Deep Reinforcement Learning for VNF placement

and chaining of Cloud Network services

Wided Khemili1,2 , Jalel Eddine Hajlaoui1 , Mohand Yazid Saidi2 , Mohamed Nazih
Omri3 , and Ken CHEN2
1
University of Sousse, Street of Khalifa Karwi, Sahloul 4, 4002, Sousse, Tunisia
2
L2TI - Institut Galilée, Université Sorbonne Paris Nord, 93430 Villetaneuse, France
3
MARS Laboratory, University of Sousse, Tunisia
[email protected],hajlaouijalel [email protected],[email protected]
[email protected],[email protected]

Abstract. With Network Function Virtualization (NFV), network services can

be swiftly and efficiently constructed by instantiating Virtual Network Func-
tions (VNFs) on host servers. In order to optimize resource utilization (such as
bandwidth) and meet Service-Level Agreements (end-to-end delay for instance),
efficient solutions are required for VNF placement, consolidation, and chaining.
To tackle these challenges, we propose two approaches: (1) a novel Formal Con-
cept Analysis (FCA)-based method focused on placing VNFs on appropriate
virtual machines, and (2) a Deep Reinforcement Learning-based approach that
integrates two parallel modules for VNF placement and chaining: Markov De-
cision Process (MDP) and Long Short-Term Memory (LSTM).
The parallel operation of these modules enables the extraction and capture of
the current NFV environment and historical transitions. In our proposal, we
employ Policy Gradient for agent training, aiming to identify suitable hosts for
each VNF in the Service Function Chain (SFC), thereby enhancing various qual-
ity metrics such as latency, resource cost, and throughput. Simulation results
show the effectiveness of our approach, achieving a 47% improvement in rewards
compared to the deep VNF approach and a 52% improvement compared to the
First Fit approach.

Keywords: Network Function Virtualization, VNF Placement and Chaining,

VNF Consolidation, resource allocation, Service-Level Agreement.

1 Introduction
Cloud computing represented a big leap in technological evolution. Where virtual-
ization is considered the main driver of cloud computing, which serves to separate
physical resources to meet the demands of a gigantic data center. Network Function
Virtualization (NFV) creates virtual copies of network services (Firewall, router, Deep
Packet Inspection (DPI)) and run them on virtual machines. In addition, NFV works
to facilitate the execution and distribution of virtualized network functions (VNFs) on
different servers or move and migrate them dynamically from server to server according
to request, ie anywhere on the network. In another sense, it allows to create hybrid con-
structor where the network functions and resources can coexist and consolidate from
one host machine to another as needed. This flexibility and dynamism contributes to
the possibility of controlling and managing virtual resources easily to optimize cost
and energy consumption. These features provide a suitable environment to exploit and
apply placement and consolidation algorithms which is mainly a migration [2] of vir-
tualized network functions or resources from one host or source machine to another
according to their requirements and capacity. Indeed, the emergence of new technologies
2 Authors Suppressed Due to Excessive Length

that require a very high network throughput, reduced delay and latency (particularly
for real-time applications like Voice over IP, telemedicine, etc.) presents a new chal-
lenge and requires the finding of suitable solutions and harmonics with the problem
of consolidation and placement of VNF. To confront this challenge, several placement
and consolidation approaches based on different resources. But, most of this research
has focused on gain and minimizing cost before quality of service [15][8]. Despite the
numerous lines of research dedicated to addressing the consolidation and placement
of Virtual Network Functions (VNFs), certain issues remain either inadequately ad-
dressed or overlooked. One such overlooked aspect is the VNF chaining problem, which
intricately ties in with the challenges of VNF placement and consolidation.
These observations highlight the need for more comprehensive research endeavors
that are cognizant of the intricacies of this problem. Further exploration is warranted,
marked by increased experience and heightened awareness to comprehensively address
the multifaceted nature of the VNF consolidation and placement challenges.
As the Service Function Chain (SFC) structure consists of a set of VNFs linked in a
chain according to a predefined order, it poses another problem of chaining in parallel
with placement and consolidation. This chaining imposes new rules for the placement
of VNFs which require compliance with a predefined order of a set of VNFs to deploy
the expected network service and to respond to an SFC request.

