree in Electrical and Computer Engineering at the
[18] P. Zhang, H. Yao, and Y. Liu, “Virtual network embedding based on the Department of Systems and Computer Engineering,
degree and clustering coefficient information,” IEEE Access, vol. 4, pp. Carleton University, Canada, in 2021, respectively.
8572–8580, 2016. His main research interests include communication
[19] H. Cao, L. Yang, and H. Zhu, “Novel node-ranking approach and multiple networks, cloud/edge computing, parked vehicle
topology attributes-based embedding algorithm for single-domain virtual edge computing (PVEC), Internet of Vehicles (IoV),
network embedding,” IEEE Internet of Things Journal, vol. 5, no. 1, pp. software-defined networks (SDN), network function
108–120, Feb 2018. virtualization (NFV), containerization technologies
[20] M. Feng, J. Liao, J. Wang, S. Qing, and Q. Qi, “Topology-aware virtual and machine learning.
network embedding based on multiple characteristics,” in 2014 IEEE
International Conference on Communications (ICC), 2014, pp. 2956–
2962.
[21] X. Mi, X. Chang, J. Liu, L. Sun, and B. Xing, “Embedding virtual
infrastructure based on genetic algorithm,” in 2012 13th International
Conference on Parallel and Distributed Computing, Applications and
Technologies, Dec 2012, pp. 239–244.
[22] Z. Zhou, X. Chang, Y. Yang, and L. Li, “Resource-aware virtual
network parallel embedding based on genetic algorithm,” in 2016
17th International Conference on Parallel and Distributed Computing,
Applications and Technologies (PDCAT), 2016, pp. 81–86.
[23] P. Zhang, H. Yao, M. Li, and Y. Liu, “Virtual network embedding
based on modified genetic algorithm,” Peer-to-Peer Networking and
Applications, vol. 12, no. 2, pp. 481–492, Mar 2019. [Online]. Available:
https://doi.org/10.1007/s12083-017-0609-x
[24] L. Boyang, W. Muqing, and Z. Haosen, “Virtual network embedding Changcheng Huang received his B. Eng. in 1985
based on hybrid adaptive genetic algorithm,” in 2019 IEEE 5th Interna- and M.Eng. in 1988 both in Electronic Engineer-
tional Conference on Computer and Communications (ICCC), 2019, pp. ing from Tsinghua University, Beijing, China. He
1197–1202. received a Ph.D. degree in Electrical Engineering
[25] I. Pathak and D. P. Vidyarthi, “A model for virtual network embedding from Carleton University, Ottawa, Canada in 1997.
across multiple infrastructure providers using genetic algorithm,” Science From 1996 to 1998, he worked for Nortel Networks,
China Information Sciences, vol. 60, no. 4, p. 040308, Mar 2017. Ottawa, Canada where he was a systems engineering
[Online]. Available: https://doi.org/10.1007/s11432-016-9015-3 specialist. He was a systems engineer and network
[26] P. Zhang, X. Pang, G. Kibalya, N. Kumar, S. He, and B. Zhao, “Gcmd: architect in the Optical Networking Group of Tellabs,
Genetic correlation multi-domain virtual network embedding algorithm,” Illinois, USA during the period of 1998 to 2000.
IEEE Access, vol. 9, pp. 67 167–67 175, 2021. Since July 2000, he has been with the Department
[27] Q. Lu, K. Nguyen, and C. Huang, “Distributed parallel algorithms for of Systems and Computer Engineering at Carleton University, Ottawa, Canada
online virtual network embedding applications,” International Journal of where he is currently a full professor. Dr. Huang won the CFI new opportunity
Communication Systems, vol. n/a, no. n/a, p. e4325, e4325 dac.4325. award for building an optical network laboratory in 2001. He is an associate
[Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/dac. editor of Springer Photonic Network Communications. Dr. Huang is a senior
4325 member of IEEE.
[28] S. Haeri and L. Trajković, “Virtual network embedding via monte carlo
tree search,” IEEE Transactions on Cybernetics, vol. 48, no. 2, pp.
510–521, 2018.
[29] H. Yao, B. Zhang, P. Zhang, S. Wu, C. Jiang, and S. Guo, “Rdam: A
reinforcement learning based dynamic attribute matrix representation for
virtual network embedding,” IEEE Transactions on Emerging Topics in
Computing, pp. 1–1, 2018, doi: 10.1109/TETC.2018.2871549.
[30] P. Zhang, C. Wang, G. S. Aujla, and X. Pang, “A node probability-
based reinforcement learning framework for virtual network embedding,”
in 2020 IEEE 21st International Symposium on ”A World of Wireless,
Mobile and Multimedia Networks” (WoWMoM), Cork, Ireland, 2020, pp.
421-426, doi: 10.1109/WoWMoM49955.2020.00077.
[31] P. Zhang, C. Wang, C. Jiang, and A. Benslimane, “Security-aware virtual
network embedding algorithm based on reinforcement learning,” IEEE
Transactions on Network Science and Engineering, pp. 1–1, 2020, doi:
10.1109/TNSE.2020.2995863.
[32] H. Yao, S. Ma, J. Wang, P. Zhang, C. Jiang, and S. Guo, “A continuous-
decision virtual network embedding scheme relying on reinforcement
learning,” IEEE Transactions on Network and Service Management,
vol. 17, no. 2, pp. 864–875, 2020.
[33] P. S. Felipe, M. Vashisht, C. Edoardo, L. Joel, O. S. Kenneth, and
C. Jeff, “Deep neuroevolution: Genetic algorithms are a competitive
alternative for training deep neural networks for reinforcement learning,”
in arXiv:1712.06567, 2018, pp. 1–16.
[34] M. Mitchell, An Introduction to Genetic Algorithms. Cambridge, MA,
USA: MIT Press, 1998.
[35] H. Mühlenbein, “Parallel genetic algorithms in combinatorial
optimization,” in Computer Science and Operations Research,
O. BALCI, R. SHARDA, and S. A. ZENIOS, Eds. Amsterdam:
Pergamon, 1992, pp. 441 – 453. [Online]. Available: http://www.
sciencedirect.com/science/article/pii/B9780080408064500344
[36] E. W. Zegura, K. L. Calvert, and S. Bhattacharjee, “How to model an
internetwork,” in Proceedings of IEEE INFOCOM ’96. Conference on
Computer Communications, vol. 2, March 1996, pp. 594–602 vol.2.