Distributed Edge Computing - IoT - PHD - Thesis

Distributed edge computing for enhanced IoT devices
and new generation network efficiency

Mohammed Laroui
To cite this version:

Mohammed Laroui. Distributed edge computing for enhanced IoT devices and new generation network
efficiency. Networking and Internet Architecture [cs.NI]. Université Paris Cité; Université Djillali
Liabès (Sidi Bel-Abbès, Algérie), 2022. English. �NNT : 2022UNIP7078�. �tel-04216998�
HAL Id: tel-04216998

https://theses.hal.science/tel-04216998
Submitted on 25 Sep 2023
HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
Ph.D. Thesis in
Computer Science
By
r
M . Mohammed Laroui
Distributed Edge Computing For

Enhanced IoT Devices and New
Generation Network Efficiency
Thesis Committee :
Pr. Hassine Moungla University of Paris Cite Thesis Director
Dr. Zohra Slama UDL SBA Supervisor
Pr. Hind Castel Telecom SudParis Reviewer
Pr. Ken Chen Sorbonne University Reviewer
Pr. Hossam Afifi Telecom SudParis Examiner
Pr. Badr Benmammar University of Tlemcen Examiner
Dr. Nassim Dennouni University of Chlef Examiner
Year : 2022
Université Paris Cité
En cotutelle avec L’Université Djillali Liabes
de Sidi Bel Abbes
École Doctorale Informatique, Télécommunication et Électronique «Edite de Paris : ED130»
Laboratoire d'Informatique Paris Descartes (LIPADE)
Distributed Edge Computing For Enhanced IoT

Devices and New Generation Network
Efficiency
Par Mohammed Laroui
Thèse de doctorat en Informatique (Réseaux)
Dirigée par Hassine Moungla

Et par Zohra Slama
Présentée et soutenue publiquement le 21 Juillet 2022
Devant un jury composé de :
Hind Castel, Professeur, Telecom SudParis, Rapporteur
Ken Chen, Professeur, Sorbonne Université, Rapporteur
Hossam Afifi, Professeur, Telecom SudParis, Examinateur
Badr Benmammar, Professeur, Université de Tlemcen, Examinateur
Nassim Dennouni, Docteur, Université de Chlef, Examinateur

Acknowledgement
I SINCERELY THANK MY PARENTS , MY BROTHER , MY SISTERS AND

THE REST OF MY FAMILY FOR THEIR HELP AND SUPPORT, UNCONDI -
TIONAL LOVE FOR YOU .
I WOULD LIKE TO THANK THE RESEARCHERS WHO HAVE CONTRIBUTED

TO MY P H .D DURING ALL MY STUDIES YEARS .
F IRST, I WOULD LIKE TO THANK MY SUPERVISORS , THEIR EXPERTISE

AND KNOWLEDGE HAVE GUIDED ME FOR MY RESEARCH .
I N ADDITION , I WOULD LIKE TO THANK THE JURY MEMBERS , REVIEW-

ERS AND EXAMINERS FOR ACCEPTING THE INVITATION AND FOR THEIR
HELPFUL FEEDBACK AND INSIGHTFUL COMMENTS .
F INALY, I WOULD LIKE TO THANK ALL MY FRIENDS AND PERSONS

HELPED ME .
i
List of Abbreviations
Abbreviation Description
IoT Internet of Things.
CC Cloud Computing.
MCC Mobile Cloud Computing.
EC Edge Computing.
MEC Mobile Edge Computing.
VEC Vehicular Edge Computing.
SDN Software-Defined Networking.
NFV Network Functions Virtualization.
VNF Virtual Network Functions.
VNO Virtual Network Operators.
SFC Service Function Chaining.
UAV Unmanned Aerial Vehicle.
QoS Quality of service.
QoE Quality of Experience.
AI Artificial Intelligence.
ML Machine Learning.
DL Deep Learning.
QL Q-Learning .
DRL Deep Reinforcement Learning.
DQN Deep Q-Network.
LSTM Long Short-Term Memory.
GRU Gated Recurrent Units.
SGD Stochastic Gradient Descent.
RFID Radio Frequency Identification.
WSN Wireless Sensor Networks.
SaaS Software as a Service.
PaaS Platform as a Service.
IaaS Infrastructure as a Service.
BTS Base Transceiver Station.
VM Virtual Machine.
gNb GNodeB.
RSU Roadside Unit.
AP Access Point.
GPS Global Positioning System.
RAN Radio Access Network.
DNS Domain Name System.
V2V Vehicle-to-Vehicle.
V2I Vehicle-to-Infrastructure.
ii
V2X Vehicle to Everything.
V2N Vehicle to Network.
V2D Vehicle to Device.
V2G Vehicle to Grid.
V2P Vehicle to Pedestrian(Person).
SCADA Supervisory Control And Data Acquisition .
ETSI European Telecommunications Standards Institute.
ISG Industry Specification Group.
3GPP 3rd Generation Partnership Project .
MEPM Mobile Edge Platform Manager.
MEO Mobile Edge Orchestrator.
AMF Application Rules Management Functions.
BCS Backup Computing Server.
RPM Remote Patient Monitoring.
HDMI High Definition Multimedia Interface.
VR Virtual Reality.
AR Augmented Reality.
MR Mixed Reality.
CPU Central Processing Unit .
GPU Graphics Processing Unit .
RAM Random Access Memory.
KPIs Key Performance Indicators.
ULLC Ultra Low Latency Communication.
CDN Content Delivery Network.
MVNO Mobile Virtual Network Operator.
AMC Adhoc Mobile Cloud.
AME Adhoc Mobile Edge.
VME Virtual Mobile Edge.
VMES Virtual Mobile Edge Server.
SCP Set Cover Problem.
WSC Weighted Set Cover Problem.
MDP Markov Decision Process.
CMDP Constrained Markov Decision Process.
VEnPA Optimal Virtual Edge nodes Placement Algorithm.
SO-VMEC Service Offloading in Virtual Mobile Edge Computing.
OSPV Optimal SFC Placement in IoT-VMEC.
MVEC Mobile Vehicular Edge Computing.
OFMU Optimal Follow Me UAV.
OVEAP Optimal Virtual Edge-Autopilot Placement.
ESPV Efficient SFC Placement Algorithm.
AFMU Follow Me UAV Algorithm.
iii
DVEAP Deep Reinforcement Learning for Edge Autopilot Placement.
LQC Low Quality Cloud.
LQE Low Quality Edge.
MQC Medium Quality Cloud.
MQE Medium Quality Edge.
HQC High Quality Cloud.
HQE High Quality Edge.
Table 1 List of abbreviations.
iv
Abstract
Traditional cloud infrastructure will face a series of challenges due to the centralization of
computing, storage, and networking in a small number of data centers, and the long-distance
between connected devices and remote data centers. To meet this challenge, edge computing
seems to be a promising possibility that provides resources closer to IoT devices.
In the cloud computing model, compute resources and services are often centralized in
large data centers that end-users access from the network. This model has an important
economic value and more efficient resource-sharing capabilities. New forms of end-user
experience such as the Internet of Things require computing resources near to the end-user
devices at the network edge.
To meet this need, edge computing relies on a model in which computing resources are
distributed to the edge of a network as needed, while decentralizing the data processing from
the cloud to the edge as possible. Thus, it is possible to quickly have actionable information
based on data that varies over time.
In this thesis, we propose novel optimization models to optimize the resource utilization
at the network edge for two edge computing research directions, service offloading and
vehicular edge computing. We study different use cases in each research direction. For the
optimal solutions, First, for service offloading we propose optimal algorithms for services
placement at the network edge (Tasks, Virtual Network Functions (VNF), Service Function
Chain (SFC)) by taking into account the computing resources constraints. Moreover, for
vehicular edge computing, we propose exact models related to maximizing the coverage
of vehicles by both Taxis and Unmanned Aerial Vehicle (UAV) for online video streaming
applications. In addition, we propose optimal edge-autopilot VNFs offloading at the network
edge for autonomous driving. The evaluation results show the efficiency of the proposed
algorithms in small-scale networks in terms of time, cost, and resource utilization.
To deal with dense networks with a high number of devices and scalability issues, we propose
large-scale algorithms that support a huge amount of devices, data, and users requests.
Heuristic algorithms are proposed for SFC orchestration, maximum coverage of mobile
edge servers (vehicles). Moreover, The artificial intelligence algorithms (machine learning,
deep learning, and deep reinforcement learning) are used for 5G VNF slices placement,
v
edge-autopilot VNF placement, and autonomous UAV navigation. The numerical results
give good results compared with exact algorithms with high efficiency in terms of time.
Keywords: Internet of things (IoT), Cloud Computing, Edge/Fog Computing, Mo-
bile Edge Computing, Artificial Intelligence, Optimization.
Résumé
Dans le cloud computing, les services et les ressources sont centralisés dans des cen-
tres de données auxquels l’utilisateur peut accéder à partir de ses appareils connectés.
L’infrastructure cloud traditionnelle sera confrontée à une série de défis en raison de la
centralisation de calcul, du stockage et de la mise en réseau dans un petit nombre de centres
de données, et de la longue distance entre les appareils connectés et les centres de données
distants.
Pour répondre à ce besoin, l’edge computing s’appuie sur un modèle dans lequel les
ressources de calcul sont distribuées dans le edge de réseau selon les besoins, tout en
décentralisant le traitement des données du cloud vers le edge autant que possible. Ainsi,
il est possible d’avoir rapidement des informations exploitables basées sur des données qui
varient dans le temps.
Dans cette thèse, nous proposons de nouveaux modèles d’optimisation pour optimiser
l’utilisation des ressources dans le edge de réseau pour deux domaines de recherche de
l’edge computing, le "service offloading" et "vehicular edge computing". Nous étudions
différents cas d’utilisation dans chaque domaine de recherche. Pour les solutions optimales,
Premièrement, pour le "service offloading", nous proposons des algorithmes optimaux pour
le placement des services dans les serveurs edge (Tasks, Virtual Network Functions (VNF),
Service Function Chain (SFC)) en tenant compte des contraintes de ressources de calcul.
De plus, pour "vehicular edge computing", nous proposons des modèles exacts liés à la
maximisation de la couverture des véhicules par les taxis et les Unmanned Aerial Vehicle
(UAV) pour les applications de streaming vidéo en ligne. De plus, nous proposons un edge-
autopilot VNFs offloading dans le edge de réseau pour la conduite autonome. Les résultats de
l’évaluation montrent l’efficacité des algorithmes proposés dans les réseaux avec un nombre
limité d’appareils en termes de temps, de coût et d’utilisation des ressources.
vi
Pour faire face aux réseaux denses avec un nombre élevé d’appareils et des problèmes
d’évolutivité, nous proposons des algorithmes à grande échelle qui prennent en charge une
énorme quantité d’appareils, de données et de demandes d’utilisateurs. Des algorithmes
heuristiques sont proposés pour l’orchestration SFC, couverture maximale des serveurs
edge mobiles (véhicules). De plus, les algorithmes d’intelligence artificielle (apprentissage
automatique, apprentissage en profondeur et apprentissage par renforcement en profondeur)
sont utilisés pour le placement des "5G VNF slices", le placement des "VNF-autopilot" et
la navigation autonome des drones. Les résultats numériques donnent de bons résultats par
rapport aux algorithmes exacts avec haute efficacité en temps.
Mots clés: Internet des objets (IoT), Cloud Computing, Edge/Fog Computing,
Mobile Edge Computing, Intelligence Artificielle, Optimisation.
vii
Thesis Publications
Journals
1. Mohammed Laroui, Hatem Ibn-Khedher, Moussa Ali Cherif, Hassine Moungla, Hos-
sam Afifi, and Ahmed E Kamel. SO-VMEC: Service offloading in virtual mobile edge
computing using deep reinforcement learning. Transactions on Emerging Telecommu-
nications Technologies (ETT), e4211, 2021. [1]
2. Mohammed Laroui, Boubakr Nour, Hassine Moungla, Moussa Ali Cherif, Hossam
Afifi, and Mohsen Guizani. Edge and Fog Computing for IoT: A Survey on Current
Research Activities & Future Directions. Computer Communications (ComCom), 2021.
[2]
Conferences
1. Mohammed Laroui, Hatem Ibn-Khedher, Hassine Moungla, and Hossam Afifi. Au-
tonomous UAV Aided Vehicular Edge Computing for Service Offering. In IEEE Global
Communications Conference (Globecom), 7-11 December 2021, Madrid, Spain. [3]
2. Mohammed Laroui, Hatem Ibn-Khedher, Hassine Moungla, and Hossam Afifi. Artifi-
cial Intelligence Approach for Service Function Chains Orchestration at The Network
Edge. In IEEE International Conference on Communications (ICC), (pp.1-6), 14-23
June 2021, Montreal, QC, Canada. [4]
3. Hatem Ibn-Khedher, Mohammed Laroui, Mouna Ben Mabrouk, Hassine Moungla,

Hossam Afifi, Alberto Nai Oleari, and Ahmed E Kamal. Edge Computing Assisted
Autonomous Driving Using Artificial Intelligence. In IEEE International Wireless
Communications and Mobile Computing (IWCMC), (pp.254-259), 28 June-2 July 2021,
Harbin City, China. [5]
4. Mohammed Laroui, Hatem Ibn Khedher, Hassine Moungla, Hossam Afifi, and Ahmed
E Kamal. Virtual Mobile Edge Computing Based on IoT Devices Resources in Smart
viii
Cities. In IEEE International Conference on Communications (ICC), (pp.1-6), 7-11
June 2020, Dublin, Ireland. [6]
5. Mohammed Laroui, Moussa Ali Cherif, Hatem Ibn Khedher, Hassine Moungla, and
Hossam Afifi. Scalable and Cost Efficient Resource Allocation Algorithms Using
Deep Reinforcement Learning. In IEEE International Wireless Communications and
Mobile Computing (IWCMC), (pp.946-951), 15-19 June 2020, Limassol, Cyprus. [7]
6. Mohammed Laroui, Boubakr Nour, Hassine Moungla, Hossam Afifi, and Moussa Ali
Cherif. Mobile Vehicular Edge Computing Architecture using Rideshare Taxis as a
Mobile Edge Server. In IEEE 17th Annual Consumer Communications & Networking
Conference (CCNC), (pp.1-2), 10-13 January 2020, Las Vegas, NV, USA. [8]
7. Mohammed Laroui, Aicha Dridi, Hossam Afifi, Hassine Moungla, Michel Marot, and
Moussa Ali Cherif. Energy Management For Electric Vehicles in Smart Cities: A Deep
Learning Approach. In IEEE International Wireless Communications and Mobile
Computing (IWCMC), (pp.2080-2085), 24-28 June 2019, Tangier, Morocco. [9]
8. Mohammed Laroui, Akrem Sellami, Boubakr Nour, Hassine Moungla, Hossam Afifi,
Sofiane Boukli Hacene. Driving Path Stability in VANETs. In IEEE Global Communi-
cations Conference (GLOBECOM), (pp.1-6), 9-13 December 2018, Abu Dhabi, United
Arab Emirates. [10]
ix
Table of contents
List of figures xiii
List of tables xv
1 General Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Edge Computing For IoT: Overview & Related Works 5

2.1 Edge Computing for IoT: Overview . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Internet of Things (IoT) . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Cloud Computing (CC) . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.3 Mobile Cloud Computing (MCC) . . . . . . . . . . . . . . . . . . 13
2.1.4 Edge Computing (EC) . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1.5 Mobile Edge Computing (MEC) . . . . . . . . . . . . . . . . . . . 23
2.2 Edge Computing for IoT: Related Works . . . . . . . . . . . . . . . . . . . 29
2.2.1 Service Offloading In EC . . . . . . . . . . . . . . . . . . . . . . . 29
2.2.2 Vehicular Edge Computing . . . . . . . . . . . . . . . . . . . . . . 35
3 Optimal Solutions For Edge Computing Networks 47

3.1 Service Offloading In EC . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.1.1 UseCase 1: 5G VNF Slices Placement in EC . . . . . . . . . . . . 48
3.1.2 Use Case 2: Task Offloading in Virtual Mobile Edge Computing
(VME) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Table of contents
3.1.3 Use Case 3: Service Offloading in Virtual Mobile Edge Computing

(SO-VMEC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.1.4 Use Case 4: Service Function Chains Orchestration in Virtual Mobile
Edge Computing (VMEC) . . . . . . . . . . . . . . . . . . . . . . 73
3.2 Vehicular Edge Computing . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.2.1 Use Case 1: Edge Computing Assisted Vehicular Networks for
Service Offering . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.2.2 Use Case 2: Edge Computing Aided Autonomous Vehicles . . . . . 83
3.2.3 Use Case 3: Edge Computing Assisted Autonomous UAV . . . . . 89
4 Large Scale Solutions For Edge Computing Networks 99

4.1 Service Offloading In EC . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.1.1 Use Case 1: 5G VNF Slices Placement in EC . . . . . . . . . . . . 99
(SO-VMEC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.1.3 Use Case 3: Service Function Chains Orchestration in Virtual Mobile
Edge Computing (VMEC) . . . . . . . . . . . . . . . . . . . . . . 106
4.2 Vehicular Edge Computing . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Service Offering . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.2.2 Use Case 2: Edge Computing Aided Autonomous Vehicles . . . . . 109
4.2.3 Use Case 3: Edge Computing Assisted Autonomous UAV . . . . . 112
5 Conclusion and Perspectives 115
References 117
xii
List of figures
2.1 IoT Application Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2 3-Tiers IoT Architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Categories of Cloud Computing Services. . . . . . . . . . . . . . . . . . . 13
2.4 Cloud Computing Architecture for IoT. . . . . . . . . . . . . . . . . . . . 14
2.5 Mobile Cloud Computing Architecture. . . . . . . . . . . . . . . . . . . . 15
2.6 Edge Computing Architecture for IoT. . . . . . . . . . . . . . . . . . . . . 18
2.7 Mobile Edge Computing Framework. . . . . . . . . . . . . . . . . . . . . 24
2.8 Mobile Edge Computing Architecture. . . . . . . . . . . . . . . . . . . . . 26
3.1 VNF Slices Placement in Edge Computing Over 5G Network [7]. . . . . . 48

3.2 Virtual Mobile Edge Computing Architecture [6]. . . . . . . . . . . . . . . 51
3.3 Failed Tasks [6]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4 Service Time [6]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.5 Processing Time [6]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.6 Average Edge Devices Utilization [6]. . . . . . . . . . . . . . . . . . . . . 56
3.7 Failed Tasks due to Edge Device Capacity [6]. . . . . . . . . . . . . . . . . 57
3.8 Virtual Mobile Edge Computing Architecture [1]. . . . . . . . . . . . . . . 59
3.9 SO-VMEC Communication Model [1]. . . . . . . . . . . . . . . . . . . . 60
3.10 Mobility Prediction for Three Mobility Models (Low, Medium, and High) [1]. 69
3.11 Energy Prediction for Three Energy Models (Low, Medium, and High) [1]. 70
3.12 Energy Prediction Impact on VNF Slices Execution [1]. . . . . . . . . . . . 71
3.13 SO-VMEC Evaluation Results [1]. . . . . . . . . . . . . . . . . . . . . . . 72
3.14 Proposed IoT-VMEC Architecture. . . . . . . . . . . . . . . . . . . . . . . 74
3.15 Energy Prediction for Three Energy Models (Low, Medium, and High). . . 79
xiii
List of figures
3.16 Mobility Prediction for Two Different Mobility Models (Low and High). . . 79
3.17 MVEC architecture overview [8]. . . . . . . . . . . . . . . . . . . . . . . . 81
3.18 CEC with/out Edge Servers [8]. . . . . . . . . . . . . . . . . . . . . . . . 83
3.19 OVEAP Performance Evaluation [5]. . . . . . . . . . . . . . . . . . . . . . 90
3.20 Proposed FMU Architecture [3]. . . . . . . . . . . . . . . . . . . . . . . . 91
3.21 Abstraction Model For FMU [3]. . . . . . . . . . . . . . . . . . . . . . . . 92
3.22 System Evaluation for Different Communication Scenarios [3]. . . . . . . . 96
4.1 RL application to VNF placement in Edge Computing [7]. . . . . . . . . . 100

4.2 Performance Evaluation of VNF Slices Placement in the edge using ILP,
Q-Learning, DQN in Large Scale Networks [7]. . . . . . . . . . . . . . . . 103
4.3 Deep Reinforcement Learning for VNF Slices Offloading [1]. . . . . . . . . 104
4.4 VME Based on ILP and DQN Models For VNFs Slices Offloading [1]. . . . 105
4.5 Total Resources Utilization for Different Flow Rates of SFC VNFs. . . . . . 108
4.6 Real Routes With Taxis Coverage [8]. . . . . . . . . . . . . . . . . . . . . 109
4.7 OVEAP and DVEAP for autopilots VNFs Offloading [5]. . . . . . . . . . . 111
4.8 OFMU & AFMU Evaluation [3]. . . . . . . . . . . . . . . . . . . . . . . . 113
xiv
List of tables
1 List of abbreviations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
2.1 CC & MCC vs EC & MEC . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.2 Summary of the existing works for service offloading in EC. . . . . . . . . 35
2.3 Summary of the existing works for V2X-assisted EC. . . . . . . . . . . . . 41
2.4 Summary of the existing works for UAV-assisted EC. . . . . . . . . . . . . 44
3.1 The notations used in the proposed VNF slices placement model. . . . . . . 49
3.2 The notations used in the proposed VEnPA model. . . . . . . . . . . . . . 53
3.3 VEnPA parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4 The notations used in the proposed SO-VMEC model. . . . . . . . . . . . . 63
3.5 SO-VMEC parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.6 The main strong points of VME against AME and AMC. . . . . . . . . . . 68
3.7 The parameters used in the proposed OSPV model. . . . . . . . . . . . . . 76
3.8 MVEC parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.9 OVEAP Parameters and Decision Variables . . . . . . . . . . . . . . . . . 85
3.10 Autonomous Vehicles Configuration. . . . . . . . . . . . . . . . . . . . . . 87
3.11 Summary of autopilot edge servers (AES) parameters. . . . . . . . . . . . . 88
3.12 The used notations in the FMU model. . . . . . . . . . . . . . . . . . . . . 91
3.13 FMU parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.1 Servers Configuration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

4.2 VNF Configuration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.3 Summary of the parameters. . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.4 OFMU and AFMU Algorithms Complexity. . . . . . . . . . . . . . . . . . 112
xv
Chapter 1
General Introduction
1.1 Motivation
The Internet of Things (IoT) has been a huge factor in the development of the high-tech
industry. With the large number of devices and profits to be made over the coming years,
the IoT will have a profound and dominant impact on the high-tech industry in general and
the semiconductor industry in particular. IoT devices can actually be any type of sensors
and chips with different capabilities made by different manufacturers, and there are many
applications that can be built to enable smart cities, smart transportation, smart homes, and
smart healthcare.
In the IoT, connected devices can generate an enormous amount of data at very high
speeds and some applications can require very low latency. The data is directly sent to the
cloud for storage and processing in the data centers. Traditional cloud infrastructure will face
a series of challenges due to the centralization of storage, computing, and networking in a
small number of data centers, and the long-distance between connected devices and remote
data centers.
To meet this challenge, edge computing seems to be a promising technology that provides
computing resources closer to IoT devices.
In this thesis, we are motivated to identify and troubleshoot problems of edge computing
architecture and the integration of this architecture with IoT applications. Further, the
1
General Introduction
literature lacks general edge-IoT architectures, this leads us to design recent architectures,
especially for service offloading and vehicular edge computing.
Furthermore, despite the optimization tasks importance, such efficient functions are
missing in the global edge-IoT architectures.
For this, we propose novel different optimization models (exact and large-scale algorithms)
at the edge layer that solve both service offloading problems related to placement at the
network edge, and vehicular edge computing problems that uses vehicles as mobile edge
servers to provide services to end-users such as caching, computing, and streaming.
1.2 Contribution
Our contribution presents new optimization algorithms to improve the quality of service in
the Internet of things based on edge computing networks to achieve maximum reliability and
efficiency for both service offloading and vehicular edge computing.
First, for service offloading, we propose optimal placement models at the network edge for
end-users tasks (ie. Optimal Virtual Edge nodes Placement Algorithm (VEnPA)), VNFs
slices (ie. Service Offloading in Virtual Mobile Edge Computing (SO-VMEC)), and SFCs (ie.
Optimal SFC Placement in IoT-VMEC (OSPV)). Then, for vehicular edge computing, we
propose optimal model for mobile edge servers (Taxis and UAV) coverage for video stream-
ing application (ie. Mobile Vehicular Edge Computing (MVEC), Optimal Follow Me UAV
(OFMU)), and Edge autopilot for autonomous driving (ie. Optimal Virtual Edge-Autopilot
Placement (OVEAP)).
Further, to cope with the high complexity problems of optimal algorithms in dense networks,
we propose large scale algorithms for both service offloading and vehicular edge computing
use cases. The artificial intelligence algorithms reinforcement learning and deep reinforce-
ment learning are used for the placement of end-users tasks and VNFs slices (Q-learning and
Deep Q-learning (DQN)), while an Efficient SFC Placement algorithm (ESPV) based on the
Bin Packing model is implemented for SFCs placement in virtual edge servers.
Moreover, for vehicular edge computing, we propose different heuristic algorithms to cover
the high requirement in dense networks. For the first use case, to maximize the client vehicles
coverage using mobile edge server (Taxis) we propose an heuristic algorithm that select
2
1.3 Outline
only the taxis that share the same path segment(s) with the client vehicle to decrease the
complexity of the algorithm. For the second use case, we propose heuristic algorithm for
FMU (AFMU) based on the Weighted Set Cover (WSC) model to increase the UAV avail-
ability during the service. Finally, for the last use case, we propose Efficient Edge Autopilot
Placement (DVEAP) based on deep reinforcement learning to place efficiently the autopilot
chains in the edge server.
To assess the efficiency of the proposed algorithms for both small scale and large scale
networks we use different tools including programming language, platforms for artificial
intelligence models, and simulators for both networking and computing.
1.3 Outline
The remainder of this thesis is organized as follows. Chapter (2) consists of two parts, first,
we overview the Internet of things (IoT) and the use of cloud to provide computing services
to IoT devices. Moreover, because of the latency issue of the cloud especially for real-
time applications, we introduce the edge computing technology with its main applications
and use cases. Then, we present the related works proposed by the research community
for both service offloading and vehicular edge computing to show the main contributions
and enhancements to place efficiently the services in the edge servers, and maintain the
high availability of the services in a vehicular environment. Then, in the chapter (3), we
present the optimal algorithms for both service offloading (VEnPA, SO-VMEC, OSPV),
and vehicular edge computing (MVEC, OFMU, OVEAP) with its variables, constraints and
objective function. Chapter (4) outlines the large-scale solutions for both service offloading
(Q-learning, DQN, ESPV), and vehicular edge computing (Maximum coverage, AFMU,
DVEAP). Finally, chapter (5) present the conclusion and the future topics and issues for edge
computing networks.
3
Chapter 2
Edge Computing For IoT: Overview &