In this context, the paper’s remaining sections are organized as follows: In section II,
we present significant works related to the VNF placement and chaining problem. We
define and formulate the VNF placement problem in section III. Then, we formulate
and present the chaining issue as extension of VNF placement problem in section
IV. Section V describes FCA methodology and deep learning-based solution we’ve
suggested and its strategy. Section VI illustrates and supports the effectiveness of our
FCA and deep learning based technique compared to the current approaches. Section
VI serves as the paper’s conclusion.

2 Related work

In fact, the growing demand on telecommunication networks represent the first reason
for the advent of virtualization. NFV creates virtual copies of network services and
run them on virtual machines. In addition, NFV works to facilitate the execution and
distribution of virtualized network functions on different servers or move and migrate
them dynamically from server to server according to request [17]. This characteris-
tics provides a suitable environment of the consolidation which contributes to reduce
the number of active physical devices and therefore the number of virtual devices, to
improve the exploitation of the resources and to minimize the energy consumed. [20]
tackles the problem of consolidation under interference constraint. We notice that the
dimensions or the members of consolidation are different from one approach to an-
other. In [16], the consolidation is performed between the VM and the server, [13] [21]
between VM and tasks, [14] between VNFs and servers, [7] between VNFs and VMs ,
in [5] between the Software Licenses. The proposed approaches are also characterized
by the use of different technologies for example (VCMM) [14], heuristics [3] [4], genetic
algorithms [6] [18] , “Gossiping” [9], etc. Some approaches indicate the use and exploit-
ing of resources by tackling only one or some resources. However, some works do not
consider the importance of this factor which also requires a large amount of energy. As
an extension of the VNF placement problem, it cannot solve the VNF placement prob-
lem without taking into account the chaining characteristics of the VNFs. A flow of
VNFs is routed in chains according to a predefined order, which constitutes a network
 Title Suppressed Due to Excessive Length 3

service chain called SFC [11] [12]. Furthermore, the flow on the router and switch is a
challenge at the level of the control unit (SDN) [14] [10]. Thus, the growing traffic re-
quires a high-performance routing strategy to minimize the routing computation time.
Lately, a multiplicity of academic research that has addressed the problem of VNF
selection and chaining. In [21], Coa et al developed a log-competitive online COATS
algorithm to direct traffic in an SDN network. This algorithm aims to control the traffic
by considering the time of arrival and departure of traffic. However, it does not address
the bandwidth availability issue considering the VNF instance multiplicity. In [19] a
new strategy based on deep learning has been proposed to solve the VNF selection
and chaining problem to minimize the end-to-end delay and get the best SFC routing
path. Moreover, these two studies [19] [21] focus on the role of SDN to control traffic
in multi-instance VNF network. [1] focused on improving shared resource utilization
of edge servers and physical links within latency bounds. The variety of this research
proves that the VNF placement and chaining problem is a complex optimization prob-
lem that requires in-depth study.

3 Formulation and description of VNF placement and

chaining problem

3.1 Description

To understand the issue of placement and chaining of VNF in SFC of requested service,
we divide it into two unseparated problems, one is the extension of the other. The first
problem consists in the placement and consolidation of VNF in the most appropriate
VM to minimize the number of active virtual machines and to put the machines which
are emptied in standby in order to minimize the energy consumed at the cloud data
centers and to minimize the latency and the cost of resources. The second problem
consists in placing the VNFs in a chain according to a predefined order to meet the
requested service requirements (SFC). The traffic and direction of the VNFs is deter-
mined by taking into consideration the multiplicity of VNF instances, the congestion
at the link level (according to the bandwidth) and at the VM node level (according to
the memory and storage capacity). Our SFC is involved four VNFs used for network
security which are: Firewall(FW) which monitors traffic to ensure network security,
deep packet inspection(DPI) which is an application used to process the packet deep
in the load of IP data, intrusion prevention system (IPS) is a framework that provides
a security system checking errors, and deception system (DS) is a system that defends
hacker attacks based on the identification of the signature of hackers.