Related Works
The rapid increase of smart devices in recent years and the change of applications’ require-
ments led to the innovation of current network architecture to satisfy the needs of user
applications to allow fast data processes and reliable services. Moreover, due to the per-
petual emergence of greedy Internet applications and user demands that are challenging
communication and computation, Edge Computing has been proposed to overcome cloud
computing issues and to bring data computation closer to end-users. The first part of this
chapter highlight the evolution of edge computing to cope with IoT applications by analyzing
the need of edge servers with the cloud to overcome the existing issues. Then, the second
part presents the state of the art for the two research directions that are studied in this thesis,
service offloading and vehicular edge computing.
2.1 Edge Computing for IoT: Overview
2.1.1 Internet of Things (IoT)

The development of communication technology across generations and the competition of
various international companies in order to provide fast and advanced communication services
with the lowest cost led to the rapid evolution of connected vehicles, embedded systems, and
mobile devices which resulted in the emergence of a smart world of interconnected devices
5
Edge Computing For IoT: Overview & Related Works
that collaborate, collect data, sense, and take decisions without the need to interact with the
human. This smart world is called the Internet of things.
Definition
The Internet of things (IoT) [11, 12] allows connecting any things including vehicles, drones,
devices, solar panels, etc that exchange data with each other. IoT is one of the important
topics of future network technologies, in addition, it has an important interest from companies
in different domains [13, 14] such as agriculture, medicine, airlines, energy, and telecommu-
nication. The IoT aims to create a smart world based on different systems of interconnected
objects. The different connected objects communicate with each other without the need to
interact with a human using different models such as Bluetooth, Zigbee, and Wi-Fi. The main
characteristic of IoT is the use of different communication technologies (e.g., identification,
tracking, and actuator networks, wireless and wired sensors, etc.) to improve the interaction
and the cooperation of various technologies.
An IoT system, for example, a smart home works by sending, analyzing, and receiving
data continuously. Depending on the type of IoT system, analyzes can be performed by
humans or by artificial intelligence aided by machine learning (AI / ML), in near real-time or
over a long time. In the example of the smart home, to predict the optimal time to set the
thermostat before return to home, the IoT system can connect to the Google Maps API to
obtain real-time traffic modeling in the user area, or use data relating to the usual journeys
that the car has collected over a long time. In addition, IoT data collected by each customer’s
thermostat can be analyzed by energy providers to optimize their services on a larger scale.
From an IT perspective, IoT solutions allow businesses to enhance their existing systems
and create new points of connection with customers and partners. However, they also bring
their share of new computer problems. A system of connected devices can generate massive
amounts of data, which is referred to as “Big Data”. However, integrating this Big Data
into existing systems and setting up analyzes that allow it to be exploited is a challenge. In
addition, security can become a critical issue depending on the degree of openness of the
future IoT platform. Despite this, the benefits that the IoT brings to businesses more than
outweigh the efforts required to implement it, and businesses in virtually every industry are
already using it successfully.
6
In agriculture, the IoT is revolutionizing the world of agriculture on several levels,

especially with the use of humidity sensors. By installing a network of humidity sensors in
their fields, farmers are now able to receive more accurate data that allows them to determine
the best time to irrigate their crops. The IoT can be taken advantage of further in this case,
by connecting humidity sensors to IoT applications that control irrigation systems; Then
irrigation is triggered automatically from the data received from the sensors, without any
human intervention.
As with all technological progress, users remain wary of the IoT, however, because
even if this trend opens up new and very interesting possibilities, it raises questions about
the confidentiality and security of its data. So, cannot ignore this aspect when planning to
implement an enterprise-wide IoT project, especially when targeting the general public as
the end-user.
Figure 2.1 displays the IoT application domains.
Smart City
Security & Surveillance Smart Home
Internet of things
FOOD CHAIN...
application domains
Industrial Automation Smart Healthcare
Smart Vehicle
Fig. 2.1 IoT Application Domains
7
Essential Technologies
The IoT consists of three different components: (a) actuator: is the hardware of sense. (b)
storage module and data analytics tools, and (c) visualization and interpretation tools. The
following technologies represent the components defined above:
• Radio Frequency Identification (RFID) [15]: RFID enables the development and the
design of microchips. It allows the numerical identification of anythings (e.g. card, vehicle,
smart device, etc.) attached with a barcode. There are two main components of the RFID
system, (a) RFID tags: are devices represented as microchips used to send data using
wireless transmission. (b) RFID reader recovers the incoming signal from the microchips, is
responsible for the object identification. The RFID system is one of the most used devices
in IoT applications [16–18] including social applications, smart healthcare, and controlling
privacy.
• Wireless Sensor Networks (WSN) [19]: is a set of sensors that collect data (e.g.,
temperature, movement, etc.), process, and transfer the collected information. WSN is
composed of the following components: (a) the sensor (unit of capture) is responsible for
data collecting from an environment as signals, and transfer it into digital data. (b) processing
unit: allows the analysis of the captured data. (c) transmission unit: responsible for data
transmission/reception, and (d) energy control unit: represent an essential part of the WSN
system, responsible for optimal energy distribution to the other modules. WSN is used in
different domains including medical application (e.g., remote medical monitoring), military,
commercial, and environmental monitoring (e.g., forest fires, landslides, pollution, etc).
• Data Storage and Analytics [20]: the use of sensors and connect the different objects
in the IoT network lead to generating a massive amount of data that require a huge storage
capacity. Because of this reason, the data storage is an important issue in the IoT network. To
overcome this issue, different solutions have been proposed to provide efficient data storage
and analytic. Moreover, in the last years the use of the cloud is become more attractive
and popular, where nowadays the use of cloud for data analytics and storage is mostly
preferred because it provide huge storage capacity as well as high computing resources for
data analytics.
8
Architecture
Different IoT architectures have been proposed for generic and specific use cases. In general,
the basic IoT architectures consists of three parts [21] as follow: (a) perception layer: includes
RFIDs, sensors, cameras, etc., (b) network layer: is responsible for transmitting the data
collected from the perception layer, and (c) application layer that represents the application
such as smart home and smart healthcare. Figure 2.2 display the 3-tier IoT architecture. The
rapid development of IoT networks with the new generation of applications led to generate
a huge amount of data that can not be processed in both the source devices and the access
points because of it’s computing requirement that exceeds the existing computing limit. For
this, a 2-tier architecture called cloud computing has been deployed to process data in the
cloud that offer a high capacity for storage and computing.
2.1.2 Cloud Computing (CC)

The rapid development of the new generation of applications the IoT led to new challenges
for data processing and storage to satisfy the resources required for these applications. Cloud
computing is an innovative technology that provides high computing resources for IoT
applications including fast processing and huge storage capacity of data.
Overview
The use of computers and smartphones has dramatically increased in the last years. This
growth led to the need for efficient architectures to provide the required processing, especially
for data processing and storage. Moreover, support the increase of devices as well as the
development of applications. Cloud computing [22, 23] is an innovative technology that
allows providing a centralized environment for data processing including high computing
resources and huge data storage capacity. Different types of cloud have been deployed [24]
including public, private, community, and hybrid. The public cloud offers different services
to connected users on the Internet. On the contrary, the private cloud provides services to
specific organizations. Where the community cloud allows to offers services to a group of
companies. The last type is the hybrid cloud that offers services where it’s a good choice for
companies because of the balance between both resource control and the utilization cost.
9
Application Layer Smart Smart Smart

City Healthcare Home
Network Layer
Router/Gateway
Router/Gateway
Router
Router
Perception Layer
IoT Devices
(sensors,...)
Fig. 2.2 3-Tiers IoT Architecture.
Cloud Computing Deployment Models
1. Public Cloud:
The public cloud is a fully connected network of high computing resources accessible
to the general public users via the Internet. It allows the storage and processing of data,
but also the development of applications and various technologies (Artificial Intelligence,
Machine Learning, etc.). Public cloud services are provided by service providers, besides,
these services can be free or billed peruse. The largest providers offer many services on their
public cloud platforms, to guarantee their users an ever richer experience and an ever greater
variety of products.
10
The public cloud has several advantages: ease of configuration and use, great flexibility,
quick adaptation to business needs, and pay-as-you-go according to user needs.
2. Private Cloud:
The private cloud is a storage or data processing space only accessible by its owner,
whether he is a single customer or a group of companies. Only users linked to this client can
access the data and applications developed in this private area.
This configuration has several advantages: increased confidentiality, better control of the
stored data, and a feeling of security.
3. Hybrid Cloud:
The hybrid cloud generally describes a cross form between a public and a private cloud.
As a result, some of the applications and data processed on company servers and some on
the servers of a dedicated provider. Ideally, a hybrid cloud combines both private and public
clouds together in a transparent and symbiotic fashion. The company alone decides which
type of data kept locally and the data sent to the cloud servers, for example, preserve the
data protection files locally and store the remaining data on the cloud servers. while some
other companies prefer to outsource only cloud computing and keep full storage locally. Or
conversely, storing the data in the cloud to allow remote access, and use the computing power
locally.
Cloud Computing Service Models
The cloud provides different services categorized by the type of service. The three categories
of cloud services can be defined as follows:
1. Software as a service (SaaS): [25] is an online service in the cloud allowing developers
to create their applications via the Internet without the need to install any software on their
local computer. When users use SaaS, the data is permanently stored on the cloud. In
addition, users do not have the possibility to control or manage the network, storage of data
and the operating system used. SaaS has different advantages for both service providers and
end users. Service providers enjoy maintenance, installation, and control of versions, while
the end users enjoy access the services anywhere and anytime. Also, collaborate and share
data becomes more easily.
11
2. Platform as a service (PaaS): [26] this type provides users the control of different
resources in the cloud over the Internet including network capacity, operating systems, hard-
ware and storage management, etc,. The PaaS offers a virtualized environment for users
to design, deploy, and run their applications without the need to extend the capacity of the
resources in their computers. This platform consists of software that includes development
tools, middleware and databases. The main advantages of PaaS are: manage and update regu-
larly the operating system features, provide an efficient scalability, offer quick development
of softwares and applications, and provide efficient and better technology to businesses.
3. Infrastructure as a Service (IaaS): [27] offers to users the possibility to control and
manage both hardware and software in the cloud. Moreover, IaaS provides online access
to platforms and applications anywhere, using any connected device from any network,
provides virtual infrastructure, load balancing, as well as a large capacity for computing. The
storage services provide a large amount of data storage and retrieval services from anywhere
using any device. In addition, provides security and compliance features that meet the most
stringent data security requirements. Moreover, offers many advantages such as managing
data with optimized costs. The computing services offers a resizable computing capacity
in the cloud. It is designed to offers developers easier services such as a large capacity
of computing and tools to build different applications, In addition, it guarantees flexibility,
security, and a reliable environment. The use of IaaS has different benefits where users can
access applications and platform (that are always available) from anywhere, using any device,
and from any network, provide virtual infrastructure (i.e. network virtualization, storage, and
servers), as well as provide services of load balancing and a large capacity for computing.
Figure 2.3 display the three categories of cloud computing services.
CC for Internet of Things
Both IoT and cloud computing are two innovative technologies that offer various services to
our daily life. The cloud can serve the IoT technology by providing high computing resources
for processing and a huge capacity for storage, that represent the main requirements for IoT
to support the new generation of applications. For this, the CloudIoT paradigm is the merge
of the two technologies into one architecture that connects the IoT devices directly with the
cloud. Botta et al. [28] presented an integration of IoT with the cloud by providing services
12
Private Cloud Infrastructure as Platform as a Software as a

(On Premise) a service (IaaS) service (PaaS) service (SaaS)
Applications
Customer
Applications Applications Applications
Customer
Data Data Data Data
OS OS OS OS
Customer
Provider
Virtualization Virtualization Virtualization Virtualization
Provider
Servers Servers Servers Servers
Provider
Storage Storage Storage Storage
Networking Networking Networking Networking
Fig. 2.3 Categories of Cloud Computing Services.
to IoT such as computing resources and data storage. Similarly, Neagu et al. [29] proposed a
monitoring service for healthcare that offers different medical facilities by using cloud-IoT
architecture. Ismail et al. [30] presented a study for task scheduling algorithm and virtual
machine placement in Cloud-IoT architecture where they focused on the consumed energy
in data centers. In the CloudIoT system, the devices connect directly with the cloud over
the Internet and start exchanging data where the cloud provides processing resources and
storage capacity to the end-user devices which form an interconnected system represent the
IoT network. Figure 2.4 outline the CloudIoT architecture.
2.1.3 Mobile Cloud Computing (MCC)

Despite the rapid increase in mobile computing usage, it is difficult to exploit all its resources
because of its inherent problems including frequent disconnections, mobility, and resource
scarcity. Mobile cloud computing can resolve these issues by using providers’ resources to
execute mobile applications outside to the end-user devices (mobile devices).
13
Cloud Computing
Cloud
Controller
Cloud
Controller
Data Center
Data Center
Internet
Satellite Access Point

BTS
Fig. 2.4 Cloud Computing Architecture for IoT.
Overview
In a time of profound change in the computing field, the main developments have been on
both mobile Internet and cloud computing. These two technologies overlap in mobile cloud
computing.
The mobile cloud computing (MCC) [31–33] is a technology that combines both mobile
computing and cloud computing to bring rich computational resources to network operators,
cloud computing providers as well as mobile users. In MCC technology, data storage, and
data processing happen outside of end-user devices (mobile devices) which offer many
benefits such as extend battery life, improved scalability, and reliability, improve processing
14
power and data storage capacity, as well as access to data from anywhere. MCC is also
widespread within companies by making cloud services available to mobile personnel who
can access them in the field or on the move.
Architecture
Based on the general concept of MCC [34], figure 2.5 display the typical architecture of
MCC. From the figure, the mobile devices (smartphones, laptops,...) are connected to the
base station (satellite, access point, base transceiver station, etc) that allows to create the
communication link, control the connection state, and functional interfaces between the
mobile devices and the network. The requests of mobile users are transmitted to the mobile
network that provides local services such as authorization, accounting, and authentication.
For computing requests, the users demands are forwarded from the mobile network to the
cloud over the Internet. In the cloud, the controllers process the users requests to provide
services including computing, database servers, web applications, and virtualization, etc.
Mobile Network A
Mobile Network Services

Database
Satellite
Servers
Agent Cloud Computing
BTS
Authentication
Authorization
Accounting
Cloud
Access Point Controller
Cloud
Internet
Controller
Data Center
Mobile Network B Data Center
Mobile Network Services

Database
Satellite
Servers
Internet Service
Agent
Providers (ISPs)
BTS
Authentication
Authorization
Accounting
Access Point
Mobile Users Network Operators

Gateway
Fig. 2.5 Mobile Cloud Computing Architecture.
The centralized architecture of the cloud is not efficient to process large size of data
generated by the IoT devices that require a short response time. To overcome such issues,
an alternative paradigm namely edge computing has been used that processes the data in
connected devices or the local gateways.
15
2.1.4 Edge Computing (EC)

In the incoming years, the number of connected devices in the IoT network will increase
rapidly [35], this lead to generate a huge quantity of data [36]. Moreover, the new generation
of applications such as online gaming, video streaming,.., etc. requires a very short response
time to satisfy the quality of service requirements and provide efficient service to end-users.
In addition, energy consumption in wireless communication is an important issue because
of the limited capacity of the battery in IoT devices, which by consequence lead to the
impossibility of processing in the local device. Thus, the centralized platform does not
satisfy the IoT applications requirements by providing computation, storage, and networking
resources in the cloud owned by companies that require payment for any service. Moreover,
the huge amount of data sent to the cloud need to be processed in a short time which can not
be provided in a centralized architecture.
Overview
The edge computing [37, 38] paradigm aims to be an innovative computing solution for
future networks. Thus, it allows solving critical issues including computation-based and
time-constrained applications. The main benefits of processing data at the edge of the network
are: reduce the energy consumption of mobile devices, reduce communication latency and
network load, give the new inventors chances to help future network innovations, provide
security, privacy, and reliability, as well as eliminating the congestion within the network
core.
Concepts & Relationships
To cope with the ambiguity related to the difference between fog computing and edge
computing, we describe both fog and edge computing from different perspectives.
In the research community, some researchers considered edge computing and fog comput-
ing as the same technology with differences only in their names [39]. While other researchers
defined the two as different concepts. Chiang et al. [40] said the edge refers to the edge
network, with equipment such as edge routers, home gateways, and base stations. While fog
computing is an end-to-end platform that distributes the computing, storage, control, and
16
networking function closer to end-user devices over the cloud-to-things. On the other hand,
Goscinski et al. [41] claims that fog computing is located out between the edge cloud and the
cloud, and thereby include the edge, while edge computing focuses on the processing of data
at the edge. Moreover, Pan et al. [42] said that edge computing pushes applications, services,
and data from the core to the network edge, based on the core-edge topology [36, 43]. The
smart city, video analytics, cloud offloading and smart home are examples of edge computing
applications. While fog computing is a background of the IoT, it extends cloud computing
and different services to the devices such as multiplexers routers, switches, etc.
In a nutshell, both fog and edge computing have the same research topics and both of
them aim at providing the computing services closed to end-users by the decentralization of
data processing from the cloud to the edge of the network.
Architecture
The fog/edge layer is located between the end-users and the cloud, as is shown in the
figure 2.6, is composed of the following components [44]:
(a) Authentication and Authorization: identify the access policies and the control roles,
(b) Offloading Management: defines the manner to design an optimal offloading scheme, the
partition for offloading, and the types of information in the offloading process, (c) Location
Services: allows learning the mobility model by mapping the physical locations with the
network, (d) System Monitor: offer information such as energy, usage, and workload to the
other components, (e) Resource Management: is responsible for resource allocation and
discovery, dynamic joining and leaving of the fog node, provisioning, and maintaining of
resources, and (f) VM Scheduling: that aims to provide an optimal strategy and management
for virtual machine scheduling.
EC for Internet of Things
The edge offers networking services, computation, and storage to end-users in IoT networks
[45]. For example, an intelligent surveillance system doesn’t need to send all recorded data to
the cloud, the facial recognition and movement detection algorithms running on the fog/edge
layer, hence saving bandwidth and storage space. So, the fog/edge will be the best option for
applications that require temporary storage. In addition, the fog/edge can play an important
17
Cloud Computing
Cloud
Controller
Cloud
Controller
Data Center
Data Center
Internet
Edge Layer
Ed Ed Ed
ge ge ge
Se Se Se
rve rve rve
r r r

BTS
Connected Devices
Fig. 2.6 Edge Computing Architecture for IoT.
role in the future, where the wireless sensors networks and IoT present with the integration
of heterogeneous devices and protocols for enhancing services [46]. The edge computing is
an appropriate architecture for the current and future Internet in general and the IoT networks
in particular in which this distributed architecture may provide high-level communication
and efficiency in the edge of the network rather than at the cloud-level. It supports a wide
range of today’s IoT applications, especially those that require a short response time.
18
Edge Computing for IoT: Use Cases & Applications
The edge computing offers an efficient platform for current and future IoT applications
including smart cities, smart homes, smart grid, smart healthcare, etc. Different applications
have been proposed for IoT based on edge computing as follow:
(a) Smart City Systems: A smart city [47] is an urban area composed of connected sectors
that pull together to provide services to users through the data analysis that is collected from
different sources. Smart cities aim to decrease traffic congestion and energy consumption
which contribute to improving the quality of life.
Different technologies are used in smart cities that will create an important economic
market which by consequence that reflected on the economic development of countries.
Though, different issues and challenges can still face the deployment of smart cities such
as large-scale sensing networks deployed in a large geographic area, big data analysis,
cooperative communication in a wireless network, and energy and mobility management in
intelligent traffic systems.
Smart cities require intelligent and efficient data monitoring and analysis in order to
achieve an automated and fast decision without any human intervention [48]. Hereby, ensure
people’s safety and the reliability of components. Furthermore, smartness has some require-
ments such as efficient platforms with advanced algorithms including artificial intelligence
models (machine learning [49], deep learning [50], and deep reinforcement learning [51]).
The edge computing technology is an efficient platform for smart cities [52], due to its
architecture, the process of data will happen at the edge of the network instead of the cloud
[36] and enhance both the network delay and the user experience. This is an important and
essential element to build reliable and efficient smart cities that can manage the following
challenges:
• Huge Quantity of Data: A huge amount of data will be generated by future applications
in smart cities varies from smart traffic lights, utility, transport, etc. The process of all this
data in the cloud is inefficient because of the long distance between the cloud and end-users
which takes more time for data transfer. For this, the process of data at the edge of the
network close to end-users will resolve the issues of data processing in the centralized cloud
by offering an efficient processing service that is required by current and future applications
in smart cities.
19
• Low Latency: The latency requirement is one of the most important issues in smart
cities, where most applications such as health emergencies and public safety require a low
latency to provide efficient services to end-users and enhance the quality of experience. The
edge computing provide a promising platform for smart cities, it could decrease the time
for data transmission and organize the network structure. Both pre-processing and decision
could be made in the edge layer which lead to decrease the latency compared to the use of
the centralized cloud only that is far from end-users which increase the latency [36].
• Location Awareness: Most smart cities applications such as intelligent transportation
systems and utility management are geographic-based applications in nature. The edge
computing provide efficient platform for the location awareness in smart cities by collecting
the data based on geographic sites which increase the QoS and prevents the data access delay,
as well as providing the transparency of the transferred information [36].
(b) Smart Home Systems: The smart home [53] allows controlling intelligent smart
devices inside homes including fridge, cooker, air-conditioner, TV, etc. The smart services
[54] inside a smart home can be classified into three categories:
• Home Management Services: such as smart control of connected devices such as cooker
and TV.
• Home Automation Services: allows the automation of smart devices such as air-
conditioners, cleaners, etc. And lastly,
• Home Security Services: such as detection of potential crimes, prevention from gas
leakage, and explosions.
The connected devices inside a smart home generate a massive amount of data for
decisions and smart control inside the home. The process and the analysis of the generated
data require many resources and storage which need an efficient architecture to ensure the
continuity and the quality of service. The edge computing provides a distributed architecture
for smart home systems [55, 56] by processing data at the network edge in an efficient way
without any degradation which provides: efficient data processing, low latency, and less
energy consumption.
The edge computing platform aided smart home systems aims to provide an efficient
architecture for future IoT applications, especially real-time applications that require short
20
response time such as online streaming [57] and intelligent surveillance inside the home [58],
etc.
(c) Smart Transportation Systems & Vehicular Networks: A smart vehicle system [59, 60]
consists of a connected vehicle embedded with devices for computing and storage that allow
the vehicle to send and receive data from other vehicles which is known as a vehicle-to-
vehicle network (V2V) also with roadside units known as vehicle-to-infrastructure (V2I), and
communication with other connected devices known as vehicle-to-everything (V2X). The
edge computing provides an efficient platform for smart vehicles system, it can support both
the interactions among vehicles, the high mobility, collisions and obstacles detection, etc.
Different application types can be developed and deployed for smart transportation systems
aided edge computing, including:
• Safety Applications: This type of application can be used to send warnings and noti-
fications about crashes, traffic violations, pre-crash sensing, as well as they could include
sensing of approaching emergency vehicles.
• Convenience Applications: include advice on congestion situations [61], personal
routing, in addition to some incidents such as power failure and network breakdown. In
some scenarios, the smart vehicle can provide SOS and emergency calls, as well as can play
important role in the monitoring of weather and road by sharing information from vehicle
sensors.
• Smart Traffic Lights: smart lights synchronize the smart vehicles by sending warning
messages using wireless networks to share traffic data which can help vehicles in different
situations such as warning messages when approaching the pedestrian corridor, as well as
the different problems related to the traffic light [62].
• Smart Parking Systems: Most big cities face traffic congestion, consequently, finding an
empty place in parking is expensive and difficult [63]. To overcome this issue, it is important
to build smart parking connected directly with smart vehicles to automatically check available
places inside the parking. The parking system sends notifications messages using a wireless
network to inform smart vehicles if there are empty places inside the parking. This aims
peoples to gain both time and effortlessly find a place in the parking.
• Commercial Applications: represents paid location-based services [64] such as adver-
tisements and entertainment, as well as services of diagnostics for vehicle problems and
21
updates of social networks. These services are provided by companies by offering special
sensors or applications installed in the vehicles which can connect directly to the edge server
of the associated company.
The edge computing provides an efficient platform that offers different benefits to smart
vehicles and intelligent transportation system. The edge servers are localized nearby to
connected vehicles, which provide efficient and rapid service to enhance the quality of
service in smart cities.
(d) Smart Grid Systems: [65] composed of smart meters in different locations used to
measure the status of electricity distribution. This information is analyzed by a server called
SCADA [66], that manage the system by sending information about emergency and responds
to different change requests to protect the power grid and stabilize the system.
The edge computing can be integrated with smart grid systems [67] which may offer great
services to enhance the efficiency of this smart system. With the edge computing paradigm,
the power generators (e.g., wind farms, solar panels, etc.) can be integrated with the main
power grid which offers efficient control of the power networks. Moreover, the SCADA can
be linked with the edge which can improve the security [68] and the network cost [69].
The edge computing aided smart grid will turn into a hierarchical system (multi-tier
architecture) that aims to create interactions among different SCADA [70]. In this system,
the edge layer manages the micro-grid, and exchange data and information with neighboring
edge servers and the higher tiers.
(e) Smart Healthcare Systems: Remote Patient Monitoring (RPM) [71, 72] aims to
create smart system allows remote patients monitoring. This system is composed of three
components: data acquisition module, diagnostics module, and visualization module. To
get data, the sensors are deployed on patient (e.g., blood glucose sensor). The first step
is acquiring the data, then it will be sent to the diagnostic module (processing unit) for
processing. And finally, the visualization module display the calculated analytic. In healthcare
system [13], the data management is an important issue because the patient data contain
private and important information that need to be processed in a secure and an efficient way.
The edge computing system provides efficient platform for smart medical systems which
is one of the important technology in the Internet of things ecosystem. The patient data
22
analysis requires short response time and security, etc., which can be provided only by
efficient platform and full distributed system offered by the edge computing architecture.
2.1.5 Mobile Edge Computing (MEC)