3.2 Problem formulation

A.Placement: To formulate the VNF placement problem, suppose our system is a

context K = (V M, V N F, I). V M = {vm1 , vm2 , ....vmi } is the set of objects, V N F =
{vnf 1, vnf 2, ...vnfj }, and I ∈ (V M, V N F ) is the binary relationship between VM
objects and VNF attributes. A concept of context (VM,VNF, I) is a couple of object
and attribute (vm1,vnf1) where, vm1 ⊆ V M and vnf 1 ⊆ V N F which verifies the
application vm′ = vnf and vnf ′ = vm. For a concept (vm, vnf ), vm is said to be its
extension and vnf its intention. Let app be the function of placing a VNF partition in
virtual machines with VM × VNF → app. We define VNF placement problem formally
by the next triplet:

P P rV N F ≡≺ V M, V N F, app ≻ (1)
4 Authors Suppressed Due to Excessive Length

The app application represents the operation of placing a set of VNFs into a set of
VMs. Each vmi requires a vnfj to run. After the identification and formulation of the
problem, we will apply the approach based on the FCA methodology to extract the
concept (vmi , vnfj ) which indicates the best placement of VNF in VM.
B.Chaining: After solving the VNF placement problem, we extract the candidate
VNFs to use them to deploy the requested network service according to a predefined
order. We formulate the chaining problem as follows.
Let be a model of an undirect network graph G = (V, E), V is the node set V M , E is the
edge set. Each V Mi ∈ V is equipped with R = 1, 2, 3 resources (memory, storage, cpu).
F represents the set of virtualized network functions (VNFs) where F = f 1, f 2, fi ..., ft
which is characterized by latency di,j between node i and j, where E ∈ (i, j). Let S be
the service function chain (SFC) which contains a series of VNFs fi which are placed
according to a predefined order, where f is the it h VNF in S and t is the total number
of VNFs in S. Service function chain request S is defined by R, this request has a
source node v and destination node v. We consider Nf as the set of candidate VM
nodes for hosting VNF. The binary value xi,j shows if the edge e on the path from
ingress node v to egress node v. Also, the binary value y identifies if VNF f is mapped
to node v.

4 FCA based approach

4.1 Placement and consolidation strategy

Generally, the strategy proposed is based on three main steps of FCA: (i) the first step
consists on the organization of the VNF according to a hierarchy, described by a lattice
of concepts, in order to group the VNF and the dependent NS in the same concept. (ii)
the second step is the selection of the candidate concepts (cc.int) of FCA and to attach
them and use as VNF to the second process of PBDRL and (iii) the third step is to
construct the lattice of Galois, where the circle represents the concept, and the arcs
between the rectangles express the relationship from the more general (above) to the
more specific (below). The Galois lattice aims to eliminate unnecessary and redundant
concepts without losing information.

4.2 FCA (Formal concept analysis)

We apply FCA to place VNFs in the minimum and most suitable VM. We use the con-
text K = (V M, V N F, I) which is defined in section(3.2) according to Table 1. Given
a concept C = (V N F, V M ) of context K, we define the placement problem by eq(1)
and solve the rest of the placement problem as follows.
1. Determine the dependency between two VNFs jointly dependent on one or more
VMs:

V NF
Dependencyvnfi ,vnf j
= Count(vmvnfi ∩ vmvnfj ) (2)
This dependency is explained by the number of network service VM that use both vnfi
and vnfj . Conversely, the dependency between two VM jointly dependent on one or
more VNF which can formulate as follows:

VM
Dependencyvmi ,vmj
= Count(vnfvmi ∩ vnfvmj ) (3)
To extract the candidate concept, we first define the concept weight and then the
maximum coverage.Let ci be any concept, where ci.Int is the intention of the concept
 Title Suppressed Due to Excessive Length 5

ci . The weight of a concept ci , denoted by W (ci ), is written as follows:

|ci .Int|
W (ci ) = (4)
|vnfi |
And the maximum coverage that covers the set of V M objects denoted by MC is
defined as follows: 
1 if ∪i∈I ci.Int=1
MC = (5)
0 else
Let C = cc1 , ..., ccp be a set of concept. We define the candidate concept, that has a
maximum weight and cover all VNF, noted by Cand(CC), as follows:
 P
1 if k=0 W (ccp ) = 1andM C(cc) = 1
Cand(cc) = (6)
0 else

Table 1: table of context

O/A vnf1 vnf2 vnf3 vnf4
vm1 X X
vm2 X X
vm3 X X
vm4 X

5 Deep reinforcement based approach: PBDRL

In this section, we describe the second proposed deep reinforcement learning approach
which is based on two parallel modules LSTM and MDP. This approach is presented
as an extension of the FCA approach using candidate VNFs as the input for the deep
neural network used in the NFV environment. DRL is a combination of Deep Learning
and Reinforcement Learning(RL). This combination makes it possible to adapt the
proposed environment (NFV) of Reinforcement Learning with the characteristics of
neural network to transform the set of service input function chain requested (set of
VNF in predefined order) into a set of outputs for deploy SFC while respecting the
constraints of available capacities and resources. Reinforcement Learning allows control
of inputs through the agent receiving the VNF of SFC in which an agent learns to make
decisions through trial and error. MDP is used to extract motion state from included
NFV environment transitions (state, action, and reward) to be learned. LSTM is used
to extract the motion state of historical transitions. We combine MDP and LSTM to
extract the Snew state. At each timestep, the agent receives a sn ew state, takes an
action a, receives a reward and transitions to next state according to the dynamic
characteristics of the environment. DRL algorithms are based on deep learning for
the high-dimensional MDP problem where π(a/snew ) represents the policy or function
acquired as a neural network.

6 VNF placement and chaining approach for requested SFC

based on different timeslots.
NFV network dynamic transitions and movements are expressed in different time slots.
At each time slot (ie. time step), the agent and the environment exchange information
6 Authors Suppressed Due to Excessive Length

Fig. 1: Deep reinforcement learning as a combination of deep neural network (DNN)

and reinforcement learning (RL).

and learn it. To reduce the discrete action space, the deep reinforcement learning
strategy in our approach is based on serialization and backtracking methods to VNFs
in SFCs. Only one VNF is treated in each state transition. If a VNF cannot be placed
owing to insufficient resources, delay, or bandwidth constraints, the request can not be
fulfilled. Thus, our system backtracks to the previous network state. Note that between
every two-time slots, there are two cases:
(1) Intra time slot: our system takes one VNF of an SFC after another to conserve
the sequence order. MDP state is changed when a VNF is rejected or accepted knowing
that MDP model is formally defined as: ≺ S, A, P, R, γ ≻ where S is the collection of
discrete states, A is the collection of discrete action, P : S × A × S is the distribution of
the transition probability. R : S × A is the reward function and γ ∈ [0, 1]is a discount
factor of future rewards. As shown in Figure 3, when two requests come (such as
SF C1 and SF C2) in the time slot, V N F 1 is processed by the NFV agent which tries
to place it in a node satisfying the bandwidth and resource constraints. An action is
then passed to set the V N F 1 on the appropriate node. No rewards are returned to
our system agent as SF C1 is not fully deployed. Therefore, the system enters state
st+1 , where the reward U (st , a) = 0. In state St+1 , if no action is taken then SF C1
is rejected. This is either because there are no VM node satisfying all the constraints
related to V N F 2. By applying the same method, SF C2 is successfully deployed at
time slot st+3 with a reward U (st+3 , a) that is the throughput of SF C2 minus the cost
of the resource consumption: U (st+3 , a) = SF Ci Bli τ Cost(St+3 , a). To avoid multiple
VNF instances, in order to improve the resource utilization and choose the nearest
node minimizing the traffic and the delay, many SFCs can share a same VNF that is
deployed on the same node. In this case, the same procedures are put in place to place
the VNFs in the chain in the intra and inter slots as shown in Figure 3. (2)Inter time
slot:In the last time slot τk (K ∈ N +), SFC requests stop arriving at the agent. In
this case, the state does not change and no action occurs. For every two time slots, the
system removes the timeout and releases the resources. In the next time slot, a new
request arrives and the agent takes state transition and according to constraints gives
an action or rejected the SFC. Accordingly, a reward U (st , a) is calculated equation
 Title Suppressed Due to Excessive Length 7

(7):

−σt latence(st , a) + wri , si ri is accepted
U (st , a) = (7)
0 ri is rejected or not fully deployed.

Fig. 2: Placement of VNFs for each SFC request at different time slots.