The path toward defining the next generation of cellular networks systems encompasses
all the requirements of the communication network to cope with massive data generated
that needs efficient processing so as to allow operators to offer services and provide content
efficiently and profitably. Driven by the visions of cellular networks and the Internet of
things, recent years have seen a rapid increase in computing requests due to the development
of applications which led to a paradigm shift in mobile computing, from mobile cloud
computing to mobile edge computing.
Overview
Mobile edge computing (MEC) [73, 74] also know as Multi-access Edge Computing in-
troduced by the ETSI ISG (European Telecommunications Standards Institute, Industry
Specification Group) standardization organization [75, 76]. It provides a combination of
computing on a radio access network (RAN) on the network edge. The MEC aims to
push storage, mobile computing, and network control to the edge of the network (e.g., ac-
cess points and base stations) so as to provide resources especially for latency-critical and
computation-intensive applications. MEC promises a spectacular reduction in mobile energy
consumption and latency, tackling the main challenges for materializing the new generation
of cellular networks vision. The promised benefits of MEC have motivated global efforts in
both industry and academia on developing and deploying this technology.
MEC Framework & Architecture
The MEC framework (shown in figure 2.7), displays the involved high-level functional
entities. These entities are grouped into the network-level, system-level, and host-level
entities.
The network-level contains related entities are the external networks, local networks,
and the 3rd Generation Partnership Project (3GPP) cellular network. This layer allows
23
connectivity to cellular networks, local area networks, and external networks such as the
Internet.
The ME host-level consists of the management entity and the ME host that also includes
the virtualization infrastructure, the ME platform, and ME applications.
The top-level is the ME system-level that has the overall visibility to all ME system. This
layer consists of ME management to run applications on the operator network.
ME System
UE
Level
ME System-Level Management
Third
Party
ME App
ME Host Level
ME
ME App
Platform
ME ME
Applications Host-Level
Management
Virtualization
Infrastructure
ME
Host
Networks
3GPP Local External

Network Network Network
Fig. 2.7 Mobile Edge Computing Framework.
The reference MEC architecture defined in figure 2.8, represents a thorough understanding
of the mobile edge (ME) system. It defines in detail the functional entities with the relations
between them. The ME entity is a host consisting of the ME platform and the network
resources, storage, and computing virtualization infrastructure. Besides, the ME host can
offer permanent storage and real-time information related to applications. The data plane
included in the virtualized infrastructure routes traffic between the services, networks, and
applications, as well as forward the rules received by the ME.
The ME platform consists of a collection of functionalities that are important to run
applications on ME and to enable ME applications to advertise, offer, and discover the ME
24
services. Both the applications and platform provides ME services, while the applications
and the platform use ME services.
In the ME platform, a local domain name system (DNS) server/proxy can be configured
to direct the user requests to selected ME applications. Many ME platforms can communicate
with each other or grouped together to form a communication grid through an Mp3 interface.
On top of the virtualization infrastructure in the ME host, the ME applications are running
as virtual machines (VMs), where the applications communicate with the platform through
an Mp1 interface to use the services offered by the platform. The Mp1 interface is also
used to prepare the relocation for the application state in the case of handover, in addition,
indicating the application availability. Each ME application has its own resources or services
needed, the system level validates these requirements, then, the selection of ME host(s) is
performed based on the availability of application requirements in the host(s).
The ME platform manager (MEPM) consists of ME application lifecycle management,
ME application rules management functions (AMF). and ME platform element management
(PEM). The ALM provides the application events to the ME orchestrator (MEO). The AMF
includes DNS configurations, traffic rules, authorizations, as well as resolving issues in the
case of conflict between rules. The reference point between the MEPM and the ME platform
is to manage the relocation of applications, configure the rules and the platform, and support
the application lifecycle.
Strategic Relevance of MEC
MEC led to an evolution of telecommunications networking, and mobile base stations by

shifting the load of the cloud to local servers in the network edge. Which enhancing the
quality of experience (QoE) for end-users and helps reduce both latency and congestion
on mobile networks. MEC allows applications to benefit from local computing resources,
and real-time information related to the network. By deploying various caching content
and services near mobile devices, the core of the networks can efficiently serve applications
purpose which will act for interesting revenues for the industry including vendors, operators,
and third-parties. Use cases include:
1. Connected Autonomous Vehicle Functions
25
CFS Mx1 Operations Support
ME System
Portal Mm8 System
User
Level
UE
App Mm1
App Mx2
LCM Mm9
Proxy ME
Orchestrator
Mm2 Mm3
Mm4
Mp3
Other ME ME Platform
ME Host
Platform Mp1 ME Management
Mp1
Level
Platform
Mm5
ME ME ME
App App App Mm6
Mp2
Virtualization Mm7 Virtualization
Infrastructure Infrastructure Manager
Other ME
ME Host Host
Fig. 2.8 Mobile Edge Computing Architecture.
Used to share information about positions of cars, animals, pedestrians, as well as

the weather conditions and the road infrastructure directly to the vehicles. instead
of interact with the central cloud, the combination of MEC with AI/ML will allow
autonomous vehicles to be real-time synchronized with their surrounding situations
and critical changes such as accidents, and incoming emergency vehicles, etc. The
MEC provides low latency to autonomous vehicles system in order to operate safely
which is important and critical as vehicles cannot wait for the data to be processed in
the cloud.
2. Virtual Reality (VR), Augmented Reality (AR) and Mixed Reality (MR)
The AR/VR can be used by the workers to repair and conduct maintenance tasks
remotely. The MEC would provide an efficient platform that offers an overlay of
important information related to a particular problem they are required to repair. The
workers can fixe the problem on the screen field display on mobile devices. The 3D
models require lot of resources because are too heavy so cannot be done in both the
26
remote cloud and the mobile device, as the latency is high. The MEC allows rendering
the broken down device as 3D models of as well as process the data near to end-users
(workers) devices, which allows a remote expert to comment the displayed model as
well as increase the number of digital models displayed on the screen of the worker
device.
Another use case for MR assisted MEC for enterprise application, is a collaboration of
multi-user for construction, architecture, and engineering teams. It allows real-time
cooperation by rendering 3D models for projects that will be done by the team on
jointly in the network edge. This reduces the latency, especially for the 3D models
that are often massive files. The MEC system allows easy collaboration between
stakeholders, over a distributed network that offer efficient resources near to team
devices.
3. Cloud Gaming and Multiplayer Gaming
With MEC, the intensive graphics/compute processing would be moved from the
gaming console to the edge of the network. This allows gamers to play different games
with high graphics quality from a near edge server from anywhere and any time without
the need to massive dedicated resources in their gaming console. Since the MEC offers
a very low latency, cloud gaming will give to audience and developers a wider access
to gaming studio and enjoy high gaming experiences, as well as create a potentially
new income models for emerging companies in the field of electronic games.
4. Real-time Unmanned Aerial Vehicles (UAV) Detection
The UAVs are getting a lot of attention from both researchers and developers because
of its development in different field such as security, data transmission, packages
delivery, etc. There is an important increasing necessity for innovative solutions that
can detect when an UAV has entered a geo-fenced zone or a secure zone to trigger
alarms to inform the teams that secure the site. As well as, this technology can be
used in hospitals and prisons because the respond to threats must be immediately in
these locations, which is provided by the MEC system. The use of MEC offers a lower
latency for detecting a foreign UAV and get its path to show if its close to an exclusion
27
zone, where the MEC keeps the UAV data closer to its source, decreasing the taken
time to react for any threat or a security breach.
5. Video Analytics
The video surveillance is one of the most systems used in smart cities such as the
surveillance of road, homes, markets, and enterprises, with the massive increase of
data volume due to both video quality and the number of cameras. MEC enables data
analysis in the edge of the network instead of send the videos traffic to remote central
processing. This allows the collect of video streams from different cameras types, as
well as, enable analytics of other video applications such as facial recognition. MEC
decrease the volume, cost and time it takes to send the video chunks from the camera
to the edge server, and allows real-time triggers during the analysis.
Table 2.1 outlines the main characteristics of the cloud computing and mobile cloud
computing compared with edge computing and mobile edge computing.
CC MCC EC MEC
Private
Ownership entities Private entities, companies, individuals
Network
Deployment core Network edge, devices
Servers,
user
Hardware Servers devices Heterogeneous servers
Service Virtualization
Net.
Architecture Centralized N-tiers, decentralized, distributed
Mobility Yes
Latency Low Average
Local
awareness Yes
Availability High
Scalability Average High
Table 2.1 CC & MCC vs EC & MEC
This part has presented an overview of cloud, edge, and fog computing for the Internet
of things. First, we detailed the IoT system and architecture. Then, we detailed the cloud
28
2.2 Edge Computing for IoT: Related Works
computing technology by presenting the necessary components of the cloud system and
mobile cloud computing mechanism. Further, we surveyed the edge computing technology
by presenting a detailed overview of edge computing and mobile edge computing for IoT
with its innovative strategies and solutions provided to next-generation networking with the
different applications in smart cities, smart homes, smart vehicles, and healthcare.
The state of the art will be the following step in the next part to fully understand the
problem and to have a detailed idea of the solutions proposed by different researchers around
the world.

The integration of edge computing with the Internet of things (Edge-IoT) faces many issues
and challenges related to QoS. For this, different research directions related to Edge-IoT have
been studied to cover all communication models to satisfy the applications’ requirements
and provide efficient services to end-users. This chapter presents the state of the art for two
Edge-IoT research directions, service offloading and vehicular edge computing.
2.2.1 Service Offloading In EC

Service offloading refers to uploading some computationally intensive services to the edge
servers for processing. Resource allocation for these uploads refers to allocating some
computing resources on edge servers depending on the requirements of the service. Moving
services to the network edge facilitates storage, service delivery, content caching, and IoT
management, which allows for improving response times and transfer rates, in this way
ensuring that users have the fastest and the best service possible. Service offloading is an
important network optimization task that still painstakingly tune heuristics to get sufficient
solution. These algorithms use data as input and output near-optimal solutions. Service
offloading in this novel concept is a complex task. It has to design efficient strategies to
optimize the service offloading and allocation decisions. Despite the huge effort in ETSI
research group, service offloading functionalities still suffer from the legacy heuristics used to
design optimal resource management (orchestration, placement, scheduling, etc.) algorithms
29
and place concurrent slices within virtual network operators (VNO). In the following, we
present a review of the existing solutions for service offloading in edge computing.
Miluzzo et al. [77] proposed mClouds, a mobile computing architecture where submitted
tasks are executed on smart mobile devices that form a cooperative computing platform.
Although he discussed the different steps of tasks execution and mClouds management, the
authors ignored the mobility impact which is missed in their work. Moreover, both problem
formulation and evaluation results are not presented in this work.
Fahim et al. [78] proposed a task offloading environment to execute tasks in mobile
devices. Real mobile phones are used to validate the proposed system. The evaluation results
prove the efficiency of the proposed system where the potential gain is high in both energy
consumption and processing time. However, the proposed system and its formulation is
limited to small network scale. Moreover, the impact of device mobility is not studied in this
work.
Van et al. [79] proposed a workload offloading based on Markov decision process
(MDP). The mobile cloudlets are used to process tasks by considering the impact of channel
properties, distance and cloudlets mobility. The obtained results show that the efficiency of
the proposed work. However, the formulation should be enhanced with energy and mobility
parameters to model the transition behavior of the MDP algorithm.
Hasan et al. [80] proposed Aura; a mobile adhoc cloud (MAC) computing architecture
for IoT task offloading. The framework consists of three main components: i) Mobile Agents
(MA) that includes smartphones and other end-user smart devices. They are used to run the
tasks that are needed to be offloaded and executed in the Aura. ii) IoT devices that form
a part of the global Aura architecture. It forms the ad-hoc cloud, and iii) Controllers that
form the link between IoT devices and MA. They ease the creation and the distribution of
sub-tasks, the programmability of IoT devices, the MA action selection strategies, etc. It is
worth mentioning that the obtained results demonstrate the efficiency of the Aura architecture
in MAC cloud computing fields in terms of memory/CPU usage, cost of execution, and task
completion time. However, the impact of IoT devices mobility (and energy) is missed in
their work.
Van et al. [81] proposed a deep reinforcement learning (DRL) approach for tasks
offloading in adhoc mobile cloud environment. The MDP theory is used to find tasks
30
processing actions. Then, the Deep-Q-Network (DQN) is used to find the efficient solution
for the MDP model. The evaluation results show that the proposed model is efficient based
on different metrics. However, the prediction of the dynamic parameters used in the MDP
formulation is missed in this work.
In [82], the authors proposed an offloading technique in mobile adhoc cloud environment
based on constrained Markov decision process (CMDP). An Q-Learning and LP schemes
are proposed to find efficient solution of the CMDP model. The LP to find the optimal
solution, while the QL is used to receive the offloading decision. The obtained results show
the efficiency of the proposed schemes. However, the application of the proposed works in
real environment is missed in this work.
Alam et al. [83] proposed a computation offloading solution in mobile fog/edge comput-
ing. They consider access points and controllers as on-demand fog nodes. A framework based
on deep Q-learning is proposed to handle the resource demand for computation offloading.
The authors prove the efficiency of the proposed solution in terms of computation offloading
cost and service latency. However, the work missed the impact of the IoT devices mobility
on the computation offloading decisions. Moreover, we believe that the IoT devices energy
should be considered to deal with resource constrained devices.
Zhang et al. [84] proposed an optimal task offloading scheme based on deep Q-learning
in heterogeneous vehicular edge computing networks. First, the authors used the intrinsic
reward variable to minimize the utility of task offloading. Then, in order to guarantee the
offloading reliability in the presence of tasks transmission failure, they proposed an adaptive
redundant offloading algorithm. The overall approach guarantees the offloading process
reliability and enhances the system utility. Further, they used a real traffic trace to evaluate
their solution. The analytical results prove the efficiency of the proposed solution in terms of
reliable offloading and latency as well as tasks offloading with optimal utility. However, the
authors did not provide details about the implementation of the proposed solution, and its
applicability in the real-world.
Wang et al. [85] proposed an optimal data offloading algorithm to ensure the resource
provider privacy and achieve efficient resource usage. In addition, the proposed model
takes into account the maximization of the resource buyer benefits and the minimization of
resource provider cost. The obtained results prove the feasibility of the proposed solution and
31
its efficiency to identifying malicious resource providers. However, this work did not provide
any details about the adaptation of this solution in a real mobile Adhoc cloud environment. In
addition, this work needs to address important issues in the mobile Adhoc cloud such as the
resources management of devices and the problem of offloading failure caused by obstacles
and non-existing devices.
Lin et al. [86] proposed Circa, a framework for code offloading between a set of neighbor
mobile devices. They used an indoor positioning system (iBeacons) for transmitting in
short-range with lower energy and cost for sharing tasks among devices without the need for
a centralized server. in addition, task allocation algorithms are proposed in order to select the
set of reliable collaborators from the neighbors’ mobile devices and offload efficiency the
tasks among the selected devices in a fair fashion. The experiment results show that Circa
reduces the execution time and preserving the mobile application performances. However, the
impact of obstacles on offloading is not studied in this work. In addition, the communication
range is very low which decreases the selected devices which by consequence leads to task
failure caused by non-existing devices.
Saha et al. [87] developed a mobile device cloud framework where the resourceful and
reputed worker devices execute the different application tasks. The user devices request
service offloading for their application (set of tasks). After getting the willingness from
the worker device, the cloudlet selects the optimal task mapping to minimize the cost and
maximize the quality of experience (QoE). The results prove the efficiency of the proposed
work. However, the evaluation in real-world large-scale networks is missing in this work.
Roy et al. [88] Proposed a task allocation framework for mobile device cloud. In
addition, a workers auction mechanism is proposed to select buyer devices. Moreover,
authors introduced the federated learning and the multi-weight subjective scheme to check
the reliability and the trustworthiness of worker devices’. The simulation results prove the
efficiency of this work in terms of buyer utility and QoE. However, the evaluation in dense
networks is missing in this work.
Table 2.2 shows the summary of the existing works for service offloading in EC.
32
Paper Main Idea Problem Edge Po- Advantages Limitations

Formula- sition
tion
Miluzzo et Tasks ex- Some equa- Mobile Execute Long re-
al. [77] ecution in tions for tasks for sponse
smart mobile service other users. time.
devices. payment,
etc.
Fahim et al. Task offload- No mathe- Mobile Efficient The mobility
[78] ing using matical for- energy impact is not
mobile devices mulation is consump- studied.
as computing used. tion and
nodes. processing
time.
Van et al. Tasks process- Markov Mobile Enhance The imple-
[79] ing in mobile decision time, and mentation
cloudlets. process total energy details is
(MDP). consump- missing.
tion.
Hasan et al. Aura, a task Equations Mobile Enhance ex- The mobility
[80] offloading for cost, ecution cost impact is not
in mobile time, and and time. studied.
ad-hoc cloud resources
architecture. measure-
ment.
33
Van et al. Tasks of- - Markov Mobile Enhance The eval-

[81] floading in Decision energy con- uation in
an ad-hoc Process sumption, large-scale
mobile cloud (MDP). payment, networks is
environment. - Deep-Q- and delay. missing.
Network
(DQN).
Van et al. Offloading Markov de- Mobile Enhance task The evalua-
[82] technique in cision pro- loss ratio, tion in real
mobile ad-hoc cess. average environment
cloud. delay, and is missing.
energy con-
sumption.
Alam et al. Computation - Markov Mobile - Mini- The mobility
[83] offloading in Decision mize the impact is not
mobile edge Process computing studied.
computing. (MDP). latency.
- Deep
Q-learning.
Zhang et al. Optimal task Deep Q- Fixed Reliable and The com-
[84] offloading learning optimal task plexity is not
in a mobile approach. offloading. discussed.
environment.
Wang et al. Optimal of- Distributed Mobile Identify The Imple-
[85] floading model opti- malicious mentation
in mobile mization resource detail is
Adhoc cloud algorithm. providers. missing.
environment.
34
Lin et al. Circa, an Tasks allo- Mobile Enhance the High risk
[86] offloading cation algo- execution of commu-
framework rithms. time. nication
among mobile failure.
devices.
Saha et al. Mobile de- Optimal Mixed Low cost. The eval-
[87] vice cloud model and uation in
framework. greedy large-scale
algorithms networks is
for tasks missing.
provision-
ing.
Roy et al. Task allocation Task al- Mobile Efficient The eval-
[88] framework for location both buyer uation in
mobile device. algorithms. utility and large-scale
QoE. networks is
missing.
Table 2.2 Summary of the existing works for service offloading in EC.
Service offloading algorithms in the literature are based on formulating exact models to
measure the optimal behavior of such a solution. However, in practice, network operators
always design intelligent heuristics to deal with scalability problems in large-scale networks
with huge amounts of data. Currently, these heuristics are near-optimal solutions and
sometimes are very far from optimal. Further, heuristics are unused when system input
(e.g. demand matrix, etc..) changes and gives wrong and fatal errors when unexpected input
arrives in the system.
2.2.2 Vehicular Edge Computing