7 Experiments and results analysis

To evaluate our approach of two algorithms FCA and PBDRL, we first prepare the
necessary data for the framework of each algorithm. For FCA, the dataset is made up
of a set of VMs whose number varies from 100 to 1000 and a set of VNFs between 150
and 1100 which are distributed as illustrated in table 2. The number of VMs chosen by
expert is always lower than the number of VNFs. For PBDRL, the dataset is mainly
made up of a set of SFC request inputs and a set of VNFs which are grouped and
distributed into evaluation, training and result datasets. The values of the resources
requested for each VNF in terms of memory, CPU and I/ O are chosen randomly in
the range [1000, 5000]. Whereas, the values of resource capacity in terms of memory
for each VM, CPU, and I/ O are chosen by chance from the range [10, 000, 50, 000]. In
implementation, we applied our solution of placement of VNF by the principle of FCA,
where each VNF is automatically migrated to the most appropriate VM on the basis
of the received stimulus and its internal response. After, the VNF candidate extracted
which will use in the following step.

7.1 Evaluation and control parameters

The different evaluation and control parameters for the results of this work are:
– The first criterion is packaging efficiency which is shown from the quantity of reduc-
tion in number of active virtual machines and reflects on the total energy consumption.
– The second criterion is energy consumption for server and network.
8 Authors Suppressed Due to Excessive Length

– The performance evaluation is based on the comparison of computational time of

different solutions for a set of test problems of different sizes and complexities.
- Number of VM: the number of nodes that directly affects the cost of resources (CPU,
memory), and therefore controls the reward. Obviously, when the number of nodes
decreases, the reward is increases.
- Number of VNFs: the number of VNFs affects the number of trials and mainly the
number of episodes. Indeed, when the number of VNFs increases in each request, the
number of episodes also increases, as well as the requirements of each episode in band-
width, latency, throughput, and resources.
- Resources (CPU, memory): the resources also present themselves as a setting param-
eter and control the reward and the cost of our network.

7.2 Results analysis

From the FCA algorithm database, we were able to extract and determine the evolution
of the average latency and reduction in the number of active virtual machines by
comparing it with the MultiSwarm algorithm. For the FCA hybridization result with
PBDRL approach, we were able to extract the reward results obtained, the average
throughput and the error rate by comparing it with the First Fit and NFV deep
algorithm.
Table. 2 shows the number of active virtual machines (VM) versus the total number
of incoming network services to Cloud for each time period (test). The results clearly
show the effectiveness of the clustering which is directed by FCA to reduce the number
of active virtual machines (Table 2) as well as to minimize the unused resources (UR) as
shown in Table 3. The percentage of unused resources exhibits a slight increase with the
growing number of virtual machines (VMs), impacting the Packaging Efficiency (PE)
rate, as depicted in Table 3. These findings affirm the effectiveness of our approach in
minimizing the number of active virtual machines, thereby demonstrating efficiency in
reducing energy consumption.

Table 2: Number of active VM compare to the total incoming VM number and Unused
resources by our FCA algorithm.
Sets of tests Number of VM Number of active VM
S1 100 60
S2 200 78
S3 300 116
S4 400 212
S5 500 204
S6 600 182
S7 700 322
S8 800 406
S9 900 448
S10 1000 449

Compared to the MultiSwarm algorithm, our approach achieved lower average la-
tency. As illustrated in table 4, the average latency of our approach at level 150 VNF
is 0.016ms while it exceeds 0.050ms for Multiswarm algorithm which confirms the
effectiveness of our approach.
 Title Suppressed Due to Excessive Length 9

Table 3: Packaging efficiency (PE) evolution according to Resource unused(RU) evolu-

tion for our approach FCA.
Sets of tests Number of VM PE% RU%
S1 150 2.5 77
S2 250 3.20 71.5
S3 350 3.01 71.65
S4 450 2.12 80.29
S5 600 2.94 74.23
S6 700 3.84 64.63
S7 800 2.48 77.35
S8 900 2.21 80.004
S9 100 2.23 79.96
S10 1100 2.44 77.45

Table 4: Efficiency of our approach in term of average latency and acceptance rate
Average Our approach First Fit
VNF latency of our Average latency acceptance acceptance
Test number approach of MultiSwarm rate rate

S1 150 0.016 0.052 0.85 0.69

S2 250 0.012 0.05 0.65 0.6
S3 350 0.0089 0.02 0.7 0.58
S4 450 0.0047 0.009 0.6 0.55
S5 600 0.0049 0.0084 0.71 0.53
S6 700 0.0054 0.0081 0.57 0.5
S7 800 0.0031 0.007 0.48 0.48
S8 900 0.0024 0.0066 0.46 0.46
S9 1000 0.0022 0.005 0.44 0.43
S10 1100 0.002 0.0045 0.40 0.41