Smart vehicles are considered as mobile devices equipped with sensors, having the capability
of data collection, computation, and communication. The information is collected from both
35
in-vehicle sensors and the external environment. Edge computing can provide an efficient
and scalable architecture for vehicular networks by improving data processing and traffic
in real-time. We present in the following parts a review of the existing edge and fog-based
solutions for vehicular networks.
V2X and EC Convergence
The V2X model allows the passing of information from the connected vehicle to any other
connected devices including other vehicles, UAV, end-user devices, edge servers, cloud,
etc. Also, it incorporates other types of vehicular communication such as V2N (vehicle-to-
network), V2P (vehicle-to-pedestrian), V2D (vehicle-to-device), V2V (vehicle-to-vehicle),
V2I (vehicle-to-infrastructure), and V2G (vehicle-to-grid). The main motivations for the
V2X model are traffic efficiency, energy savings, and road safety. In the following we present
the existing works for V2X assisted Edge computing.
Kim et al. [89] presented parking management in a vehicular network based on the
Fog-Cloud environment. The proposed model consists of many parking slots, the status of
slots inside the parking lot (i.e. reserved or vacant) are sent to the fog server that is located
at different areas (e.g., shopping malls, restaurants, banks, etc). The fog server provides
information of their managed vacant spaces to the RSUs. At each RSU, the roadside cloud
and fog computing communicate with the fog servers aiming to direct drivers to the optimal
space. The performance evaluation shows that this approach is reliable for finding a vacant
slot. However, this approach requires a scalable fog system.
Zhang et al. [90] proposed a task offloading mechanism in a vehicular edge computing
framework. They formulated the proposed mechanism as a stackelberg game, and developed
a distributed algorithm in order to obtain optimal strategies of the vehicular edge computing
servers. In the proposed algorithm, each vehicular edge computing server (VEC) starts
selecting the selling price and the amount of resources to purchase from the backup computing
server (BCS) where both the price and the resources are limited in an interval. Numerical
results show that the proposed mechanism guarantees an optimal delay of the computation
tasks. However, this work did not take into account the risk of link failure between vehicles
and RSUs.
36
Zhu et al. [91] proposed a solution for task allocation in vehicular fog computing, namely
Fog Following Me (Folo). They formulated the task allocation as an optimization problem
and solved it using mixed integer linear programming. The Folo composed of two types
of fog nodes: stationary fog nodes located with WiFi access points and RSUs, and mobile
fog nodes that are mobile edge nodes carried by vehicles. Moreover, there are other types
of vehicles, named client vehicles, which generate tasks and send them to the fog nodes.
Furthermore, a service zone is used in the Folo system as a service coverage zone inside the
city. Although The obtained results show that the proposed solution reduces the latency, the
proposed solution for task allocation requires computation resources.
Gillam et al. [92] Proposed a distributed computing architecture for off-vehicle and
on-vehicle computation to support autonomous and connected vehicles. They suggested
a local computation in the edge rather than the central cloud to reduce the latency. The
evaluation is missing in this work to validate the efficiency of the proposed architecture.
Caminha et al. [93] proposed SensingBus, a smart city architecture using buses and a fog
system. The SensingBus architecture contains three levels: (i) sensing level that is composed
of sensing nodes installed in buses, allows the gathering of data about the city, registers the
geographic coordinates from the GPS; (ii) fog level that consists of nodes installed in strategic
locations around the city, e.g., bus stops, these nodes are responsible for pre-processing data
that has been received from the sensing nodes and forwards it to the cloud; and (iii) cloud
level that is considered as the final destination of collected data and consists of nodes, each
node contains a set virtual machine. The performance analysis of the proposed prototype
proves the enhancing of memory and CPU utilization. However, SensingBus is limited
because the buses cover only some areas in the city, and can-not provide wide coverage.
In [94], authors studied connected vehicle-based congestion identification techniques.
Computing strategies for vehicular networks are studied to develop feasible solutions to detect
congestion that performs efficiently for multiple scenarios with large coverage. The authors
expected that the designed system can detect congestion to achieve traffic management
objectives that are important in transportation systems. However, the proof of work is
missing in this study.
Ning et al. [95] proposed a three-layer traffic management model in a vehicular fog
environment to minimize the response time of the events that are reported and collected
37
in the city by vehicles. Moreover, an offloading scheme using optimization techniques is

formulated by leveraging parked and moving vehicles as fog nodes. The obtained results are
evaluated based on real-world taxi trajectory data. However, the vehicle’s mobility impact is
not studied in this work.
Moubayed et al. [96] proposed a V2X service placement in a mobile edge computing
environment. An optimal model is formulated for service placement. Hence, to deal with
dense networks they proposed a heuristic algorithm that offers efficient service placement
with less complexity compared with the optimal model. The numerical results validate
the efficiency of the proposed approach in terms of delay/latency and resource utilization.
However, the impact of vehicle’s mobility is not studied in this work.
Similarly, Shaer et al. [97] targeted V2X service placement in an edge computing
environment. An optimal model is formulated to minimize the delay and satisfying the
requirements of V2X services. The obtained results prove the soundness of the proposed
approach in terms of delay. However, important metrics are missing in this work such as
resource utilization.
In [98], authors proposed a real-time autonomous driving system in the edge computing
environment. an offloading strategy is proposed to decide where and when to offload the
tasks of autonomous driving. The authors considered only the edge power consumption
minimization. However, more important metrics are not studied in this work such as resource
utilization and QoS.
Hameed et al. [99] proposed a dynamic clustering approach in vehicular fog environment
by creating clusters based on the vehicles’ network. Further, a load balancing technique
is used for inter and intra-cluster for IoT jobs. The division of vehicle networks into
clusters increases the efficiency of the resource management, as well as provides an efficient
distribution of job load among vehicles’. The performance evaluation proves the efficiency
of the proposed approach in terms of QoS and energy. However, the authors used only one
mobility model in this work.
Table 2.3 shows the summary of the existing works for V2X-assisted EC.
38

Formula- sition
tion
Kim et al. Parking man- Many- Fixed Reliability to The scalabil-
[89] agement in to-one find available ity of the fog
fog-cloud match- parking slot. system is not
environment. ing game studied.
model.
Zhang et al. Task offload- Stackelberg Fixed Optimal The risk of
[90] ing in mobile game. delay for link failure is
environment. tasks compu- not studied.
tation.
Zhu et al. Task allocation An exact Mixed Enhance the Requires
[91] mechanism model end-to-end high com-
in mobile for task latency. putation
environment. allocation resource.
problem.
Gillam et al. Distributed N/A Fixed Reduce No eval-
[92] computing latency. uation is
architecture presented.
for vehicles
Caminha et SensingBus, a N/A Mobile Enhance The buses
al. [93] data collecting CPU and can not
system from Memory provide wide
sensors on utilization. coverage.
urban buses.
39
Thakur et Connected N/A Fixed System The proof of

al. [94] vehicle-based can detect work is miss-
congestion congestion ing
identification to achieve
techniques traffic man-
agement
objectives
Ning et al. Traffic man- Exact of- Fixed Reduce the The vehicle’s
[95] agement floading response mobility im-
model in a model time pact is not
vehicular fog studied
environment
Moubayed V2X service Optimal Fixed Enhance de- The impact
et al. [96] placement in V2X lay/latency, of vehicle’s
a mobile edge service and resource mobility is
computing placement utilization not studied
environment model
Shaer et al. V2X service Optimal Fixed Minimize de- Resource
[97] placement model for lay utilization
in an edge service metric is not
computing placement studied
environment
Tang et al. Real-time au- Task of- Fixed Efficient Resource uti-
[98] tonomous driv- floading resource lization and
ing system in model utilization QoS metrics
the edge com- and power are not stud-
puting environ- consumption ied
ment
40
Hameed et Dynamic Optimal Mobile Enhance One mobility

al. [99] clustering model for QoS and model is
approach in network’s energy used
vehicular fog utilization consumption
environment
Table 2.3 Summary of the existing works for V2X-assisted EC.
Current convergences between V2X and EC technology should take into consideration
safety metrics and real-time reactions in critical situations. Therefore, despite the existing
works in EC-assisted vehicular communications, next-generation applications such as road
safety and autonomous driving require efficient deployment to achieve ultra-low latency in
dense vehicular networks.
UAV and EC Convergence
Unmanned Aerial vehicles (UAVs) are introduced in almost wireless communication fields
to aid the EC technology due to their deployment, mobility, and cost-effective. The UAV
meets EC to increase the efficiency of the offered services to end-users to achieve ultra-low
latency especially for new-generation of applications. Next, we overview the existing works
for UAV-aided EC technology.
Zhou et al. [100] tackled the UAV-enabled mobile edge computing system where the
UAVs are used as mobile edge servers to offers both offloading services and energy to mobile
end-users. The results prove that the proposed solution achieves efficient energy consumption.
However, the mobility impact for both UAVs and mobile end-user devices is not studied in
this work.
Hu et al. [101] proposed a UAV-enabled mobile edge computing where mobile end-users are
served by UAVs that have more computing resources. The user tasks can be offloaded to the
UAV or served locally depending on the task type and its requirements. The results prove the
efficiency of the proposed approach. However, the mobility of both UAVs and end-users is
not studied in this work.
Similarly, Xiong el al. [102] proposed a UAV-assisted edge computing solution where UAV
41
equipped with edge server and employed as a mobile base station to serve mobile end-users
by offering tasks processing service. The performance evaluation demonstrates the efficiency
of the proposed approach. However, the mobility is not studied in this work.
Zhan el al. [103] proposed a UAV-assisted edge computing solution where UAVs are con-
sidered as edge servers to serve IoT by offering computing resources to process tasks. The
obtained results show that the proposed solution is efficient by ensuring the trade-off between
completion time and energy consumption. However, the impact of mobility is not studied in
this work.
Wu el al. [104] proposed a UAV-assisted edge computing architecture where UAVs are
used as edge servers to accomplish IoT tasks. The obtained results prove the feasibility and
the efficiency of the proposed solution in terms of energy consumption. However, more
evaluation metrics need to be considered in this work.
Yang el al. [105] proposed a UAV-aided mobile edge computing system where a group of
UAVs cooperates to provide service as edge server to IoT by considering both tasks pro-
cessing and load balancing. Simulation results prove the efficiency of the proposed solution.
However, both UAVs’ energy consumption and computing metrics are not studied in this
work.
Deng el al. [106] proposed a tracking system based on edge computing where UAV track
target vehicle inside the city and send the captured video to the edge server. Moreover, when
the target vehicle is out of the UAV coverage, it stays tracked by the camera deployed inside
the city. The results show that this approach is efficient in terms of tracking timeliness and
reliability. The UAV energy consumption is not studied in this work.
In [107], authors studied the integration of UAVs with the edge computing environment. The
video captured by UAVs is offloaded to the edge server for processing. The proposed frame-
work is efficient since the UAV intensive computing is offloaded to the edge which decreases
resource utilization in the UAV, consequently maintain energy for a long time. However,
UAVs require real-time processing with ultra-low latency which is an open challenge.
Table 2.4 shows the summary of the existing works for UAV-assisted EC.
42

Formula- sition
tion
Zhou et al. End-user Optimal Mobile Efficient The mobility
[100] devices com- model for energy con- of both UAV
putation energy, re- sumption. and users is
offloading in source, and not studied.
UAV-enabled trajectory.
MEC system.
Hu et al. UAV-enabled Optimal Mobile Minimize la- The mobility
[101] MEC system model tency. of both
to serve mobile for tasks UAVs and
end-users. offloading. users is not
studied.
Xiong el al. UAV-aided Optimal Mobile Efficient The mobility
[102] edge comput- model for energy con- impact is not
ing to serve energy and sumption. studied.
IoT devices. resources
allocation.
Zhan el al. UAV-assisted Optimal Mobile Efficient The impact
[103] edge comput- model to energy con- of mobility is
ing to serve minimize sumption not studied.
IoT devices. resource and resource
utilization. utilization.
Wu el al. UAV-assisted Optimal al- Mobile Efficient The resource
[104] edge com- gorithm for energy con- utilization is
puting to UAV posi- sumption. not studied.
accomplish tion.
IoT tasks.
43
Yang el al. UAV-aided mo- Optimal Mobile Efficient UAV energy

[105] bile edge com- and deep tasks offload- consump-
puting system reinforce- ing. tion is not
to provide ser- ment studied.
vice to IoT. learning
algorithms.
Deng el al. Tracking sys- Set cov- Fixed Efficient UAV energy
[106] tem based on erage tracking consump-
edge comput- and short- timeliness tion is not
ing and UAV. est path and reliabil- studied.
models. ity.
Chen el al. UAV-edge Framework Fixed Decreases Ultra-low la-
[107] integration for au- resource tency is not
for video tonomous utilization. studied.
offloading. UAV and
video
offloading.
Table 2.4 Summary of the existing works for UAV-assisted EC.
Future network generation architectures based on UAV-EC require intelligent optimization

algorithms that decide the optimal point of operations where UAV should be deployed in
dense networks without neglect the energy factor that represents the important issue in UAV
networks.
Conclusion
This chapter presents an overview of IoT, Cloud Computing, Mobile Cloud Computing,
Edge Computing, and Mobile Edge Computing with the need of these technologies for
next generation applications to satisfy the users requirements and enhance the Quality of
Service. Moreover, in the second part, I presented the state of the art for both studied research
44
directions, service offloading and vehicular edge computing. In the following chapter, we
presents our optimal proposed models for service placement in edge servers, maximize
the vehicles coverage, and service offering to resolve different issues at the network edge
including latency, computing resource utilization, energy consumption, etc.
45
Chapter 3
Optimal Solutions For Edge Computing

Networks
The network optimization procedure is still an important task that needs to be carefully
studied due to rapid networks evolutions and user demands. In this chapter, we present our
exact models for different use cases in service offloading and vehicular edge computing
research directions. For each use case, we present a detailed optimal model with the used
constraints, objective function, and variables.
3.1 Service Offloading In EC

With the rapid evolution of network generations and applications requirements, service
offloading is considered as one of the most important network tasks that allow to offload
resource-intensive services to optimal computation hardware in both edge servers and the
central cloud for processing. This complex task requires efficient communication, optimiza-
tion strategies, and virtualization technologies to deal with different services and provide
efficient QoS for users. In the following, we present our optimal algorithms for service
offloading in EC networks.
47
Optimal Solutions For Edge Computing Networks
3.1.1 UseCase 1: 5G VNF Slices Placement in EC

ETSI recently expected the benefits of MEC architecture to run all of its applications and
entities as VNFs in an NFV environment [108].
The cloud VNF placement issue has gained more attention. It is similar to the placement
of virtual machines (VMs), where VNFs are made up of VMs or containers that can perform
network functions. Indeed, the placement of VNF at the edge of the network offers many
advantages, in particular the proximity of the end-users, which decreases the latency and
improves the quality of the services.
5G network slicing allows multiplexing of virtualized and independent networks, where
each slice corresponds to an isolated network to meet application requirements.
The Proposed 5G-edge Computing Architecture For VNF Slices Placement
The proposed architecture allows the placement of VNF slices at the edge of the network
in the context of 5G network where the physical network is divided into isolated logical
networks called slices, each slice corresponds to a type of application including augmented
reality (AR), IoT, unmanned aerial vehicles (UAVs), and so on.
Figure 3.1 shows the proposed 5G-edge computing architecture for VNF slices placement.
Access point Edge Server VNF

ce
Router
Sli
Cloud
intermediate
8:30
nd
yournamehere
node
ba
VNF
d
Liked by yourfriend, yourfriend and 20 others

user_name your caption here
#yourhashtag #yourhashtag #yourhashtag
View All 10 Comments
2 HOUR AGO SEE TRANSLATION

oa
yournamehere
Br
ile
4G VNF
SMART HOME
CONTROL SYSTEM
ob
M
ce
Sli
VNF
es
icl
eh
VNF
lV
ri a
Ae
d
VNF
ne
an
m
Un
VNF
e
lic
sS
ng
VNF
hi
fT
to
ne
VNF
r
te
In
Fig. 3.1 VNF Slices Placement in Edge Computing Over 5G Network [7].
48
VNF Slices Placement Model Over 5G
In this part, we present the formulation of the exact model for VNF slices placement in a
5G-edge computing environment. Table 3.1 display the used notations in the proposed model.
The model formulation is as follows:
Table 3.1 The notations used in the proposed VNF slices placement model.
Notations Definition
N The set of servers in the edge.
M The set of vnf for the placement.
wj The number of slices in the vnf
j ∈ M.
ci The number of cpu slots in the
server i ∈ N.
αi Server characteristic fixed by net-
work operators.
Decision variables Definition
yi A binary variable indicate if the
server i ∈ N is used.
xi j A binary variable that assigns the
vnf j ∈ M to the server i ∈ N.
We present in the following the decision variables used to solve the proposed model:
• The binary variable yi indicates if the server i is used or not. It is defined as follows:

1

if the server i is used;
yi = (3.1)
0

Otherwise.
• The binary variable xi j indicates the assignment of the vnf j to the optimal server i. It
is defined as follows:

1

if vnf j is assigned to the server i;
xi j = (3.2)
0

Otherwise.
n
min ∑ αi yi (3.3)
i=1
49
Subject to
m
∑ w j xi j ≤ ciyi, i ∈ N = {0, ..., n} (3.4)
j=1
n
∑ xi j = 1, j ∈ M = {0, ..., m} (3.5)
i=1
yi ∈ {0, 1} , i∈N (3.6)
xi j ∈ {0, 1} , i ∈ N, j∈M (3.7)
The equation 3.3 formulates the proposed objective function. It represents the number of
used servers. Note that each server is characterized by αi which also depends on the server
i and the operator’s policy. The equation 3.4 guarantees that each server cannot exceed its
computing capacity represented in the form of CPU slots. The equation 3.5 ensures that each
VNF can be associated with at most one server.
The formulated problem is NP-hard because of our complex system, thus, the proposed
exact algorithm is difficult to provide a decision for VNF slices placement in dense networks.
Accordingly, efficient heuristic algorithms will be proposed in the next chapter.
3.1.2 Use Case 2: Task Offloading in Virtual Mobile Edge Computing

(VME)
Several edge computing architectures have been proposed by using edge servers in fixed,
mobile, or mixed positions each one offers benefits to the next network generations. We
proposed to use the connected IoT devices in smart cities as a virtual edge server to process
data at the network edge. Each selected device can execute sub-task(s) depending on its
available resources for computing and the sub-tasks requirements.
VME Architecture
Figure 3.2 outlines the proposed VME architecture where IoT devices including drones,
laptops, smartphones, etc. are connected in a smart city. All the devices are connected to an
access point that is connected to a central cloud. When a request comes to the access point
for task execution, it selects a set of IoT devices to accomplish the submitted task according
to the task processing requirements and the resource availability in the connected devices that
50
form a virtual edge server. If the task processing requirements are higher than the available
resources in the IoT devices, the task is forwarded directly to the cloud.
The aim of the network architecture design is to clarify the network communication
scenarios, including the different entities that participate in the communication such as IoT
devices. Thus, the proposed architecture allows to provide data processing in virtual edge
servers made up of connected IoT devices, consequently, improves the efficiency of the
offered services in the network by reducing the computation cost.
Fig. 3.2 Virtual Mobile Edge Computing Architecture [6].
Virtual Edge nodes Placement Algorithm (VEnPA)
We designed an optimal algorithm for task allocation, where the main objective is to maximize
tasks allocation in virtual edge servers.
The proposed algorithm Optimal Virtual Edge nodes Placement Algorithm (VEnPA) give
the optimal solution for tasks placement by allocate IoT devices to execute sub-tasks. The
model input takes the network topology that consists of: IoT devices, network elements, and
the cloud.
1. Problem Definition
First, we overview the system hypotheses, after that, we present the tasks offloading
model based on the exact optimization technique.
51
The submitted task is fragmented into sub-tasks dispatched to the IoT devices (virtual
edge server) to be executed. After that, the received results of sub-tasks are combined
together as the main result of the task.
Moreover, each selected IoT device has its available resources for processing (CPU,
GPU, RAM, and Storage).
2. Mathematical formulation
(a) VEnPA: System parameters

In this part, we present the model parameters and constraints that are proposed to
formulate the optimization model. This formulation defines the optimal location
for task offloading and partitioning.
We present in Table 3.2 the main VEnPA parameters and decision variables.
(b) VEnPA: Decision variables
i. The binary variable x represents the sub-tasks placement, and its offloading
from the client device ud to the virtual edge device ve It is formulated as:

1, if service subtask (ud, k) is placed on the virtual edge ve

ve
xud,k =
0, Otherwise

(3.8)
ii. The binary variable y represents the virtual edge usage. It is defined as
follows: 
1, if the virtual edge ve ∈ V is used

yve = (3.9)
0, Otherwise

(c) VEnPA: Algorithm constraints

The formulation of the exact partitioned and offloading algorithm, Optimal
Partitioned and Offloading (OPO) is as follows:
ve
max ∑ ∑ xud,k (3.10)
ud∈U ve∈V
52
Table 3.2 The notations used in the proposed VEnPA model.
U The set of User Applications.
V The set of virtual edge servers.
K The set of user application
sub-tasks.
Rve Maximum RAM available in
the edge device ve.
Cve Maximum CPU available in
the edge device ve.
Sve Maximum Storage available
in the edge device ve.
Gve Maximum GPU available in
the edge device ve.
rud,k Required RAM for the appli-
cation sub-task (ud, k).
cud,k Required CPU for the applica-
tion sub-task (ud, k).
sud,k Required Storage for the ap-
plication sub-task (ud, k).
gud,k Required GPU for the applica-
tion sub-task (ud, k).
ve
xud,k A binary variable that assigns
the sub-task k ∈ K of user ap-
plication ud ∈ N to the virtual
edge ve ∈ V .
Subject to
ve
∑ xud,k ≤ 1, ∀ud ∈ U, k ∈ K (3.11)
ve∈V
ve
∑ cud,k xud,k ≤ Cve , ∀ve ∈ V (3.12)
ud,k
ve
∑ gud,k xud,k ≤ Gve ∀ve ∈ V (3.13)
ud,k
53
ve
∑ rud,k xud,k ≤ Rve , ∀ve ∈ V (3.14)
ud,k
ve
∑ sud,k xud,k ≤ Sve , ∀ve ∈ V (3.15)
ud,k
ve
xud,k ∈ {0, 1} (3.16)
Algorithm constraints represent conditions that need to be verified. Equations

(3.12), (3.13), (3.14), and (3.15) guarantee that the selected virtual edge device
has enough amount of CPU, GPU, RAM, and storage respectively to execute the
submitted user application sub-tasks. Besides, the constraint (3.11) ensure the
offloading of each sub-task to only one virtual edge device.
(d) VEnPA: Algorithm complexity

The formulated OPO model is NP-hard because of the system constraints, the
number of IoT devices, and the available resources that change over time in each
IoT device. Accordingly, an efficient algorithm will be proposed in the next
chapter.
VEnPA: evaluation of the OPO model
1. Simulation Setup
The performance of the proposed VME architecture is evaluated using three different
tasks (applications) in a simulated environment. We used the simulator EdgeCloudSim
[109] to simulate the VME communication and the offered services. For each appli-
cation we define parameters including CPU, GPU, RAM, storage, and the number of
connected devices. Table 3.3 outlines the VEnPA parameters.
2. Metrics
The performance of the VME system is based on resources utilization. Therefore, we

focused on the measurement of Processing Time, Service Time, Average Edge Devices
Utilization, Failed Tasks, and Failed Tasks due to Edge Devices Capacity.
54
Table 3.3 VEnPA parameters.
Parameter Value
CPU slots (Min/Max) (1/15)
GPU slots (Min/Max) (1/10)
RAM (Min/Max) (Megabytes) (20/2000)
Storage (Min/Max) (Megabytes) (50/20000)
Number of connected devices (Min/Max) (100/1000)
3. Simulation Results
45
Adhoc Mobile Cloud (AMC)
40 Adhoc Mobile Edge (AME)
Virtual Mobile Edge (VME)
35
30
Failed Tasks (%)
25
20
15
10
0
100 200 300 400 500 600 700 800 900 1000
Number of Mobile Devices (Client Devices)
Fig. 3.3 Failed Tasks [6].
4
Adhoc Mobile Edge (AME)
3.5
Service Time (sec)
2.5
1.5
100 200 300 400 500 600 700 800 900 1000
Fig. 3.4 Service Time [6].
Figures 3.3, 3.4, 3.5, 3.6, and 3.7 shows a comparison among the proposed architecture
VME, with both Adhoc Mobile Cloud (AMC) that represent the architecture proposed
55
4
Adhoc Mobile Edge (AME)
3.5 Virtual Mobile Edge (VME)
Processing Time (sec)

3
2.5
1.5
1
100 200 300 400 500 600 700 800 900 1000
Fig. 3.5 Processing Time [6].
100
90
Average Edge Devices Utilization (%)
80
70
60
50
40
30
20
0
100 200 300 400 500 600 700 800 900 1000
Fig. 3.6 Average Edge Devices Utilization [6].
in [77, 78, 80, 81]. And Adhoc Mobile Edge (AME) architecture that allow to create
edge servers based on end-user devices in Adhoc mode.
Figure 3.3, show the percent of failed tasks. From the figure we show that the failed
tasks increase at the same time when the number of mobile devices increase for the
three architectures. For the VME, the increase level is low compared to the AMC and
AME because of the proposed offloading algorithm can select virtual edge server for
each task. Wherein the offloading take a long time in AMC and AME wich leads to
task failure.
56
100
Failed Task due to Edge Devices Capacity (%)

90
80
70
60
50
40
30
20
0
100 200 300 400 500 600 700 800 900 1000
Fig. 3.7 Failed Tasks due to Edge Device Capacity [6].
Figures 3.4 and 3.5, display the service/processing time for task execution. From the
figure we show that the VME service/processing time is low compared to the AMC
and the AME when the number of mobile devices is less than 500. However, when the
number of mobile devices exceed 500, the VME is greater than the AME because the
AME failed tasks is high compared the VME (as shown in figure 3.3) which signify
that the VME execute tasks more than AME this is because the VME take more time.
Beside, in the AMC, the end-user devices are used as cloud which leads to high level of
tasks failure (as shown in figure 3.3) due to insufficient mobile cloud devices or high
tasks requirements.
Figure 3.6 display the Average Edge Devices Utilization, we show that the edge devices
utilization increase when the mobile device increase, because of the submitted tasks,
in addition, the VME has the highest value of edge utilization because it has lowest
failed tasks (as shown in figure 3.3).
Figure 3.7, show the Failed Tasks due to Edge Devices Capacity, we show that the
VME has the lowest failed tasks due to edge capacity because of the proposed VEnPA
model that ensure efficient offloading and placement.
57

(SO-VMEC)
As mentioned before, edge computing allows creating an intermediate layer between end-user
devices and the cloud. This layer consists of access points, servers, routers, switches, etc.
These devices can be replaced by IoT devices that can be used as virtual mobile edge servers
(VMES).
The VMES is an edge server that offers different services such as caching and processing.
It consists of connected IoT devices such as connected vehicles, UAVs, tablets, and smart-
phones, etc. When client devices require computation service, it offloads the VNF to the
access point, after that, the VNF is split into slices. Then, the access point selects a set of IoT
devices that satisfy the slices computing requirements including GPU, CPU, RAM, Energy,
and Storage. These selected devices form a VMES. If the slice computing requirements are
high compared with available resources in IoT devices, the slice is offloaded to the cloud.
After computation, the results are sent back to the access point that combines the results to
form the main result for the VNF. And finally, the result is forwarded to the source device.
The VME can reduce latency, network delay, and the total cost. Further, it guarantees services
with a short response time which is required in the new generation of applications.
SO-VMEC Architecture
The proposed SO-VMEC architecture consists of three layers as follows:
1. Cloud Layer: It is the top layer allows to process VNFs that require high processing
resources.
2. Edge Layer: It is the medium layer composed of connected IoT devices. It provide
processing services at the network edge according to the available resources in IoT
devices.
3. End-users Layer: It is the bottom layer, represents the clients devices that require
processing due to insufficient computing resources.
Figure 3.8 show the Virtual Mobile Edge Computing Architecture.
58
Cloud
Virtual Mobile Edge Server
RAMi CPUi GPUi STORAGEi 4 3 Device1 (PC)

Device1 Ram1 Cpu1 Gpu1 Storage1
2 3
Device4 Ram4 Cpu4 Gpu4 Storage4 4 3

Device2 (Laptop)
Device7 Ram7 Cpu7 Gpu7 Storage7 4 3

SDN Controller
3 Device3 (iPad)
4
4
Device4 (SmartPhone)
5
Device5 (Tablet)
1
Device6 (Drone)
Access Point (AP)
Device7 (Smart Vehicle)
End-User Device
(Client Device)
Fig. 3.8 Virtual Mobile Edge Computing Architecture [1].
SO-VMEC Communication Model
Figure 3.9 display the main steps of the SO-VMEC communication model that are detailed
as follows:
1. VNF Submission and Slicing: client devices send their VNF for execution to the access
point that allows resources discovery and selection of a VMES for slice allocation.
2. Resources Discovery in IoT Devices: when the access point receives the VNF from
the client device, it divides the VNF into slices. After that, it starts the selection
of IoT devices to form the VMES. The selection is based on the slices’ computing
requirements in terms of GPU, CPU, Storage, Energy, and RAM. This step taking into
account the mobility of IoT devices by using the prediction technique for both mobility
and energy consumption as follow:
(a) IoT devices mobility prediction: The IoT device mobility prediction is an es-
sential step to prevent the network from crashes during the processing such as
59
Access Point (AP)

End-User Virtual Mobile Edge Server
SDN Controller (Smart-phones, Vehicles, Cloud
(Client Device)
(SO-VMEC) Drones,...)
VNF Submission
(Slicing) Resources Discovery in Available IoT
Connected IoT Devices Devices
Slices Offloading and

Alloca�on Procedures
Lack of Resources
Slice Computa�on Results

Forward the
Results
Fig. 3.9 SO-VMEC Communication Model [1].
disconnection, which decreases the efficiency of the proposed system, especially

where the IoT devices are used as edge servers for processing. For this we used
three different mobility models depend on the IoT device type as follow: i) Low
Mobility Model: includes smart devices such as tablets, smartphones, laptops,
etc. ii) Medium Mobility Model: This type of device includes drones, city buses,
etc. iii) High Mobility Model: includes UAVs, vehicles, and all transportation
systems in general, etc.
(b) IoT devices energy prediction: Most IoT devices are based on batteries as an
energy source, for this, energy management is an important feature in the SO-
VMEC when using the IoT devices as an edge server to process VNFs that require
high processing resources include energy. The prediction technique is used to
predict the IoT device energy consumption which can prevent the edge device
from disconnection because of discharge during VNF slices processing.
3. Slices Offloading and Resources Allocation: When the VMES is selected and is ready
for processing, the access point starts the slice offloading process by allocating devices
from the selected VMES that satisfy the slices processing requirements. An optimiza-
tion algorithm is used for devices allocation according to the network requirements.
60
Moreover, the cloud computing is used when no device can process slices due to
resource miss.
4. Slice Computation Results: when the slices processing is finished in the VMES or/and
the cloud, the results are sent to the access point that merges all results to form the
main result of the submitted VNF.
5. Results Forward: in the last step, the access point sends the VNF processing result to
the source device.
IoT Devices Mobility & Energy Prediction
1. Algorithms Overview:
(a) Long Short-Term Memory (LSTM): The LSTM proposed for language modeling
[110], that came to resolve the vanishing gradient problem in Recurrent Neural
Networks. The LSTM is composed of memory blocks called gates that contain
multiplicative units. Each memory block is composed of three gates: the input
gate, the output gate and the forget gate. LSTM based neural network techniques
have been proposed to solve difficult problems in different fields, such as acoustic
modeling, speech recognition, and traffic prediction.
The LSTM model is important to understand what to forget, what to remember,
and what to output. So when a new input arrives in the neural network module,
the model first forgets any long-term information it decides it no longer needs.
Then it learns which parts of the new input are worth using, and saves them into
its long-term memory.
(b) Gated Recurrent Units (GRU): The GRU model is an improved version of the
vanilla recurrent neural network, proposed to solve the vanishing gradient prob-
lem. Basically, it uses update and reset gates to store and filter the incoming
information. It has gates responsible for modulating the information flow inside
the central blocks. A GRU block consists of a couple of gates as follows: an
update gate that is responsible for specifying the block updates content or acti-
vation, and a reset gate that makes the block read each input sequence without
considering the previous sequence (i.e., forget the previous state).
61
We used both LSTM and GRU to predict the IoT devices mobility and energy
consumption for the three mobility and energy consumption models based on real
data-sets. The prediction step is required later to feed the proposed SO-VMEC
optimization algorithms.
2. Data-sets Description:
(a) Real mobility traces: Three datasets (Low mobility, medium mobility, and high
mobility) are used for the mobility prediction.
i. Low mobility (Smartphone): The data represents the walker mobility recorded
every 0.6 seconds in the Ostermalm area inside Stockholm city, Sweden
[111].
ii. Medium mobility (City Bus): The dataset represents buses mobility in Rio
de Janeiro, Brazil [112]. The data is represented as follows: the date and
time, the bus id, the speed, the bus line, the longitude, and the latitude. The
CSV file contains the positions of more than 12000 buses.
iii. High mobility (Vehicle): The data represents taxis positions for one month
in Rome city, Italy [113]. The data is classified in a text file represented as
follows: taxi id, date and time (timestamp), and taxi position.
(b) Real IoT devices energy consumption traces: We used three datasets for different
battery types (low capacity (smartphone), medium-capacity (laptop), and high
capacity (smart TV)).
i. Low energy consumption (Smartphone): the dataset contains mobile phone
usage of 342 persons for 3 months in Spain [114]. The dataset contains
the events of 25 sensors such as location and received notifications. The
sensors such as data consumption, phone calls, cell tower, airplane mode,
Wifi details, SMS events, etc.
ii. Medium energy consumption (Laptop): The dataset contains the electrical
devices power consumption collected in Darmstadt city, Germany [115]. the
data is grouped into directories and stored in CSV files where each directory
corresponds to a device type. The data is represented as follows: date, time,
and power consumption.
62
iii. High energy consumption (Smart TV): The dataset contains real power
consumption data of electrical devices inside 22 houses in Korea [116]. The
data is grouped in directories where each directory contains the data for a
house. The data contains the timestamp and the power.
Optimization Model for SO-VMEC
In this section, we present the exact optimization model for the proposed SO-VMEC. The
model takes as inputs the results of the mobility and the energy prediction, as well as the
system capacity in terms of network, capacity, computing, and storage. It allows to optimally
place the VNF slices in the selected VMES.
Table 3.4 The notations used in the proposed SO-VMEC model.
Edge notations Definition

U The set of User Devices, It represents user demands in term
of VNF slices.
V The set of virtual edge servers.
K The set of (user) VNF slices.
Eve The maximum Energy available in the edge device ve (i.e.,
at a specific time slot according to the periodicity of the
algorithm).
Mve The predicted (required) Mobility for the edge device ve
during the VNF slice service time. It is represented as a
Boolean matrix, detecting if the edge device is available or
not during the service period.
Rve The maximum RAM available in the edge device ve.
Cve The maximum CPU available in the edge device ve.
Sve The maximum Storage available in the edge device ve.
Gve The maximum GPU available in the edge device ve.
VNF slices notations Definition
eud,k The predicted (required) Energy for the VNF slice (ud, k).
rud,k The required RAM for the VNF slice (ud, k).
cud,k The required CPU for the VNF slice (ud, k).
sud,k The required Storage for the VNF slice (ud, k).
gud,k The required GPU for the VNF slice (ud, k).
ve
xud,k A binary variable that assigns the slice k ∈ K of an end-user
device ud ∈ U to the virtual edge ve ∈ V .
63
1. Algorithm Constraints and Objective
The studied problem considers VNF slices offloading to VMES served by the network
operator. The problem objective is to determine where to offload the submitted VNF
slices in an optimal way by taking into account the resource allocation and the mobility
of the devices.
We considered different constraints for the proposed SO-VMEC optimization model

such as GPU, RAM, Storage, CPU, and IoT devices energy/mobility prediction. The
main objective is to maximize the offload of VNF slices.
2. Model Formulation
In this part, we present the defined parameters, constraints, and the proposed optimiza-
tion model. This formulation represents the optimal offloading of VNF slices to the
available devices that form the VMES.
(a) Decision Variables: We quote in Table 3.4 the used parameters and variables.
They are defined as follows:
i. The binary variable x represents the VNF slices placement, and its offloading
from the client device ud to the optimal device ve in the VMES. It is defined
as: 



1 if VNF slice (ud, k) is placed

ve
xud,k = on the virtual edge ve (3.17)



0

Otherwise
ii. The binary variable y represents the VMES usage. It is formulated as follows:

1 if the virtual edge ve ∈ V is used

yve = (3.18)
0

Otherwise
(b) Constraints: depending on the proposed communication model, the step of VNF
slices placement is constrained by the computing resources (GPU, RAM, CPU,
Storage, Energy), and the mobility of the devices. Hereafter, we present the
constraints related to the proposed SO-VMEC.
64
i. We propose that only one optimal VMES ve ∈ V should execute the offloaded
VNF slice (ud, k) ∈ (U, K). The constraint formulation is as follows:
ve
∑ xud,k ≤ 1, ∀ud ∈ U, k ∈ K (3.19)
ve∈V
ii. According to the energy prediction, the selected device should have the
necessary energy to execute the VNF slice(s), the formulation is as follow:
ve
∑ eud,k xud,k ≤ Eve , ∀ve ∈ V (3.20)
ud,k
iii. The available RAM in the server ve should be enough to execute the VNF
slice(s), the constraint formulation is as follow:
ve
∑ rud,k xud,k ≤ Rve , ∀ve ∈ V (3.21)
ud,k
iv. The available storage in the server ve should be enough, the constraint
formulation is as follow:
ve
∑ sud,k xud,k ≤ Sve , ∀ve ∈ V (3.22)
ud,k
v. The available CPU slots of the server ve should be enough to process the
VNF slice(s), the constraint formulation is as follow:
ve
∑ cud,k xud,k ≤ Cve , ∀ve ∈ V (3.23)
ud,k
vi. The GPU capacity of the server ve should be enough to process the VNF
slice(s), the constraint formulation is as follow:
ve
∑ gud,k xud,k ≤ Gve ∀ve ∈ V (3.24)
ud,k
65
vii. Guaranteeing the non-negativity of decision variables:
ve
xud,k ∈ {0, 1} (3.25)
(c) Objective Function Formulation: The general model formulation allows maxi-
mizing the VNF slices placement in the VMES. The formulation of the objective
function is as follows:
ve
max ∑ ∑ xud,k (3.26)
ud,k ve∈V
SO-VMEC Performance Evaluation
To evaluate the proposed SO-VMEC models, we used Keras 1 and TensorFlow 2 platforms
for the devices mobility and energy prediction. Moreover, we used CPLEX for python tool 3
and EdgeCloudSim [109] to assess the efficiency of the SO-VMEC in both small and dense
networks. Table 3.5 show the SO-VMEC parameters.
Table 3.5 SO-VMEC parameters.
Parameter Value
CPU slots (Min/Max) (1/30)
GPU slots (Min/Max) (1/15)
RAM (Min/Max) (Megabytes) (20/8000)
Storage (Min/Max) (Megabytes) (50/50000)
Number of Devices (Min/Max) (100/1000)
1. Key Performance Indicators (KPIs)
To evaluate the efficiency of the proposed approach, we used the following KPIs:
(a) Failed VNFs Execution: We proposed to measure the failed VNFs to evaluate the
proposed model. Because the increase of failed VNFs decreases the efficiency of
the offered service. We defined the successfully allocated VNF slices (NOS) as
follow:
ve
NOS = ∑ ∑ xud,k (3.27)
ud∈U k∈K
1 https://keras.io/
2 https://www.tensorflow.org/
3 https://pypi.org/project/cplex/
66
In the case of resource lack on VMES, the VNFs slices are offloaded to the cloud.
The formulation of offloaded slices is as follows:

ve
∑ ∑ 1 − xud,k . (3.28)
ud∈U k∈K
(b) Service Time: It allows measuring the total service duration that includes the
VNF submission, slicing, VMES selection, slices offloading, and the results
forwarding to the client device.
(c) Processing Time: represents the duration required to process the VNF slices in
the selected VMES.
(d) Failed VNFs Execution Due To Edge Capacity: resource availability is important
during processing to decrease the failed VNFs. When the GPU, RAM, CPU, and
Storage satisfy the VNFs requirements, it increases the QoS. We defined the cost
of the resources as follow:
ve
∑ud,k cud,k xud,k
CPU Cost = (3.29)
∑ve∈V Cve
ve
∑ud,k gud,k xud,k
GPU Cost = (3.30)
∑ve∈V Gve
ve
∑ud,k rud,k xud,k
RAM Cost = (3.31)
∑ve∈V Rve
ve
∑ud,k sud,k xud,k
Storage Cost = (3.32)
∑ve∈V Sve
(e) Failed VNFs Execution Due To Mobility: the IoT devices mobility impacts the
SO-VMEC system which leads to decrease the QoS. For this, the prediction
is used to estimate the future devices movement and prevent the system from
disconnection because of devices mobility.
(f) Failed VNFs Execution Due To Energy: most IoT devices use battery as an energy
source, for this we used the prediction to estimate the energy consumption to
prevent the system from disconnection that leads to failed VNF slices processing.
67
The IoT device energy cost is formulated as follows:
ve
∑ud,k eud,k xud,k
Energy Cost = (3.33)
∑ve∈V Eve
(g) Average Edge Devices Utilization: it represents the average number of used
VMES for VNF slices processing. it depends on the available resource in the
VMES, and the number of clients requests.
2. Comparative Analysis
To assess the behavior of the proposed SO-VMEC, we compared the efficiency of the
proposed approach with other networking architectures from the state-of-the-art. We
summarize these architectures as follows:
(a) Adhoc Mobile Cloud (AMC): proposed in [77, 78, 80, 81, 85, 86], this architec-
ture allows using connected devices as adhoc cloud for processing. However, the
risk of offloading failure is high in the case of unavailable devices, or when the
VNF requirements are high compared with available resources.
(b) Adhoc Mobile Edge (AME): This architecture allows using connected devices
as on-demand edge servers in adhoc mode, the submitted VNF is processed in
the selected edge servers, wherein in the case of resource unavailability, the VNF
is offloaded to the cloud. The risk of VNF failure is high in this architecture
because of device disconnection.
Table 3.6 show a summary of the main problems studied in this work compared with
existing architectures.
Architecture Mobility impact Energy impact Failure offloading management

AMC ✗ ✗ ✗
AME ✗ ✗ ✗
VME ✓ ✓ ✓
Table 3.6 The main strong points of VME against AME and AMC.
68
Hereafter, we evaluate the proposed algorithms under the above KPIs, and we present
the comparison of the proposed architecture (VME) with both AMC and AME.
3. Prediction Results
1850 43.5
Real Route
LSTM Prediction
1800
GRU Prediction
43.45
1750
1700
43.4
Coordinate Y
Coordinate Y
1650
43.35
1600
1550
43.3
Real Route
1500
LSTM Prediction
GRU Prediction
1450 43.25
700 750 800 850 900 950 22.83 22.84 22.85 22.86 22.87 22.88 22.89 22.9
Coordinate X Coordinate X
(a) Smartphone (Low Mobility) (b) City Bus (Medium Mobility)

12.5
12.495
12.49
12.485
12.48
Coordinate Y
12.475
12.47
12.465
12.46
12.455 Real Route

LSTM Prediction
12.45
GRU Prediction
12.445
41.83 41.84 41.85 41.86 41.87 41.88 41.89 41.9
Coordinate X
(c) Vehicle (High Mobility)
Fig. 3.10 Mobility Prediction for Three Mobility Models (Low, Medium, and High) [1].
Figure 3.10 shows the mobility prediction for three different mobility models (low,
medium, and high). The obtained results show that both LSTM and GRU are efficient
in terms of prediction accuracy.
Figure 3.11 show the energy prediction results of LSTM and GRU algorithms. From
the figure, we show that both algorithms converge to real energy values which help the
system to prevent crashes during the processing step.
69
Real Energy LSTM Prediction GRU Prediction Real Energy LSTM Prediction GRU Prediction
48
100 46
44
80
42
40
Energy (W)
Energy (%)
60
38
36
40
34
32
20
30
0 28
14-06 15-06 16-06 17-06 18-06 19-06 20-06 5 10 15 20 25 30 35 40 45 50
Date (day-month) Hours
(a) Smartphone (Low) (b) Laptop (Medium)

Real Energy LSTM Prediction GRU Prediction
87.6
87.5
87.4
Energy (W)
87.3
87.2
87.1
87
0 5 10 15 20 25 30 35 40 45 50
Hours
(c) Smart TV (High)
Fig. 3.11 Energy Prediction for Three Energy Models (Low, Medium, and High) [1].
Figure 3.12 shows the energy prediction impact on VNF slices execution. In figure
(3.12a), the selection of VMES without prediction leads to processing interruption
because the battery is running out of power during the VNF slices processing. Where
in figure (3.12b), the selection of VMES using the prediction prevents the system from
execution failure which improves the offloading efficiency on the VMES.
4. Optimization Results
Figure 3.13 shows the performance evaluation of the proposed VME architecure
compared with both AMC et AME architectures.
70
Energy Consumption Energy Consumption

70
80
Execution time range t=[58,73]
60
Execution time range t=[60,80]
60
Energy Consumption (%)

50
Energy Consumption (%)
40
40 Finish execution (device stay on)

30
20
Energy threshold 20
Energy threshold
10
Interrupt execution (device off)
0 0
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time (Min) Time (Min)
(a) Selected IoT Without Energy Predic- (b) Selected IoT With Energy Prediction.
tion.
Fig. 3.12 Energy Prediction Impact on VNF Slices Execution [1].
Figure (3.13a) displays the failed VNFs percentage when the number of client devices
increases. We show that when the number of client devices increases the failed VNFs
increase also for the three architectures. But, the percent of failed VNFs is very low
in the VME compared with the other architectures because both AMC and AME take
more time to select the edge devices which leads to a failure in VNF slices execution.
Figures 3.13b and 3.13c, shows the service/processing time for VNF slices execution.
From the figure we show that the VME service/processing time is low compared to the
AMC and the AME when the number of client devices is less than 550. However, when
the number of client devices exceed 550, the VME is greater than the AME because the
AME failed tasks is high compared the VME (as shown in figure 3.13a) which signify
that the VME execute tasks more than AME this is because the VME take more time.
While the service/processing time for the AMC is stay high because of the high failed
VNF slices execution (as shown in figure 3.13a) due to high VNF slices requirements
or insufficient edge devices.
Figure (3.13d) shows the failed VNFs processing due to edge capacity. From the figure,
we show that the VME outperforms the other architectures because it offloads VNFs
correctly (fewer VNFs failed due to edge capacity). Moreover, the proposed model
manages optimally the IoT devices resource.
71
35 4.5 4.5
Adhoc Mobile Cloud (AMC) Adhoc Mobile Cloud (AMC) Adhoc Mobile Cloud (AMC)
Adhoc Mobile Edge (AME) Adhoc Mobile Edge (AME) Adhoc Mobile Edge (AME)
30 Virtual Mobile Edge (VME) Virtual Mobile Edge (VME) Virtual Mobile Edge (VME)
4 4
25
Failed VNFs Execution (%)
3.5 3.5

Service Time (sec)
20
3 3
15
2.5 2.5
10
2 2
5
0 1.5 1.5
100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
Number of Mobile Devices (Client Devices) Number of Mobile Devices (Client Devices) Number of Mobile Devices (Client Devices)
(a) Failed VNFs Execution (b) Service Time (c) Processing Time
100 60 140
Adhoc Mobile Cloud (AMC) Adhoc Mobile Cloud (AMC)
90 Adhoc Mobile Edge (AME) Adhoc Mobile Edge (AME)
Failed VNFs Execution due to Edge Capacity (%)
Virtual Mobile Edge (VME) 120 Virtual Mobile Edge (VME)

50
Failed VNFs Execution due to Mobility (%)
Failed VNFs Execution due to Energy (%)

80
70 100
40
60
80
50 30
60
40
20
30 40
20
10
Adhoc Mobile Cloud (AMC) 20
0 0 0
100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
Number of Mobile Devices (Client Devices) Number of Mobile Devices (Client Devices) Number of Mobile Devices (Client Devices)
(d) Failed VNFs Execution Due (e) Failed VNFs Execution Due (f) Failed VNFs Execution Due
To Edge Capacity To Mobility To Energy
100 25
Adhoc Mobile Cloud (AMC) Virtual Mobile Edge (VME) with Prediction
90 Adhoc Mobile Edge (AME) Virtual Mobile Edge (VME) without Prediction
80 20
Average Edge Devices Utilization (%)
70
Failed VNFs Execution (%)
60 15
50
40 10
30
20 5
10
0 0
100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
Number of Mobile Devices (Client Devices) Number of Mobile Devices (Client Devices)
(g) Average Edge Devices Utiliza- (h) Failed VNFs Execution

tion With(out) Prediction
Fig. 3.13 SO-VMEC Evaluation Results [1].
Figure (3.13e) displays the impact of device mobility in terms of failed VNFs. The
proposed VME architecture outperforms both AME and AMC because of the used pre-
diction technique that prevents the system from disconnection which by consequence
decrease the failed VNFs.
Figure (3.13f) shows the impact of the energy consumption of IoT devices on the
failed VNFs processing. The obtained results show that the failed VNFs is very low in
the VME compared to both AME and AMC because of the used prediction technique
72
that helps the system to prevent crashes during the processing which by consequence
decrease the number of failed VNFs.
Figure (3.13g) displays the VMES utilization. From the figure, we show that the
VMES utilization increase because of the increase of incoming VNFs from the client
devices. Indeed, the VME architecture needs many VMES since it offloads a high
number of VNFs (less failed VNFs as presented in the figure 3.13a).
Figure (3.13h) shows the number of failed VNFs with and without prediction. From
the figure, we show that the failed VNFs is decreased when the prediction technique is
used for both energy and mobility as proved in the figures ((3.13e) and (3.13f)).
3.1.4 Use Case 4: Service Function Chains Orchestration in Virtual

Mobile Edge Computing (VMEC)
IoT-VMEC Architecture
1. IoT-VMEC System
The huge proliferation of IoT devices led to creates a complex network that generates
a high amount of data that require a short time for processing. Edge computing offers
a distributed architecture that provides efficient processing at the network edge. We
proposed to use IoT devices as virtual mobile edge server (VMES) for service function
chain (SFC) orchestration where artificial intelligence (AI) is used to help using both
cloud and edge in an intelligent way. Figure 3.14 show the proposed IoT-VMEC
architecture.
The integration of AI at the network edge aims at:
(a) Data collection and analysis: based on New Radio (NR) gNB network gateways.
(b) Mobility prediction of IoT devices: at the gNB, we used deep learning to predict
the IoT devices mobility to prevent the network from SFC VNFs execution failure
due to mobility.
(c) Energy prediction of IoT devices: based on deep learning, the gNB node predict
the IoT devices energy consumption to prevent the network from SFC VNFs
processing failure due to energy or battery crashes.
73
Cloud
VME 1
2
vCDN Chain
4
Vnf0 Vnf0 Vnf1
Slices
1
4
vCDN Operator
ain nf4
Ch V
AN
vR Vnf3 Migration
f2 2
Vn 3
MVNO Operator
in
Cha f5 Vnf5
wall
vFire Vnf3 V
n
Vnf3
Vnf3
1
gNB
Internet Service Provider
VME 2
Fig. 3.14 Proposed IoT-VMEC Architecture.
(d) Self-Organizing Network (SON) resources: based on the AI outputs (the

predicted mobility and energy), an optimization module is used to organize the
SFC VNFs offloading (MVNO, vCDN,...) to available IoT devices (VMES).
For the purpose of efficient SON resources management, the different network entities
cooperate together to ensure the processing of SFC parts (VNF slices) in an efficient
way as follows:
(a) Local processing of VNF slices at the VMES where each IoT device process
VNF slice(s) according to their available computing resources.
(b) Migration of VNF slices between VMESs in the case of lack of resources in the
selected VMES.
(c) Offloading of VNF slices to the centralized cloud in the case of a lack of
computing resources for both local processing and migration.
74
2. IoT-VMEC Communication Model
Figure 3.14 shows the main system entities for SFC orchestration in VMEC. The
communication steps (from 1 to 4) are presented as follows:
(a) Step (1): the distributed entities (vCDN, MVNO operators, etc) offload their SFC
to the gNB for placement.
(b) Step (2): happens in the gNB allow to find a set of IoT devices that satisfy the
requirements (vRAM, vGPU, vCPU, and vStorage) of the received SFC based
on the energy and mobility prediction to create a VMES.
(c) Step (3): If the selected VMES does not satisfy all the SFC parts (VNF slices)
requirements, the gNB tries to find IoT devices connected to the neighbor’s gNB
(using the migration) that can satisfy the rest of VNF slices requirements.
(d) Step (4): The last step is to send the VNF slices to the IoT devices in the selected
VMES. Moreover, the rest of VNFs slices where their requirements are not
satisfied in the selected VMES or the other VMES selected by migration, it
offloaded to the cloud.
OSPV: Optimal SFC Placement in IoT-VMEC
We proposed an optimal SFC placement algorithm in IoT-VMEC using optimization tech-

nique. Table 3.7 outlines the main parameters and decision variables.
The Equations (3.34) and (3.35) represent the decision variables. The binary variable x
indicates the placement of the VNF slices on the VMES. The binary variable y represent the
VMES usage.