Table 4 shows the SFC request acceptance rate in different time steps for our
hybrid approach (FCA and PBDRL) compared to First Fit algorithm. So, our approach
ensures a very high acceptance rate compared to the First Fit algorithm which reflects
the efficiency of our approach to pre-estimate for appropriate SFC. Table 5 compares

Table 5: Reward of each request

Reward/Time slot per episode 1000 2000 3000 4000 5000 6000
Our approach 2.92 3.03 3.09 3.16 3.15 3.17
NFV deep 1.4 1.38 1.41 1.37 1.42 1.42
First Fit 1.25 1.27 1.3 1.1.31 1.35 1.39

the reward of our approach with the NFV deep and First Fit approach. This comparison
shows that our approach is 60.5% more effective than the NFV deep approach and 62%
more efficient than the First Fit algorithm. So, for episodes going from 1000 time-slots
10 Authors Suppressed Due to Excessive Length

to 6000-time slots, the reward of our approach increases more quickly compared to
than that obtained with the NFV deep and First Fit approach.

Table 6: Average throughput

Average throughput/Time slot per episode 2048 4096 6144 8192 10240 12288
Our approach 1.55 1.56 1.58 1.59 1.57 1.59
NFV deep 1.4 1.41 1.42 1.41 1.43 1.43
First Fit 1.12 1.15 1.25 1.3 1.36 1.37

As shown in Table 6, the average throughput of our approach is higher than NFV
deep and First Fit at each time slot per episode. So, the average throughput of our
approach reaches 1.55 for 2048-time slots per episode, 1.4 for NFV deep approach, and
1.12 for First Fit approach.

8 Conclusion

In this paper, we first worked to solve the problem of placing VNFs independently in
order to minimize the number of active virtual machines, the cost of resources and
the energy consumed. We have applied the principle of FCA grouping by invoking
the principle of labor division in swarm intelligence in a cloud environment to extract
the candidate VNFs. In order to properly evaluate our approach, we implemented and
simulated our solution using EdgeCloudSim tools. We performed a series of tests to
verify the performance and correspondence of our proposed algorithm with the sought
conditions. The experimental results show that our virtual network function placement
algorithm performed better in terms of our objective compared to the used compar-
ison MultiSwarm algorithm. It guarantees the minimization of the number of active
virtual machines for the allocation of VNF and the quantity of unused resources, gives
high packing efficiency and reliable computation time. In advanced stage, we worked
to solve the problem of placement and chaining of VNFs taking into account the pre-
defined order of VNF in SFC series to deploy the requested network service. In this
work, we studied the consistency of the deep learning reinforcement strategy with the
problem of VNF chaining and the deployment of SFC requests in a cloud network. Our
Parallel bi-state Deep Reinforcement learning approach has achieved better results in
decreasing delay by network throughput and improving the SFC request acceptance
rate. Concisely, our approach consists of problem formulation by FCA and DRL, place-
ment prediction, and deployment of SFC requests by deep reinforcement learning based
on policy gradient (PG) for action decision-making (placement of VNF for an SFC re-
quest) considering the state of the environment and their history which is captured
using MDP and LSTM.

References
1. Abdelhamid Alleg. Service Function Placement and Chaining in Network Function Vir-
tualization Environments. PhD thesis, Bordeaux, 2019.
2. Daewoong Cho, Javid Taheri, Albert Y Zomaya, and Pascal Bouvry. Real-time virtual
network function (vnf) migration toward low network latency in cloud environments. In
2017 IEEE 10th International Conference on Cloud Computing (CLOUD), pages 798–801.
IEEE, 2017.
 Title Suppressed Due to Excessive Length 11