1 if the VNF slices vs ∈ Vs is placed

gnb,ve
xsc,vs = on the server ve of the gnb ∈ Gnb . (3.34)



0

Otherwise.
75
Table 3.7 The parameters used in the proposed OSPV model.

Sc The set of Service Chains.
Vs The set of VNFs. Each VNF is
composed of slices.
Gnb The set of gNBs (gnb).
V The set of IoT devices available at
gnb ∈ Gnb .
Ksc The number of VNFs in each ser-
vice chain sc ∈ Sc .
Cve The maximum vCPU available in
the server ve ∈ V .
Gve The maximum vGPU available in
Rve The maximum vRAM available in
Sve The maximum vStorage available
in the server ve ∈ V .
Eve The maximum Energy available in
the edge server ve ∈ V .
SFC slices notations Definition
esc,vs The predicted (required) Energy
for the SFC slice (sc, vs).
rsc,vs The required RAM for the SFC
slice (sc, vs).
csc,vs The required CPU for the SFC
slice (sc, vs).
ssc,vs The required Storage for the SFC
slice (sc, vs).
gsc,vs The required GPU for the SFC
slice (sc, vs).
Decision Variable Definition
gnb,ve
xsc,vs A binary variable that allocates the
SFC slice vs ∈ Vs of the service
chain sc ∈ Sc to the server ve ∈ V
of the gnb ∈ Gnb .
ygnb,ve A binary variable that indicates if
the server ve ∈ V of the gnb ∈ Gnb
is used.




1 if the server ve ∈ V of the gnb ∈ Gnb

ygnb,ve = is used. (3.35)



0

Otherwise.
The objective function Optimal SFC Placement in VMEC (OSPV) represented in the
equation (3.36) allow to maximize the allocated SFC by using the minimum number of
VMES.
76
gnb,ve
max ∑ ∑ xsc,vs − ∑ ygnb,ve (3.36)
sc∈Sc ,vs∈Vs gnb∈Gnb ,ve∈V gnb∈Gnb ,ve∈V
Subject to
gnb,ve
∑ xsc,vs ≤ 1, ∀sc ∈ Sc , vs ∈ Vs (3.37)
gnb∈Gnb ,ve∈V
gnb,ve
∑ esc,vs xsc,vs ≤ Eve ygnb,ve , ∀gnb ∈ Gnb , ve ∈ V (3.38)
sc∈Sc ,vs∈Vs
gnb,ve
∑ rsc,vs xsc,vs ≤ Rve ygnb,ve , ∀gnb ∈ Gnb , ve ∈ V (3.39)
sc∈Sc ,vs∈Vs
gnb,ve
∑ ssc,vs xsc,vs ≤ Sve ygnb,ve , ∀gnb ∈ Gnb , ve ∈ V (3.40)
sc∈Sc ,vs∈Vs
gnb,ve
∑ csc,vs xsc,vs ≤ Cve ygnb,ve , ∀gnb ∈ Gnb , ve ∈ V (3.41)
sc∈Sc ,vs∈Vs
gnb,ve
∑ gsc,vs xsc,vs ≤ Gve ygnb,ve , ∀gnb ∈ Gnb , ve ∈ V (3.42)
sc∈Sc ,vs∈Vs
gnb,ve gnb,ve
∑ xsc,vs = ∑ xsc,vs+1 ,
gnb∈Gnb ,ve∈V gnb∈Gnb ,ve∈V (3.43)
∀sc ∈ Sc , vs ∈ Vs \ {Ksc }
gnb,ve (3.44)
0 ≤ xsc,vs ≤ 1, ∀gnb ∈ Gnb , ve ∈ V, sc ∈ Sc , vs ∈ Vs
0 ≤ ygnb,ve ≤ 1, ∀gnb ∈ Gnb , ve ∈ V (3.45)
The constraints of the proposed OSPV model are presented as follows:

Equation (3.37) ensures that the VNF slices vs ∈ Vs of the service chain sc ∈ Sc is placed
in only one server ve ∈ V of a gnb ∈ Gnb .
77
1. Equations (3.38), (3.39), (3.40), (3.41), and (3.42) guarantees that the selected server
has enough energy, vRAM, vStorage, vCPU, and vGPU resources to execute the VNF
slices vs ∈ Vs of the service chain sc ∈ Sc .
2. Equation (3.43) guarantee that all the VNF slices vs ∈ Vs of the service chain sc ∈ Sc
are placed.
3. Equations (3.44) and (3.45) ensures the non-negativity of the decision variables.
The proposed model is adapted to small-scale networks. Because of the high complexity
in dense networks, we will propose a heuristic placement technique to solve the SFC
placement in the IoT-VMEC architecture. The heuristic algorithm will be presented in the
next chapter.
LSTM and GRU Models Evaluation
We used LSTM and GRU to predict both the energy and the mobility of IoT devices (Section
3.1.3), The obtained results are presented as follows:
1. Energy Prediction
Figure 3.15 show the energy consumption of IoT devices in three different consumption
models (low, medium, and high). The obtained prediction results prove the efficiency
of both LSTM and GRU models that is near to real energy data which is accurate
the three consumption models 3.15a, 3.15b, and 3.15c respectively which prevent the
system from problems such as crashes during the process step.
2. Mobility Prediction
Figure 3.16 display the IoT mobility prediction. We predict the coordinates (x, y)
correspond to the positions for both pedestrian (smartphone) and city bus that represents
two different mobility models. From the figures 3.16a, and 3.16b we show that the
LSTM and GRU models are efficient in terms of prediction where the two are near to
real positions, by consequence prevent the virtual mobile edge from crashes during
processing caused by mobility.
78
Real Energy LSTM Prediction GRU Prediction Real Energy LSTM Prediction GRU Prediction
36
100 34
32
80
30
Energy (%)
60 28
Energy (W)
26
40
24
22
20
20
0 18
)
)
00
00
00
00
00
00
00
00
00
00
00
00
00
0:
0:
0:
0:
0:
0:
0:
0:
0:
0:
0:
0:
0:
16
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
0
40
60
80
00
20
40
60
80
00
(0
(1
(0
(1
(0
(1
(0
(1
(0
(1
(0
(1
(0
82
82
82
83
83
83
83
83
84
6
6
-0
-0
-0
-0
-0
-0
-0
-0
-0
-0
-0
-0
-0
50
50
50
50
50
50
50
50
50
15
15
16
16
17
17
18
18
19
19
20
20
21
24
24
24
24
24
24
24
24
24
13
13
13
13
13
13
13
13
13
Day-Month (Hour) Timestamp
(a) Smartphone (Low) (b) Laptop (Medium)

Real Energy LSTM Prediction GRU Prediction
85.7
85.6
85.5
Energy (W)
85.4
85.3
85.2
85.1
85
5 10 15 20 25 30 35 40 45
Hours
(c) Smart TV (High)
Fig. 3.15 Energy Prediction for Three Energy Models (Low, Medium, and High).
1850 12.62
1800 12.6
12.58
1750
12.56
1700
Coordinate Y
Coordinate Y
12.54
1650
12.52
1600
12.5
1550
12.48
Real Route Real Route
1500 12.46
LSTM Prediction LSTM Prediction
GRU Prediction GRU Prediction
1450 12.44
680 700 720 740 760 780 800 820 840 41.78 41.8 41.82 41.84 41.86 41.88 41.9 41.92 41.94 41.96
(a) Smartphone of Pedestrian (b) Connected Vehicle
Fig. 3.16 Mobility Prediction for Two Different Mobility Models (Low and High).
79
3.2 Vehicular Edge Computing

With the recent development of urban networks, smart vehicles are considered as mobile
devices equipped with different on-board sensors. Nowadays vehicles start to have the
capability to communicate with each others (vehicle-to-vehicle) and with urban infrastructure
(vehicle-to-infrastructure), collect data, and perform computation and in-vehicle processing.
Today, typically, high computation applications are executed on the cloud. However, delay-
sensitive applications can suffer from the long-time response and may suffer bad quality of
experience.
The new vehicular networks generations allow using vehicles as an edge server in the
context of vehicular edge computing systems (VEC) for data processing, video caching, etc.
It allows to provide information to other connected devices included end-user devices, other
vehicles, etc.

Service Offering
We proposed to use rideshare taxis as mobile edge server to offer video chunks to connected
vehicles inside the city. We designed a distributed vehicular edge computing architecture that
provide data caching and low latency communication.
MVEC: Mobile Vehicular Edge Computing Architecture
The main objective of designing any network architecture or protocol is to provide efficient
communication with high scalability support, which is considered as a fundamental chal-
lenge. Adding the huge number of users, applications, and the dynamic mobility make the
design more challenging. Thus, MVEC architecture (Figure 3.17) aims to distribute the
communication and computation in a vehicular environments. This is achieved by using
rideshare taxis as Mobile Edge Server, which covers a large city and provides ubiquitous
content distribution and computation. That by consequence enhances the communication
and data delivery in delay-sensitive applications.
80
Cloud
Video Chunks
Send video chunks

(from cloud to rideshare)
Rideshare
(Mobile Edge)
Rideshare
(Mobile Edge)
Send video chunks

(from rideshare to vehicle)
Vehicle (End user)
Fig. 3.17 MVEC architecture overview [8].
Problem Formulation & Proposed Solution
We used the prediction technique in Section 3.1.3 to estimate the future positions of vehicles
inside Rome city. In addition, the formulation of the proposed model is based on the set
cover problem (SCP) that is used to maximize the coverage (shared path distance between
the taxis and the client vehicle) inside the city. The problem formulation is as follows:
Let S = {s1 , s2 , s3 , ..., sn } a set that represents the taxis routes, E = {e1 , e2 , e3 , ..., em }
S
the representation of the client vehicle route as a set of segments e where S = E, and
H = {h1 , h2 , h3 , ..., h4 } a set of handovers correspond to taxis route. The objective is to find
X = E and |X|, h( X) are small.
S S
a subset X of S where
n
min ∑ xi hi (3.46)
i=1
81
Subject to:
∑ xi ≥ 1 ∀e j ∈ E (3.47)
{i|e j ∈si }
xi ∈ {0, 1} , ∀i ∈ {1...n} (3.48)
• Complexity:
The problem formulation is a variant of the SCP, that allows maximizing the coverage
of elements. The entry of the problem is a set of elements, a list of subsets of this set, and
an integer k, and one must find k subsets such as the number of elements belonging to at
least the subset. one of these is maximized. An element is said to be covered if it belongs to
one of the selected subsets. The SCP model is one of the Karp’s 21 NP-complete problems
[117], for this, an approximate algorithm is proposed in the next chapter to support the dense
networks.
Performance Evaluation
The performance evaluation of the proposed MVEC is evaluated in a simulated environment.

We used Cloud Report4 to simulate communications with the cloud, and SUMO5 for urban
mobility, and real tracking inside Rome city. Table 3.8 outlines the parameters used in the
simulation.
Table 3.8 MVEC parameters.
Parameter Value
Execution time 30 minutes
Number of taxis per zone 50
Number of client vehicles per zone 200
Figure 3.18 shows the cloud energy consumption (CEC) with/out edge servers for two
different vehicle routes. The results are presented for three different cases. First, CEC
for only cloud utilization (OCU), represents the use of only the cloud as a source of video
chunks. Second, CEC during cloud utilization (CU), represents the use of cloud with the
edge sometimes. Finally, the CEC during edge utilization (EU), represents the use of edge as
a source of video chunks.
4 Cloud Report: www.github.com/thiagotts/CloudReports
5 SUMO: www.sumo.dlr.de/docs/index.html
82
CEC During Cloud Utilization CEC During Edge Utilization

CEC During Edge Utilization CEC During Cloud Utilization
CEC For Only Cloud Utilization * CEC For Only Cloud Utilization *
35 35
30 * *** 30 ***
** * * *
* * ** * * *
* * *
25 * * 25 * * * *
* * * * **
*
* *
Energy (Kw)
Energy (Kw)
20 20
15 * 15
Gain
10 10
Gain
5 5
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (Minutes) Time (Minutes)
(a) Route 1 (b) Route 2
Fig. 3.18 CEC with/out Edge Servers [8].
In Figure 3.18a we show that the energy consumption for both CU and OCU is the same
in the first 14 minutes because no taxis coverage exists. Then, the energy is decreased when
the taxis (edge servers) are used which increases the energy gain.
Similarly, in Figure 3.18b we show that the taxis with high coverage leads to increase the
gain of energy. Also we show that when no taxis coverage exists, the energy consumption of
cloud increase (red circles in Figures 3.18a and 3.18b) caused by the handover during the
communication which leads to send many video requests to the cloud.
3.2.2 Use Case 2: Edge Computing Aided Autonomous Vehicles

We propose to use edge computing for autonomous driving where the autopilot VNFs are
offloaded to the edge for efficient self-driving. The proposed solution is summarized as
follows:
1. We propose a reliable, end-to-end, and low-latency communication edge autopilot

architecture that allows the offloading of autopilot VNFs services to edge servers to
improve the driving efficiency for autonomous vehicles.
2. We propose a communication protocol for dense moving vehicles.
3. The optimization technique is used to model the edge-assisted autonomous driving for
optimal autopilot VNFs allocation on edge servers.
83
Proposed Edge Autopilot Architecture
The proposed architecture for edge autopilot consists of three layers as follows:
1. Centralized Cloud Layer: represents the cloud autopilot, is responsible for processing
a part of the autopilot VNFs.
2. Distributed Edges Layer: consists of edge servers responsible for autopilot VNFs
analyzing and processing.
3. Autonomous Vehicles Layer: represents the autonomous vehicles that request autopi-
lot VNFs services offloading because of the scarcity of local resources.
Proposed Edge Autopilot Protocol
1. Autopilot VNFs: the autonomous vehicles can offload some autopilot VNFs for
processing. It requests the edge server in the near gNB or RSU to enable edge resource
discovery and allocation.
2. Resources Discovery in Edge Servers: when the gNB receives the autopilot VNFs,
it selects the edge servers that satisfy the VNFs requirement in terms of computing
resources and network capabilities.
3. Autopilot VNFs Offloading: when the edge servers are selected, the gNB starts the
allocation process by placing each VNF in the optimal edge server. Still, the cloud
may represent the alternative solution in the case of a lack of resources at the network
edge to ensure the autopilot VNFs processing.
4. Vehicle-Edge/Cloud Connection: in the last step, the gNB forwards the control
commands to the autonomous vehicle by creating a communication link between the
edge/cloud and the autonomous vehicle while satisfying its requirements.
Problem Formulation for Edge Autopilot
We proposed an optimal model for autopilot VNFs placement (OVEAP) based on the
optimization technique. It takes as input the computing resources for each edge server. Then,
84
it allows to allocate optimally the autopilot VNFs upon the edge servers. The autopilot VNFs
are offloaded to the edge server to reduce the processing time and enhance driving safety.
The OVEAP model is formulated as follow:
1. OVEAP Parameters and Decision Variables
Table 3.9 outlines the parameters and decision variables of the proposed OVEAP
model.
Table 3.9 OVEAP Parameters and Decision Variables
E The set of Edge Autopilots in
terms of services chain.
V The set of Edge Autopilot VNFs.
Each Edge Autopilot VNF is com-
posed of slices.
A The set of Autonomous Vehicles.
M The set of MEC servers at the net-
work edge.
Lea The number of VNFs in each Edge
Autopilot ea ∈ E. It represents the
length of the Edge Autopilot.
Gmec The maximum computing capacity
vGPU available in the MEC server
mec ∈ M.
gea,vn f Required vGPU resources for the
VNF slices vn f of the Edge Au-
topilot services chain ea.
amec
ea,vn f A binary variable that allocates the
Edge Autopilot VNF vn f ∈ V of
the Edge Autopilot ea ∈ E to the
MEC server mec ∈ M.
bmec A binary variable that indicates if
the MEC server mec ∈ M is used to
process the edge autopilot VNFs.
The binary variable a indicates the allocation of the autopilot VNF (ea, vn f ) on the
MEC server mec.




1 if the VNF v f n of the autopilot service chain ea

amec
ea,vn f = is allocated on the MEC server mec (3.49)



0 Otherwise

Further, the binary variable y is used to track the MEC server usage. The variable
formulation is as follows:
85

1 if the MEC server mec ∈ M is used

bmec = (3.50)
0

Otherwise
2. OVEAP Model Formulation
The OVEAP objective function (3.51) allows to maximize the autopilot VNFs place-
ment in the edge servers. The general model formulation is as follows:
max ∑ ∑ amec
ea,Lea − ∑ bmec (3.51)
ea∈E mec∈M mec∈M
Subject to
∑ amec
ea,vn f ≤ 1, ∀ea ∈ E, vn f ∈ V (3.52)
mec∈M
∑ ∑ gea,vn f × amec
ea,vn f ≤ G
mec mec
b ,
ea∈E vn f ∈V (3.53)
∀mec ∈ M
∑ amec
ea,vn f = ∑ amec
ea,vn f +1 , ∀ea ∈ E,
mec∈M mec∈M (3.54)
vn f ∈ V \ {Lea }
bmec ≤ ∑ amec
ea,vn f , ∀mec ∈ M (3.55)
ea,vn f
0 ≤ amec
ea,vn f ≤ 1, ∀mec ∈ M, ea ∈ E, vn f ∈ V (3.56)
0 ≤ bmec ≤ 1, ∀mec ∈ M (3.57)
86
The model constraints are detailed as follows:
Constraint (3.52) guarantee that the VNF slices vn f ∈ V of the autopilot chain ea ∈
E is placed on a unique MEC server mec ∈ M. Constraint (3.53) ensures that the
vGPU resources in the selected MEC server is enough to process the VNF slices of
the autopilot chain. Constraint (3.54) ensure the respect of VNFs chaining during
the process in the edge servers, where the allocation of one VNF should be done
completely in the edge server to allocate the next VNF in the chain. Constraint (3.55)
are formulated to select the minimum number of edge servers for the VNFs allocation.
Finally, constraints (3.56) and (3.57) guarantee the non-negativity of the decision
variables.
3. OVEAP complexity
OVEAP is NP-hard which is feasible only with a few instances within a reasonable
time in small-scale networks. For this, a suitable algorithm for dense networks is
presented in the next chapter.
Edge Autopilot Performance Evaluation
We considered different autopilot VNFs: vPerception, vLocalisation, vPlanner, and vControl.

The service chain is composed of a set of autopilot VNFs. The objective is to place autopilot
services in the edge server while guaranteeing the VNFs chaining. Tables 3.10 and 3.11 show
the different configurations used in the evaluation.
Table 3.10 Autonomous Vehicles Configuration.
Autonomous Vehicles Autopilot Chain

Vehicle 1 vPerception
Vehicle 2 vPerception, vLocalisation
Vehicle 3 vPerception, vLocalisation, vPlanner
Vehicle 4 vPerception, vLocalisation, vControl
Vehicle 5 vPerception, vLocalisation, vPlanner, vControl
Embedded Autopilots Number of servers
Embedded vehicle 1
1. Key Performance Indicators (KPIs) To evaluate the proposed model, we proposed the
following KPIs:
87
Table 3.11 Summary of autopilot edge servers (AES) parameters.
Parameter Min value Max value

Number of AES 3 5
Number of Autonomous vehicles 5 (small scale) 100
Available vCPU in AES 50 (slot) 200 (slot)
Available vGPU in AES 10 (slot) 150 (slot)
Available vRAM in AES 50 (Gigabytes) 200 (Gigabytes)
Available vStorage in AES 10 (Terabytes) 100 (Terabytes)
Available Bandwidth in AES 10 (Mbps) 150 (Mbps)
Latency in Autopilot Chain 1/8 (ms) 1 (ms)
(a) Average Network Delay: it represents the network delay between the edge
servers and the autonomous vehicles.
(b) Service Time: represents the end-to-end time to accomplish the service from the
autopilot VNF submission to the results forwarding to the autonomous vehicle.
It is depend on the offloading decisions and availability of resources where an
efficient service has low execution time.
(c) Processing Time: represents the duration needed to complete the execution of
the autopilot VNFs. In general, the processing time is related to the availability
of resources in edge servers.
To assess the behavior of the proposed OVEAP at the network edge, we compared
three different computing architectures. We summarize the architecture as follows:
(a) Embedded Computing: allows the processing of autopilot VNFs in the au-
tonomous vehicle (local processing) while the edge is still active to receive
VNFs.
(b) Edge Computing: this architecture prioritizes the edge servers for autopilot
VNFs processing.
(c) Cloud Computing: this architecture allows the use of the centralized cloud for
autopilot VNFs processing.
3. Obtained Results
88
Figure 3.19 outlines the performance evaluation of the proposed approach.
In Figure 3.19a, We plot the network average delay between autonomous vehicles and
the computing servers. The result shows that the delay is low for embedded computing
because the processing happens locally, while the delay of the edge is lower than
the cloud since the edge offers a short response time because is near to autonomous
vehicles. In Figure 3.19b we show that edge reduces the service time compared to
cloud architecture. Indeed, the heavy autopilot VNFs are offloaded to the near edge
for efficient processing. In Figures 3.19c and 3.19d we show that the edge computing
reduces the service/processing time compared to both cloud and embedded computing
since the edge offer efficient processing service near to autonomous vehicles.
3.2.3 Use Case 3: Edge Computing Assisted Autonomous UAV

We propose to use autonomous mobile UAVs as mobile edge servers to offer video chunks to
connected vehicles in vehicular edge computing architecture. Each connected vehicle inside
the city can receive video chunks from a UAV(s) that cover the vehicle route.
Proposed UAV-assisted V2X Architecture
Figure 3.20 display the proposed Follow Me UAV (FMU) architecture with the communication
entities described as follow:
1. Connected Vehicles: represent the end-user layer consists of a high number of vehicles.
2. Autonomous Mobile UAVs: is an intermediate layer that represents the MEC servers
responsible for reducing the communication load on the centralized cloud. It guarantees
ultra low latency communication (ULLC) and safety for sensitive delay applications.
In ULLC, UAVs serve the high dynamic connected vehicles according to their quality
requests.
3. The Centralized Cloud: represents a set of servers that provide high computing and
storing services for insensitive delay applications.
89
0.08 0.6
0.07 0.5
Average Network Delay (sec)
Service Time on Cloud (sec)

0.06 0.4
Embedded Computing Embedded Computing

0.05 Edge Computing 0.3 Edge Computing
Cloud Computing Cloud Computing
0.04 0.2
0.03 0.1
0.02 0
20 30 40 50 60 70 80 90 100 20 30 40 50 60 70 80 90 100
Number of Autonomous Vehicles Number of Autonomous Vehicles
(a) Average Network Delay (b) Service Time on Cloud

Embedded Computing Embedded Computing
Edge Computing Edge Computing
Cloud Computing Cloud Computing
3 2.8
2.8 2.6
2.6
2.4
Service Time (sec)
2.4
2.2
2.2
2
2
1.8
1.8
1.6 1.6
1.4 1.4
20 30 40 50 60 70 80 90 100 20 30 40 50 60 70 80 90 100
Number of Autonomous Vehicles Number of Autonomous Vehicles
(c) Service Time (d) Processing Time
Fig. 3.19 OVEAP Performance Evaluation [5].
FMU Optimization Algorithm
1. The Optimal FMU (OFMU) Algorithm The FMU objective is to find the shared path
between UAVs and connected vehicles to maximize the vehicles coverage using avail-
able UAVs. For this, we propose an optimal model based on optimization technique.
Table 3.12 displays the used notations in the FMU model.
Let S = {s1 , s2 , s3 , ..., sn } represent UAVs routes as a collection of sets.
Let W = {w1 , w2 , w3 , ..., wn } a collection of sets represent the energy consumed in

each UAV routes.
90
Cloud
Video Chunks
UAV for
Surveillance
UAV for
Journalism
UAV for
Delivery
Co
nn
ec
te
d
Ve
hicl
e
UAV Communication Range
Fig. 3.20 Proposed FMU Architecture [3].
Table 3.12 The used notations in the FMU model.
C Cloud service.
CV The set of Connected Vehicles cv ∈ CV = {1..|CV |}
U Set of connected UAVs in service.
Dv,u Vehicle v ∈ V consumes the content from u ∈ U
Cv,c Vehicle v ∈ V is connected with c ∈ C
huav The handover delay required by UAV uav ∈ UAV
wuav The energy consumption required by UAV uav ∈ UAV
uav
Mmax Maximum association links of the UAV uav ∈ UAV
θ It represents the minimum threshold of the covered
connected vehicles cv ∈ CV
E = {e1 , e2 , e3 , ..., em } is a set of adjacent segments ei , i ∈ {1..m} represent the vehicle

S
route where S = E.
H = {h1 , h2 , h3 , ..., hm } a set of handovers of UAVs route.
91
X = E and |X|, h( X) are

S S
The main objective is to find a subset X of S such that
small as possible while minimizing the UAVs energy consumption and the number of
handover.
To clarify better the proposed OFMU algorithm, we show in Figure 3.21 an example
of UAV selection to cover the vehicle route while minimizing the number of handovers
and the UAVs energy cost. We have as an input the predicted vehicle’s route using
deep learning (detailed in the Section 3.1.3), and the UAVs routes with predefined
energy cost for each UAV. The OFMU algorithm selects the most cost-effective UAVs
according to the algorithm constraints.
Connected Vehicle Route (E)
e1 e2 e3 e4
Connected
Vehicle
s1
Handover (h)
UAV for
Journalism
s2
UAV for The Parts of UAVs Routes
Security Shared With The Vehicle
s3
UAV for
Delivery
Fig. 3.21 Abstraction Model For FMU [3].
2. Decision variables
The binary variable x indicates the UAV usage. It is formulated as:


1

if the UAV uav is active
xuav = (3.58)
0

Otherwise
92
The binary variable y represents the association link between the UAV and the vehicle.
It is defined as follow:

1

if CV cv is associated with UAV uav
yuav
cv = (3.59)
0

Otherwise
The binary variable z represents the communication links among UAVs. It is defined
as follows: 



1 if UAV uav1 is within UAV uav2 ’s

zuav1 ,uav2 = communication range (3.60)



0

Otherwise
3. The FMU exact formulation and constraints
The objective function of the problem is formulated as follows:
min ∑ xuav (αhuav + (1 − α)wuav ) (3.61)

uav∈U
Where α is a metric related to the operator strategy.
Moreover, the OFMU constraints are defined as follows:
Maximum Coverage: constraint (3.62) guarantees the maximum coverage of vehicles

using UAVs:
∑ xuav ≥ 1 ∀e j ∈ E (3.62)
{uav|e j ∈suav }
Minimum Energy: guarantees that each UAV must respect the maximum energy
threshold authorized for mission achievement:
∑ xuav × wuav ≤ |N| ×W (3.63)

uav∈U
The set W represents the maximum energy consumed for each UAV.
93
Minimum Handover Delay: This constraint guarantees that the UAV route handover
must be lower as possible:
∑ xuav × huav ≤ H (3.64)

uav∈U
The set of handover H, is the number of handover for each UAV route.
Dynamic Association: this constraint ensures that the connection of connected vehi-
cles is not permitted to deleted UAVs:
∀uav, cv : yuav
cv ≤ x
uav
(3.65)
Single UAV-CV Association: ensures that a connected vehicle can be served by at

most one UAV:
∀cv ∈ CV : ∑ yuav
cv ≤ 1 (3.66)
uav∈UAV
UAV Association Capacity: represents the upper bound number of connected vehicles
that can be covered by one UAV:
∀uav : ∑ yuav uav

cv ≤ Mmax (3.67)
cv∈CV
QoE Constraints: this constraint ensures that the percentage of non successfully
served vehicles should not exceed a predefined threshold θ . The constraint formulation
is as follow:
∑ ∑ yuav
cv ≥ (1 − θ ) × |CV | (3.68)
uav∈UAV cv∈CV
Bi-connection: this constraint ensures the bi-connection between the UAVs:
∑ zuav1 ,uav2 ≥ 2 (3.69)

uav1,uav2∈UAV
94
Non negativity: this constraint ensure the non-negativity of the used variable for the
UAV usage:
xuav ∈ {0, 1} , ∀uav ∈ {1...n} (3.70)
FMU Performance Evaluation
To assess the efficiency of the proposed approach we proposed different communication

scenarios as follows:
1. Low Quality Cloud (LQC): The connected vehicles receive low-quality video chunks
for live streaming from the cloud only.
2. Low Quality Edge (LQE): The UAVs (edge servers) are introduced as intermediate
nodes to offer a low-quality video for live streaming to connected vehicles.
3. Medium Quality Cloud (MQC): The cloud offers medium-quality video for live
streaming to connected vehicles.
4. Medium Quality Edge (MQE): The intermediate UAVs offer medium-quality video
for live streaming to connected vehicles.
5. High Quality Cloud (HQC): The cloud offers high-quality video for live streaming
to connected vehicles.
6. High Quality Edge (HQE): The UAVs offer high-quality video for live streaming to
connected vehicles.
For the above cloud scenarios (LQC, MQC, and HQC), the vehicles are connected directly
to the cloud without the use of edge servers (UAVs).
Table 3.13 outlines the FMU parameters.
Figure 3.22a shows the lost video chunks due to mobility. From the figure, we show
that the use of edge servers (UAVs) decrease the lost video chunks compared to the use of
cloud only. This lead to increase both the user satisfaction index and the QoS. Moreover,
from the obtained results, the use of suitable computing is recommended (cloud and/or edge)
regarding the streaming quality and the Service Level Agreement between network operators
and Over the Top (OTT) providers.
95
Table 3.13 FMU parameters.
Parameter Value
Max Data Download 4000 Megabytes
Number of UAVs per zone (Min/Max) (2/10)
Number of Vehicle (Min/Max) (100/1000)
Max Low Quality Video Size 200 Megabytes
Max Medium Quality Video Size 400 Megabytes
Max High Quality Video Size 600 Megabytes
Figure 3.22b, we measure the average network delay for the proposed scenarios against
the number of vehicles. The result show that the UAV edge layer reduces the network delay
which enhances the quality of service. We remark that all the edge based scenarios (LQE,
MQE, and HQE) gives less network delay comparing to the cloud based scenarios (LQC,
MQC, and HQC). Therefore, the results suggest always the use of UAV as mobile edges.
Figure 3.22b shows the average network delay for different video quality. The results
show that the use of UAV as edge servers reduce the delay compared to the use of cloud only
because the content (video chunks) is near to connected vehicles which leads to decrease the
delay and enhances the QoS.
250 0.35
LQC LQC
Unreceived Video Chunks Due To Mobility (Nb chunks)
LQE LQE
0.3
MQC MQC
200
MQE MQE
Average Network Delay (sec)
HQC 0.25 HQC

HQE HQE
150
0.2
0.15
100
0.1
50
0.05
0 0
100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
The Number of Vehicles The Number of Vehicles
(a) Unreceived Video Chunks. (b) Average Network Delay.
Fig. 3.22 System Evaluation for Different Communication Scenarios [3].
96
Conclusion
We proposed different optimal models based on optimization techniques to enhance the
QoS and improve user satisfaction according to the offered services. We studied two
different communication technology-based edge computing called service offloading and
vehicular edge computing, where for each technology we proposed optimal models each one
is dedicated to a problem. The common factor between all the proposed models is the high
complexity that restricts the use of these models. This makes its usability in only small-scale
networks with a limited number of devices. For this, approximate algorithms are proposed in
the next chapter to support the huge amount of requests and connected users in large-scale
networks.
97
Chapter 4
Large Scale Solutions For Edge

Computing Networks
Large-scale algorithms are introduced as the alternative solution for the optimal solutions in
dense networks due to their feasible execution time and the given efficient solutions. Different
large-scale algorithms are proposed by the researchers to resolve several networks problems
such as services placement at the edge servers taking into account the network, storage, and
computing resources constraints. Most of these algorithms proved their efficiency in dense
networks with a huge amount of produced data and users demands.
4.1.1 Use Case 1: 5G VNF Slices Placement in EC

We proposed to use Reinforcement Learning (RL) and Deep Reinforcement Learning (DRL)
techniques for large scale VNF slices placement in edge computing.
The RL algorithm
In this part, we present our proposed scheme for online VNF slices placement with reinforce-
ment learning. We formulate the problem (3.3) and represent it as a RL model.
Q-learning is the main algorithm used for the RL model. The Q-function allows estimating
the future VNF placement reward. We define hereafter the proposed model.
99
Large Scale Solutions For Edge Computing Networks
Reward
CPU 1 CPU 2 CPU 3 CPU 4 CPU 5

Server 1
Server 2
Server 3
Agent
Acon (Placement)
Server 4
Server 5
Server 6
Server 7
State
Fig. 4.1 RL application to VNF placement in Edge Computing [7].
1. The proposed model
We consider a cluster with the CPU as a computing resource. The incoming VNF
slices to the cluster are jobs that arrive in discrete time steps. The cluster manager
places the VNF in the server according to the VNF size and the available CPU slots in
the server. Figure 4.1 show the RL application to VNF placement in edge computing.
2. The proposed formulation
We represent the system state as the current VNF slices placement in the server slot.
Figure 4.1 (right side) display the proposed state-space (edge server) that is considered
as the environment.
The action represents the placement function that takes the decision on the efficient
placement of the VNF slice by taking into consideration the server capacity in terms of
CPU slots. Moreover, the agent can not place a VNF slice in an occupied server slot.
The action allows the agent to process one by one the incoming VNFs until process all
the arrived requests.
The proposed reward represents the placement cost of VNFs. It is calculated based on
the used server after performing an action. It is defined as follows:

100 × i

if i servers are opened
Rt = (4.1)
0

Otherwise.
100
As shown in the equation 4.1, the main objective is to minimize the total used servers.
In other words, the objective of the agent is to minimize the total discounted cumulative
rewards.
In Algorithm. 1 we present the proposed RL algorithm.
Algorithm 1 : RL-driven VNF slices placement in Edge Computing Over 5G Network

[7].
1: Input: Servers, the number of slots per Server (CPUs or GPUs), the number of VNFs, and the number of
slices per VNF
2: Output: Q∗ (state(s), action(a))
3: Initialise Q(state(s),action(a)) arbitrary.
4: Observe an initial state s (current allocation and first incoming VNF slice)
5: repeat
6: Select and place a VNF slice on the edge computing server
7: r, s′ ← Observe the placement cost r and the new state s′ (new allocation and another incoming VNF
slice)
8:
Q(s, a) = Q(s, a) + α(r + γ × max Q(s′ , a′ ) − Q(s, a)) (4.2)
a′
9: s = s’
10: until no incoming VNF slices
The DRL algorithm
In the DRL algorithm, the deep neural network is used to approximate the RL algorithm. A
succession of neural network layers are used to map the input to the output (state to action).
The Stochastic Gradient Descent (SGD) algorithm [118] is used in the deep neural network to
train the deep q-learning network (DQN) [119]. The objective of the DRL model is to reduce
the complexity of both the exact model (3.3) and the RL model by reducing the iterations to
be considered in the optimization. Algorithm 2 presents the proposed DRL algorithm.
To assess the performance of the proposed algorithms, we evaluate three key performance
indicators: placement time, server utilization, and energy consumption.
For each environment, we present a set of parameters. First, servers configuration as
mentioned in table 4.1. Then, we specify the VNF configuration as depicted in table 4.2.
Figure 4.2 shows the performance evaluation of the proposed algorithms: the exact ILP
model (3.3), Q-learning, and DQN in dense networks.
101
Algorithm 2 : DRL-driven VNF slices placement in Edge Computing Over 5G Network

[7].
1: Input: Servers, the number of slots per Server (CPUs or GPUs), the number of VNFs, and the number of
slices per VNF
2: Output: Q∗ (state(s), action(a))
3: Initialize a reply memory D
4: Initialise action-value Q with random weights
5: Observe initial state s
6: repeat
7: Select an edge computing server a .
• with probability ε select a random edge computing server
• Otherwise select the server that has the maxa′ Q(s, a′ )
8: Place the VNF slice on the selected edge computing server a.
9: r, s′ ← Observe the placement cost r and the new state s′ (new allocation and another incoming VNF
slice)
10: store the experience {s, a, r, s′ } in the replay memory.
11: sample a random transition from the replay memory.
12: Calculate the target for each mini-batch transition (r + γ × maxa Q(s′ , a′ ))
13: Train the Q network using the following loss
1
Loss = ∗ (r + γ × max Q(s′ , a′ ) − Q(s, a))2 (4.3)
2 a′
14: s = s’
15: until no incoming VNF slices
Table 4.1 Servers Configuration.
Environments The number of servers The number of cpu in

servers (slots/servers)
Environment 1 (Small) 5 5/5
Environment 2 (Medium) 10 3/3, 4/3, 5/2, 6/2
Environment 3 (Large) 15 3/2, 4/3, 5/4, 6/3, 7/3
Table 4.2 VNF Configuration.
VNF Configuration The number of vnf The number of slices in vnf

(slices/vnf)
Configuration 1 5 2/3, 3/2
Configuration 2 10 1/3, 2/3, 3/4
Configuration 3 15 1/8, 2/5, 3/1, 4/1
Figure 4.2a shows the placement time for the proposed algorithms. From the figure, we
show that when the number of VNF increases, the placement time increase for the three
models while the time for the ILP is very high compared to the other algorithms that place
VNF efficiently in terms of time.
102
100 70
ILP Q-Learning
90 Q-Learning DQN
60
DQN
80
70 50
Cost (Server Utilization)

Time (Time Unit)
60
40
50
30
40
30 20
20
10
10
0 0
200 400 600 800 1000 1200 1400 1600 1800 2000 200 400 600 800 1000 1200 1400 1600 1800 2000
The Number of Slices The Number of Slices
(a) Placement Time (b) Server Utilization

100
ILP
90 Q-Learning
DQN
80
Energy Consumption in The Edge (%)
70
60
50
40
30
20
10
0
200 400 600 800 1000 1200 1400 1600 1800 2000
The Number of Slices
(c) Energy Consumption
Fig. 4.2 Performance Evaluation of VNF Slices Placement in the edge using ILP, Q-Learning,
DQN in Large Scale Networks [7].
Figure 4.2b display the server utilization for both Q-learning and DQN in dense networks.
From the figure, we show that the DQN algorithm is more efficient compared to the Q-
learning algorithm.
Figure 4.2c displays the energy consumption in the edge when using the proposed
algorithms. We show that the ILP gives less energy consumption compared to both Q-
learning and DQN when the number of VNF is less than 400. While the energy is very
high for the ILP compared to the other algorithms when the number of VNF is higher than
103
400 because the ILP uses high computation resources. Moreover, the DQN is an efficient
algorithm with lower energy consumption.

(SO-VMEC)
Deep Reinforcement Learning Model for SO-VMEC
The exact optimization algorithm is a high-cost model in terms of computation resources,

which is not adapted for dense networks. For this, we use the deep reinforcement learning
(DRL) to solve the exact model (3.26) that places the VNF slices in virtual edge servers.
1. The Proposed DRL Algorithm
We used the equation Q(s, a) = r(s) + γ × maxa′ Q(s′ , a′ ) to update the state-action
pairs. Then, the policy equation π(s) = maxa Q(s, a) is used to select the action. By
using deep learning, the Q-values are updated based on the loss value in minθ ||
r + maxa′ Q(s′ , a′ ) − Q(s, a; θ ||2 + || λ × θ ||2 where r represents the reward, and
θ represents the shared weights in the neural network. Deep Q-learning (DQN)
represents a DRL instance. It is the brain of the proposed SO-VMEC where the agent
is responsible for the placement process.
Figure 4.3 display the DRL process, The main objective is to place the VNF slices in
the VME.
Reward
CPU
GPU
Device 1 RAM
Storage
State Device 2
CPU
GPU Environment
RAM
Storage
(Selected VME)
Acon (Ooading) .
.
.
CPU
GPU
Device n RAM
Storage
Neural Network
Observaon
Fig. 4.3 Deep Reinforcement Learning for VNF Slices Offloading [1].
In the algorithm 3, we present the detail of the proposed DRL model.
104
Algorithm 3 : Deep SO-VMEC Algorithm for Service Offloading Optimization [1].

Input: The big network resources image (i.e., environment settings)
Output: Q∗ (state(s), action(a))
Do a feed-forward pass for the current states to get predicted Q-values for all actions.
Do a feed-forward pass for the next states and calculate maximum overall network outputs maxa′ Q(s′ , a′ ).
Set Q-value target for action to r + γ × maxa′ Q(s′ , a′ ) (use the max calculated in step 2). For all other actions,
set the Q-value target to the same as originally returned from step 1, making the error 0 for those outputs.
Update the weights using back-propagation. The loss function that should be minimized is || (r +
maxa′ Q(s′ , a′ ) − Q(s, a)2 ||2
To simplify the behavior of the DQN algorithm during the execution, the agent must
be able to match VNF slices to the server slots by minimizing the occupied servers.
To assess the efficiency of the proposed approach, we compare the DQN algorithm with the
exact ILP model (3.26) using the parameters mentioned in table 3.5. We measured both the
number of selected VME and the time.
12 100
ILP ILP
DQN 90 DQN
10
80
Number of Devices Selected as VME
70
8
Time (Time Unit)
60
6 50
40
4
30
20
2
10
0 0
5 10 15 20 25 30 100 200 300 400 500 600 700 800 900 1000
Number of VNFs Slices The Number of VNF Slices
(a) The Number of Devices Selected in VME. (b) VME Selection Time.
Fig. 4.4 VME Based on ILP and DQN Models For VNFs Slices Offloading [1].
Figure 4.4 display the performance evaluation of the proposed SO-VMEC based on both
DQN and ILP algorithms.
105
Figure (4.4a) displays the number of IoT devices in VME for different number of requests
(VNF slices). The results prove the efficiency of the proposed DQN algorithm since it gave a
solution equal/converge to the optimal ILP model.
Further, Figure (4.4b) shows the required time to select the VME for different VNF slices
requests. The obtained results show that the DQN algorithm is efficient in terms of time in
dense networks. In addition, the ILP is feasible since it gave a good selection time.
4.1.3 Use Case 3: Service Function Chains Orchestration in Virtual

Mobile Edge Computing (VMEC)
The exact OSPV algorithm (3.36) is a high-cost model in terms of computation resources,
which is no suitable for dense network. For this, we propose a heuristic algorithm for SFC
placement in VMEC-IoT architecture.
ESPV: Efficient SFC Placement in VMEC over AI-IoT
The optimization targets a high number of SFC instances, where the objective relies on
the Bin Packing problem [120]. The proposed algorithm allows satisfying the large SFC
instances to deal with dense networks. it is noteworthy that the optimal OSPV algorithm
could not be used for dense networks due to its complexity.
Algorithm 4 show the ESPV details.
Algorithm 4 : Efficient SFC Placement in VMEC (ESPV)

1: Input: Sc , Vs , Gnb , V , Cloud
2: Output: Z
3: repeat
4: repeat
5: Predict mobility sequences (x,y) of ve ∈ V using deep learning (Section 3.1.3)
6: Predict energy values of ve ∈ V using deep learning (Section 3.1.3)
7: until All servers ve in V are verified.
8: repeat
9: Check future positions from the prediction of ve.
10: Check resource availability (vCPU, vRAM, vGPU, vStorage, and Energy) in the server ve.
11: until Find a server ve to place the vs (True or False).
12:
13: if (resource available in ve) then
14: Place the VNF slices vs on the server ve.
15:
16: else
17: Offload the VNF slices vs to the Cloud.
18: end if
19: until no incoming SFC (VNFs).
106
To measure the efficiency and feasibility of the proposed algorithm, we propose different
system parameters to cover real-world scenarios as shown in the table 4.3.
Table 4.3 Summary of the parameters.
Parameter Min value Max value

Number of gnb 3 5
Number of IoT 3 10
Available vCPU in IoT 5 (slot) 20 (slot)
Available vGPU in IoT 1 (slot) 15 (slot)
Available vRAM in IoT 20 (Megabytes) 2000 (Megabytes)
Available vStorage in IoT 50 (Megabytes) 20000 (Megabytes)
Available Energy in IoT 25 (Energy unit) 100 (Energy unit)
1. Key Performance Indicators (KPIs)
To study the performance of the proposed algorithm, we propose to measure the

following KPIs:
(a) The Total Allocated Servers: represent the number of used servers (IoT devices)
in the VME.
(b) The Placement Time : represent the required time for SFC VNFs placement.
To assess the performance of the proposed approach, we compare the proposed ESPV
algorithm with our OSPV exact model and the exact model (other) in [121].
3. Obtained Results
Figure 4.5 displays the total allocated servers and the required time for the EPSV and
OPSV compared with other algorithm.
Figure 4.5a shows the total servers utilization for SFC VNFs placement, we show that
both the exact algorithms (OPSV and Other) give equal results for servers allocation
cost where the EPSV algorithm allocates more servers but is near the optimal model.
107
12 0.75
OSPV 0.7 OSPV
Total Allocated Servers (IoT Devices)
ESPV ESPV
0.65
10 Other Other
0.6
Placement Time (Time Unit)

0.55
8 0.5
0.45
0.4
6
0.35
0.3
4 0.25
0.2
0.15
2
0.1
0.05
0
2 4 6 8 10 12 14 16 18 2 4 6 8 10 12 14 16 18
Number of Incoming VNFs Number of Incoming VNFs
(a) VNFs Placement Cost (b) VNFs Placement Time
Fig. 4.5 Total Resources Utilization for Different Flow Rates of SFC VNFs.
Figure 4.5b displays the SFC VNFs placement time, we show that the OPSV gives
a lower time compared with the Other algorithm which proves the efficiency our
optimal OPSV model. Moreover, the ESPV algorithm gives a very low placement
time compared to the exact models, since it provides an efficient placement cost with a
lower time.

Service Offering
Rideshare Taxi Selection and Service Offering Algorithm
The formulated SCP problem (model 3.46) is NP-complete. For this, Algorithm 5 is proposed
to deal with dense networks. it uses only a subset of the main set of taxi paths which by
consequence reduces the execution time. This subset contains only the taxis that share part(s)
of the route with the client vehicle, which minimize the data processing complexity.
108
Algorithm 5 : Maximum Coverage For Rideshare Taxi Selection [8].

1: Input: RideshareTaxisRoutesList, VehiclesRoutesList
2: Output: RideshareTaxisList
3: Find rideshare taxis that cover each vehicle route.
4: for (i in VehiclesRoutesList) do
5: for ( j in RideshareTaxisRoutesList) do
6: if (Intersect(i, j) ! = Null) then
7: append j in RideshareTaxisList(i)
8: end if
9: end for
10: end for
11: return RideshareTaxisList.
We defined a set of parameters for each proposed scenario such as the number of taxis and
client vehicles where both of them use a predefined route to reach the destination. The route
of each vehicle is represented as a connected road segments in Rome city.
12.54 Without taxis coverage 12.62 Without taxis coverage

With taxis coverage With taxis coverage
12.60
12.52
12.58
12.50
Coordinate Y
Coordinate Y
12.56
12.48 12.54
12.52
12.46
12.50
12.44
12.48
41.600 41.625 41.650 41.675 41.700 41.725 41.750 41.775 41.800 41.9 42.0 42.1 42.2 42.3 42.4
(a) Taxis Coverage For Route (1) (b) Taxis Coverage For Route (2)
Fig. 4.6 Real Routes With Taxis Coverage [8].
Figure 4.6 shows the real routes with taxis coverage in two different cases. In the figure
4.6a we show that there are only some few coverage of taxis in short distance, where in
Figure 4.6b we show that the coverage of taxis is large in long distance.
4.2.2 Use Case 2: Edge Computing Aided Autonomous Vehicles

The AI-defined optimization tries to be an alternative solution for exact algorithms by the
integration of AI models in the edge to facilitate the analysis and the process of data especially
109
for the huge amount of data coming from different sources that need to be processed efficiently
in a short time.
DVEAP: DRL-based autopilot Placement Algorithm
The AI-edge for autopilot VNFs placement provides many benefits including efficient data
processing in a short time that is the important factor for autonomous driving applications,
which by consequence improve the network performance and of course the QoS.
1. The Proposed RL Model
We propose a distributed edge servers that process autopilot VNFs coming from
autonomous vehicles. At each time, the edge manager chose the efficient server to
place the VNFs. We propose an RL model to manage the VNFs placement at the
network edge. The different RL steps are described as follows:
The state space: the state st represents the current placement of autopilot VNFs on
server slots.
The action space: the action at represents the placement process of the autopilot VNF
on the server taking into account the server capacity.
The reward space: represent the placement cost r. It is measured as the number of
used servers after the VNF placement, is formulated as follows:

i

if i Edge Servers are Occupied
rt = (4.4)
0

Otherwise.
As shown in the equation 4.4, the objective of the agent is to minimize the used servers
by reducing the total rewards.
2. The Proposed DRL Algorithm
The deep neural network is used in the DRL algorithm by adding a succession of neural
network layers to map the input to the output (state to action). The main objective of
the DRL algorithm is to reduce the placement complexity compared to the optimal
OVEAP model (3.51). In algorithm 6 we present the DRL details.
110
Algorithm 6 DRL-based autopilot Placement (DVEAP) at The Network Edge [5].