3. Md Hasanul Ferdaus, Manzur Murshed, Rodrigo N Calheiros, and Rajkumar Buyya.

Virtual machine consolidation in cloud data centers using aco metaheuristic. In Euro-Par
2014 Parallel Processing: 20th International Conference, Porto, Portugal, August 25-29,
2014. Proceedings 20, pages 306–317. Springer, 2014.
4. Yongqiang Gao, Haibing Guan, Zhengwei Qi, Yang Hou, and Liang Liu. A multi-objective
ant colony system algorithm for virtual machine placement in cloud computing. Journal
of computer and system sciences, 79(8):1230–1242, 2013.
5. Leila Helali and Mohamed Nazih Omri. A survey of data center consolidation in cloud
computing systems. Computer Science Review, 39:100366, 2021.
6. Christina Terese Joseph, K Chandrasekaran, and Robin Cyriac. A novel family genetic
approach for virtual machine allocation. Procedia Computer Science, 46:558–565, 2015.
7. Wided Khemili, Jalel Eddine Hajlaoui, and Mohamed Nazih Omri. Energy aware fuzzy
approach for placement and consolidation in cloud data centers. Journal of Parallel and
Distributed Computing, 161:130–142, 2022.
8. Shankar Lal, Tarik Taleb, and Ashutosh Dutta. Nfv: Security threats and best practices.
IEEE Communications Magazine, 55(8):211–217, 2017.
9. Moreno Marzolla, Ozalp Babaoglu, and Fabio Panzieri. Server consolidation in clouds
through gossiping. In 2011 IEEE International Symposium on a World of Wireless,
Mobile and Multimedia Networks, pages 1–6. IEEE, 2011.
10. Ilias Mavridis and Helen Karatza. Combining containers and virtual machines to enhance
isolation and extend functionality on cloud computing. Future Generation Computer
Systems, 94:674–696, 2019.
11. Rashid Mijumbi, Joan Serrat, Juan-Luis Gorricho, Niels Bouten, Filip De Turck, and
Raouf Boutaba. Network function virtualization: State-of-the-art and research challenges.
IEEE Communications surveys & tutorials, 18(1):236–262, 2015.
12. Jason Nikolai and Yong Wang. Hypervisor-based cloud intrusion detection system. In
2014 International Conference on Computing, Networking and Communications (ICNC),
pages 989–993. IEEE, 2014.
13. Jianing Pei, Peilin Hong, and Defang Li. Virtual network function selection and chaining
based on deep learning in sdn and nfv-enabled networks. In 2018 IEEE International
Conference on Communications Workshops (ICC Workshops), pages 1–6. IEEE, 2018.
14. Dandan Qi, Subin Shen, and Guanghui Wang. Virtualized network function consolidation
based on multiple status characteristics. IEEE Access, 7:59665–59679, 2019.
15. Mohammad Shojafar, Claudia Canali, Riccardo Lancellotti, and Enzo Baccarelli. Minimiz-
ing computing-plus-communication energy consumptions in virtualized networked data
centers. In 2016 IEEE Symposium on Computers and Communication (ISCC), pages
1137–1144. IEEE, 2016.
16. Oussama Soualah, Marouen Mechtri, Chaima Ghribi, and Djamal Zeghlache. Energy
efficient algorithm for vnf placement and chaining. In 2017 17th IEEE/ACM International
Symposium on Cluster, Cloud and Grid Computing (CCGRID), pages 579–588. IEEE,
2017.
17. Robert Szabo, Mario Kind, Fritz-Joachim Westphal, Hagen Woesner, David Jocha, and
Andras Csaszar. Elastic network functions: opportunities and challenges. IEEE network,
29(3):15–21, 2015.
18. Maolin Tang and Shenchen Pan. A hybrid genetic algorithm for the energy-efficient virtual
machine placement problem in data centers. Neural processing letters, 41:211–221, 2015.
19. Junzo Watada, Arunava Roy, Ruturaj Kadikar, Hoang Pham, and Bing Xu. Emerging
trends, techniques and open issues of containerization: A review. IEEE Access, 7:152443–
152472, 2019.
20. Qixia Zhang, Fangming Liu, and Chaobing Zeng. Adaptive interference-aware vnf place-
ment for service-customized 5g network slices. In IEEE INFOCOM 2019-IEEE Conference
on Computer Communications, pages 2449–2457. IEEE, 2019.
21. Jingsong Zhou, Peilin Hong, Jianing Pei, and Defang Li. Multi-task deep learning based
dynamic service function chains routing in sdn/nfv-enabled networks. In ICC 2019-2019
IEEE International Conference on Communications (ICC), pages 1–6. IEEE, 2019.