1: Input: E A , V N F , Lea , M E C
2: Output: Q∗ (state(s),action(a))
3: Initialize a reply memory D
4: Initialise action-value Q with random weights
5: Observe the initial state s
6: repeat
7: Select a server a .
• with probability ε select a random server

• Otherwise select the server that has the maxa′ Q(s, a′ )
8: Place the autopilot SFC’s VNF on the Edge Computing Server a.
9: r, s′ ← Observe the incurred allocation cost r and the new edge state s′ (new allocation and another
incoming autopilot SFC’s VNFs)
10: Store the experience {s, a, r, s′ } in the replay memory.
11: Sample a random transition from the replay memory.
12: Calculate the target for each mini-batch transition (r + γ × maxa Q(s′ , a′ ))
13: Train the Q network using the following loss Loss = 21 ∗ (r + γ × maxa′ Q(s′ , a′ ) − Q(s, a))2
14: s = s’
15: until No incoming autopilot VNFs from all the SFC
In figure 4.7 we compare the DVEAP algorithm with the exact OVEAP model (3.51) in terms
of offloading time using the parameters in Tab.3.10 and Tab.3.11. From the figure we show
that the DVEAP decrease the offloading time compared with the exact OVEAP model which
proves the efficiency of the DRL algorithm for autopilot VNFs offloading.
100
OVEAP
90 DVEAP
80
70
Time (Time Unit)
60
50
40
30
20
10
0
20 40 60 80 100
Number of Autonomous Vehicles
Fig. 4.7 OVEAP and DVEAP for autopilots VNFs Offloading [5].
111
4.2.3 Use Case 3: Edge Computing Assisted Autonomous UAV

The optimal OFMU model (3.61) is NP-hard which is not suitable for dense networks
especially for the video streaming application that requires a short response time. For this, an
approximate algorithm is proposed to deal with dense networks and support ultra-low latency
applications.
The Approximate FMU (AFMU) algorithm
The proposed AFMU algorithm is based on the scalable weighted set cover problem (WSC)
[122], it aim to select the UAVs with the high energy cost to satisfy the client vehicles
requests. Algorithm 7 show the main steps of the AFMU solution where C represent the set
of UAV, and α the average energy cost. The AFMU algorithm chooses the set which is the
most cost-effective.
Algorithm 7 : Heuristic algorithm for FMU (AFMU) [3].

1: Input: C, E, S
2: Output: Picked UAV sets
3: C←0
4: while C != E do
5: Find the UAV whose energy cost effectiveness is smallest, say S
6: Let α = Cost(S)
|S−C|
7: For eachS e ∈ S −C, set cost (e) = α
8: C←C S
9: end while
Table 4.4 show the complexity and the run-time of both OFMU and AFMU algorithms.
Table 4.4 OFMU and AFMU Algorithms Complexity.
Metrics OFMU AFMU

Algorithm NP-Hard O(|E| ∗ log|S|)
complexity
Run-time A few seconds in A few second in
Small Scale Small/Large scale
Network; a few Network
minutes in Large
Scale Network
112
To assess the performance of the proposed AFMU algorithm, we compare the obtained
results with the optimal OFMU model (3.61). Figure 4.8 shows the algorithms evaluation for
both selection time and energy cost.
Figure 4.8a displays that the OFMU algorithm is efficient because it requires a short time
for UAV selection. In addition, the AFMU algorithm gives a good UAV selection time.
Figure 4.8b shows that the OFMU outperforms the approximate AFMU algorithm in
terms of cost, while both are feasible for UAV selection.
0.005 12
OFMU OFMU
AFMU AFMU
0.0045
10
0.004 UAVs Selection Energy Cost (slots)
0.0035
UAVs Selection Time (sec)
0.003
0.0025 6
0.002
4
0.0015
0.001
2
0.0005
0 0
2 4 6 8 10 2 4 6 8 10
Number of UAVs The Number of UAVs
(a) UAVs Selection Time. (b) UAVs Selection Energy Cost.
Fig. 4.8 OFMU & AFMU Evaluation [3].
Conclusion
Large-scale algorithms are suitable for the high number of user requests in dense networks.
It offers efficient solutions to deal with user satisfaction especially the new generation of
applications that requires a short response time with efficient processing. We proposed
different large-scale algorithms for both service offloading and vehicular edge computing to
deal with dense networks. The common factor of these algorithms is the lower complexity
compared to the optimal models, this makes these algorithms suitable for the high number of
113
devices and users’ requests. It gives good results for different metrics such as time and cost,
where it was efficient and feasible for all the studied problems.
114
Chapter 5
Conclusion and Perspectives
This thesis presents edge computing for the Internet of things technology. The first chapter
presents the general introduction. We reviewed in the second chapter the studied technologies,
Internet of things, cloud computing, mobile cloud computing, edge computing, and mobile
edge computing. After that, we detailed the relation and the evolution of these technologies.
The second part of the chapter presents a review of the related works of two major edge
computing research domains, service offloading and vehicular edge computing.
The third chapter presents the proposed optimal models to optimize the edge usage for
different metrics such as time and resource utilization. We proposed optimal models for tasks,
VNFs, and SFCs orchestration at the network edge (VEnPA, SO-VMEC, OSPV) by taking
into account the computing resources constraints. Moreover, we proposed other optimal
models for vehicular edge computing including maximum coverage (MVEC) for online
video streaming applications by using taxis as mobile edge servers inside the city, optimal
edge-autopilot VNFs offloading at the network edge for autonomous driving (OVEAP), and
UAV-edge model (OFMU) for online streaming by using autonomous UAV as mobile edge
servers. The numerical results show that the proposed algorithm gives good results in terms
of resource utilization at the network edge for small-scale networks.
In chapter four, to deal with dense networks we proposed large-scale algorithms that support
a huge amount of devices, data, and user requests. Heuristic algorithms are proposed for
SFC orchestration (ESPV), maximize the vehicles coverage (AFMU) by mobile edge servers
(Taxis and UAV). Moreover, The artificial intelligence algorithms (Q-learning and Deep
Q-learning (DQN)) are used for VNFs placement, edge-autopilot VNFs placement (DVEAP),
115
Conclusion and Perspectives
and autonomous UAV navigation. The obtained results prove the efficiency and the feasibility
of the proposed solutions compared to optimal algorithms.
Despite the existing solutions outlined above, several concerns and challenges must be
investigated in order to fully integrate edge computing on top of IoT applications. The
following challenges examine the major issues, as well as new concepts and guidelines that
the research community should take seriously in future works.
• Service availability must be ensured for the next generation of edge computing applications
through the use of various mechanisms such as monitoring, prediction, and system backups.
• High mobility is an important factor in the Edge-IoT system because most connected
devices, such as vehicles, drones, and mobile devices, are highly mobile, resulting in frequent
service degradations and link failure between servers and devices which decreases the QoS
of the Edge-IoT system. For this, efficient and developed models for obstacles detection and
traffic prediction, etc. need to be proposed to overcome the mobility issue.
• The edge computing architecture is made up of several distributed systems, energy con-
sumption is projected to be considerable, raising expenses. As a result, many efforts must
be made to tackle this problem by developing efficient energy models in edge computing
systems, particularly virtual and Adhoc edge systems, such as resource optimization, Relying
more on environmentally friendly energy, etc.
116
References
[1] Mohammed Laroui, Hatem Ibn-Khedher, Moussa Ali Cherif, Hassine Moungla, Hos-
nications Technologies, page e4211, 2021.
[2] Mohammed Laroui, Boubakr Nour, Hassine Moungla, Moussa A Cherif, Hossam
Afifi, and Mohsen Guizani. Edge and fog computing for IoT: A survey on current
research activities & future directions. Computer Communications, 180:210–231,
2021.
[3] Mohammed Laroui, Hatem Ibn-Khedher, Hassine Moungla, and Hossam Afifi. Au-
tonomous UAV Aided Vehicular Edge Computing for Service Offering. In IEEE
Global Communications Conference (GLOBECOM), pages 1–6, 2021.
[4] Mohammed Laroui, Hatem Ibn-Khedher, Hassine Moungla, and Hossam Afifi. Artifi-
Edge. In IEEE International Conference on Communications (ICC), pages 1–6, 2021.
[5] Hatem Ibn-Khedher, Mohammed Laroui, Mouna Ben Mabrouk, Hassine Moungla,
Communications and Mobile Computing (IWCMC), pages 254–259, 2021.
[6] Mohammed Laroui, Hatem Ibn Khedher, Hassine Moungla, Hossam Afifi, and
Ahmed E Kamal. Virtual Mobile Edge Computing based on IoT devices resources in
smart cities. In IEEE International Conference on Communications (ICC), pages 1–6,
2020.
[7] Mohammed Laroui, Moussa Ali Cherif, Hatem Ibn Khedher, Hassine Moungla, and
117
References

Mobile Computing (IWCMC), pages 946–951, 2020.
[8] Mohammed Laroui, Boubakr Nour, Hassine Moungla, Hossam Afifi, and Moussa Ali
Cherif. Mobile vehicular edge computing architecture using rideshare taxis as a
mobile edge server. In IEEE 17th Annual Consumer Communications & Networking
Conference (CCNC), pages 1–2, 2020.
[9] Mohammed Laroui, Aicha Dridi, Hossam Afifi, Hassine Moungla, Michel Marot,
and Moussa Ali Cherif. Energy management for electric vehicles in smart cities: a
deep learning approach. In IEEE International Wireless Communications & Mobile
Computing Conference (IWCMC), pages 2080–2085, 2019.
[10] Mohammed Laroui, Akrem Sellami, Boubakr Nour, Hassine Moungla, Hossam Afifi,
and Sofiane B Hacene. Driving path stability in VANETs. In IEEE Global Communi-
cations Conference (GLOBECOM), pages 1–6, 2018.
[11] Maria Stoyanova, Yannis Nikoloudakis, Spyridon Panagiotakis, Evangelos Pallis,

and Evangelos K Markakis. A Survey on the Internet of Things (IoT) Forensics:
Challenges, Approaches and Open Issues. IEEE Communications Surveys & Tutorials,
2020.
[12] Asif Ali Laghari, Kaishan Wu, Rashid Ali Laghari, Mureed Ali, and Abdullah Ayub
Khan. A review and state of art of Internet of Things (IoT). Archives of Computational
Methods in Engineering, pages 1–19, 2021.
[13] Yazdan Ahmad Qadri, Ali Nauman, Yousaf Bin Zikria, Athanasios V Vasilakos, and
Sung Won Kim. The Future of Healthcare Internet of Things: A Survey of Emerging
Technologies. IEEE Communications Surveys & Tutorials, 22(2):1121–1167, 2020.
[14] Koen Tange, Michele De Donno, Xenofon Fafoutis, and Nicola Dragoni. A systematic
survey of industrial Internet of Things security: Requirements and fog computing
opportunities. IEEE Communications Surveys & Tutorials, 2020.
[15] Ermal Elbasani, Pattamaset Siriporn, and Jae Sung Choi. A Survey on RFID in
Industry 4.0. In Internet of Things for Industry 4.0, pages 1–16. Springer, 2020.
[16] Vedat Coskun, Busra Ozdenizci, and Kerem Ok. A survey on near field communication
(NFC) technology. Wireless personal communications, 71(3):2259–2294, 2013.
118
References
[17] Mobeen Shahroz, Muhammad Faheem Mushtaq, Maqsood Ahmad, Saleem Ullah,
Arif Mehmood, and Gyu Sang Choi. IoT-Based Smart Shopping Cart Using Radio
Frequency Identification. IEEE Access, 8:68426–68438, 2020.
[18] Abdulrahman Abuelkhail, Uthman Baroudi, Muhammad Raad, and Tarek Sheltami.
Internet of things for healthcare monitoring applications based on RFID clustering
scheme. Wireless Networks, pages 1–17, 2020.
[19] Rahul Priyadarshi, Bharat Gupta, and Amulya Anurag. Deployment techniques in
wireless sensor networks: a survey, classification, challenges, and future research
issues. The Journal of Supercomputing, pages 1–41, 2020.
[20] Naveenkumar Jayakumar and Dhanashri P Joshi. Big Data & Disruptive Computing
Platforms Braced Internet of Things: Facets & Trends. In Internet of Things, Smart
Computing and Technology: A Roadmap Ahead, pages 119–150. Springer, 2020.
[21] Wan Haslina Hassan et al. Current research on Internet of Things (IoT) security: A
survey. Computer networks, 148:283–294, 2019.
[22] Thomas Welsh and Elhadj Benkhelifa. On Resilience in Cloud Computing: A survey
of techniques across the Cloud Domain. ACM Computing Surveys (CSUR), 53(3):1–36,
2020.
[23] Keke Gai, Jinnan Guo, Liehuang Zhu, and Shui Yu. Blockchain Meets Cloud Com-
puting: A Survey. IEEE Communications Surveys & Tutorials, 2020.
[24] Peter Mell, Tim Grance, et al. The NIST definition of cloud computing. 2011.
[25] Michael Cusumano. Cloud computing and SaaS as new computing platforms. Com-
munications of the ACM, 53(4):27–29, 2010.
[26] Claus Pahl. Containerization and the paas cloud. IEEE Cloud Computing, 2(3):24–31,
2015.
[27] Hamid Talebian, Abdullah Gani, Mehdi Sookhak, Ahmed Abdelaziz Abdelatif, Ab-
dullah Yousafzai, Athanasios V Vasilakos, and Fei Richard Yu. Optimizing virtual
machine placement in IaaS data centers: taxonomy, review and open issues. Cluster
Computing, 23(2):837–878, 2020.
[28] Alessio Botta, Walter De Donato, Valerio Persico, and Antonio Pescapé. Integration
of cloud computing and internet of things: a survey. Future generation computer
systems, 56:684–700, 2016.
119
References
[29] Gabriel Neagu, Ştefan Preda, Alexandru Stanciu, and Vladimir Florian. A Cloud-IoT
based sensing service for health monitoring. In IEEE E-Health and Bioengineering
Conference (EHB), pages 53–56, 2017.
[30] Leila Ismail and Huned Materwala. Energy-aware vm placement and task scheduling
in cloud-iot computing: Classification and performance evaluation. IEEE Internet of
Things Journal, 5(6):5166–5176, 2018.
[31] Niroshinie Fernando, Seng W Loke, and Wenny Rahayu. Mobile cloud computing: A
survey. Future generation computer systems, 29(1):84–106, 2013.
[32] Mazliza Othman, Sajjad Ahmad Madani, Samee Ullah Khan, et al. A survey of mobile
cloud computing application models. IEEE communications surveys & tutorials,
16(1):393–413, 2013.
[33] M Reza Rahimi, Jian Ren, Chi Harold Liu, Athanasios V Vasilakos, and Nalini
Venkatasubramanian. Mobile cloud computing: A survey, state of art and future
directions. Mobile Networks and Applications, 19(2):133–143, 2014.
[34] Hoang T Dinh, Chonho Lee, Dusit Niyato, and Ping Wang. A survey of mobile cloud
computing: architecture, applications, and approaches. Wireless communications and
mobile computing, 13(18):1587–1611, 2013.
[35] Jianli Pan and James McElhannon. Future edge cloud and edge computing for internet
of things applications. IEEE Internet of Things Journal, 5(1):439–449, 2017.
[36] Weisong Shi, Jie Cao, Quan Zhang, Youhuizi Li, and Lanyu Xu. Edge computing:
Vision and challenges. IEEE internet of things journal, 3(5):637–646, 2016.
[37] Latif U Khan, Ibrar Yaqoob, Nguyen H Tran, SM Ahsan Kazmi, Tri Nguyen Dang,
and Choong Seon Hong. Edge computing enabled smart cities: A comprehensive
survey. IEEE Internet of Things Journal, 2020.
[38] Zhuoqing Chang, Shubo Liu, Xingxing Xiong, Zhaohui Cai, and Guoqing Tu. A
Survey of Recent Advances in Edge-Computing-Powered Artificial Intelligence of
Things. IEEE Internet of Things Journal, 2021.
[39] Mohammad Aazam and Eui-Nam Huh. Fog computing: The cloud-iot\/ioe middle-
ware paradigm. IEEE Potentials, 35(3):40–44, 2016.
120
References
[40] Mung Chiang, Sangtae Ha, Fulvio Risso, Tao Zhang, and I Chih-Lin. Clarifying
fog computing and networking: 10 questions and answers. IEEE Communications
Magazine, 55(4):18–20, 2017.
[41] Andrzej M Goscinski, Zahir Tari, Izzatdin A Aziz, and Eidah J Alzahrani. Fog
computing as a critical link between a central cloud and iot in support of fast discovery
of new hydrocarbon reservoirs. In International Conference on Mobile Networks and
Management, pages 247–261. Springer, 2017.
[42] Yi Pan, Parimala Thulasiraman, and Yingwei Wang. Overview of cloudlet, fog com-
puting, edge computing, and dew computing. In Proceedings of The 3rd International
Workshop on Dew Computing, pages 20–23, 2018.
[43] Mahadev Satyanarayanan. The emergence of edge computing. IEEE Computer,

50(1):30–39, 2017.
[44] Shanhe Yi, Zijiang Hao, Zhengrui Qin, and Qun Li. Fog computing: Platform and
applications. In IEEE Workshop on Hot Topics in Web Systems and Technologies
(HotWeb), pages 73–78, 2015.
[45] Flavio Bonomi, Rodolfo Milito, Jiang Zhu, and Sateesh Addepalli. Fog computing
and its role in the internet of things. In Proceedings of the first edition of the MCC
workshop on Mobile cloud computing, pages 13–16, 2012.
[46] Behailu Negash, Amir M Rahmani, Pasi Liljeberg, and Axel Jantsch. Fog computing
fundamentals in the internet-of-things. In Fog computing in the internet of things,
pages 3–13. Springer, 2018.
[47] Yi Qian, Dan Wu, Wei Bao, and Pascal Lorenz. The internet of things for smart cities:
Technologies and applications. IEEE Network, 33(2):4–5, 2019.
[48] Dingfu Jiang. The construction of smart city information system based on the Internet
of Things and cloud computing. Computer Communications, 150:158–166, 2020.
[49] Sandeep Saharan, Neeraj Kumar, and Seema Bawa. An efficient smart parking pricing
system for smart city environment: A machine-learning based approach. Future
Generation Computer Systems, 106:622–640, 2020.
[50] Wei Liu, Yoshito Watanabe, and Yozo Shoji. Vehicle-Assisted Data Delivery in Smart
City: A Deep Learning Approach. IEEE Transactions on Vehicular Technology,
69(11):13849–13860, 2020.
121
References
[51] R Dhaya, R Kanthavel, Fahad Algarni, P Jayarajan, and Amita Mahor. Reinforcement
Learning Concepts Ministering Smart City Applications Using IoT. In Internet of
Things in Smart Technologies for Sustainable Urban Development, pages 19–41.
Springer, 2020.
[52] Haidong Luo, Hongming Cai, Han Yu, Yan Sun, Zhuming Bi, and Lihong Jiang. A
short-term energy prediction system based on edge computing for smart city. Future
[53] Adam Zielonka, Andrzej Sikora, Marcin Woźniak, Wei Wei, Qiao Ke, and Zongwen
Bai. Intelligent Internet-of-Things system for smart home optimal convection. IEEE
Transactions on Industrial Informatics, 2020.
[54] Dae-Man Han and Jae-Hyun Lim. Design and implementation of smart home energy
management systems based on zigbee. IEEE Transactions on Consumer Electronics,
56(3):1417–1425, 2010.
[55] Danni Yuan, Kaoru Ota, Mianxiong Dong, Xiaoyan Zhu, Tao Wu, Linjie Zhang, and
Jianfeng Ma. Intrusion Detection for Smart Home Security Based on Data Augmenta-
tion with Edge Computing. In IEEE International Conference on Communications
(ICC), pages 1–6, 2020.
[56] Anila Yasmeen, Nadeem Javaid, Obaid Ur Rehman, Hina Iftikhar, Muhammad Faizan
Malik, and Fatima J Muhammad. Efficient resource provisioning for smart build-
ings utilizing fog and cloud based environment. In IEEE International Wireless
Communications & Mobile Computing Conference (IWCMC), pages 811–816, 2018.
[57] Wei-Yu Chen, Po-Yu Chou, Chih-Yu Wang, Ren-Hung Hwang, and Wen-Tsuen Chen.
Live Video Streaming with Joint User Association and Caching Placement in Mobile
Edge Computing. In IEEE International Conference on Computing, Networking and
Communications (ICNC), pages 796–801, 2020.
[58] Shaojun Zhang, Wei Li, Yongwei Wu, Paul Watson, and Albert Zomaya. Enabling
edge intelligence for activity recognition in smart homes. In IEEE International
Conference on Mobile Ad Hoc and Sensor Systems (MASS), pages 228–236, 2018.
[59] Baofeng Ji, Xueru Zhang, Shahid Mumtaz, Congzheng Han, Chunguo Li, Hong
Wen, and Dan Wang. Survey on the Internet of Vehicles: Network Architectures and
Applications. IEEE Communications Standards Magazine, 4(1):34–41, 2020.
122
References
[60] Nhien-An Le-Khac, Daniel Jacobs, John Nijhoff, Karsten Bertens, and Kim-
Kwang Raymond Choo. Smart vehicle forensics: Challenges and case study. Future
[61] Jaiveer Singh and Karan Singh. Congestion control in vehicular ad hoc network: A
review. In Next-generation networks, pages 489–496. Springer, 2018.
[62] Jian Liu, Jiangtao Li, Lei Zhang, Feifei Dai, Yuanfei Zhang, Xinyu Meng, and Jian
Shen. Secure intelligent traffic light control using fog computing. Future Generation
Computer Systems, 78:817–824, 2018.
[63] G Revathi and VR Sarma Dhulipala. Smart parking systems and sensors: A survey.
In IEEE International Conference on Computing, Communication and Applications,
pages 1–5, 2012.
[64] Kai Lin, Jiming Luo, Long Hu, M Shamim Hossain, and Ahmed Ghoneim. Localiza-
tion based on social big data analysis in the vehicular networks. IEEE Transactions
on Industrial Informatics, 13(4):1932–1940, 2016.
[65] Yeliz Yoldaş, Ahmet Önen, SM Muyeen, Athanasios V Vasilakos, and İrfan Alan.
Enhancing smart grid with microgrids: Challenges and opportunities. Renewable and
Sustainable Energy Reviews, 72:205–214, 2017.
[66] Kemas M Tofani, PAA Pramana, BBSDA Harsono, Dhandis R Jintaka, and KG H
Mangunnkusumo. SCADA Systems Design to Optimize and Automate Microgrids
Systems in Indonesia. In IEEE International Conference on Technology and Policy in
Energy and Electric Power (ICT-PEP), pages 187–192, 2020.
[67] Farzad Samie, Lars Bauer, and Jörg Henkel. Edge computing for smart grid: An
overview on architectures and solutions. In IoT for Smart Grids, pages 21–42. Springer,
2019.
[68] Shehzad Ashraf Chaudhry, Hosam Alhakami, Abdullah Baz, and Fadi Al-Turjman.
Securing Demand Response Management: A Certificate based Access Control in
Smart Grid Edge Computing Infrastructure. IEEE Access, 2020.
[69] WenJing Hou, Yixin Jiang, Wenxin Lei, Aidong Xu, Hong Wen, and Songling Chen.
A P2P network based edge computing smart grid model for efficient resources coordi-
nation. Peer-to-Peer Networking and Applications, pages 1–12, 2020.
123
References
[70] Mohamed Amine Ferrag, Messaoud Babaghayou, and Mehmet Akif Yazici. Cyber
security for fog-based smart grid SCADA systems: Solutions and challenges. Journal
of Information Security and Applications, 52:102500, 2020.
[71] Prabal Verma and Sandeep K Sood. Fog assisted-IoT enabled patient health monitoring
in smart homes. IEEE Internet of Things Journal, 5(3):1789–1796, 2018.
[72] Lakmini P Malasinghe, Naeem Ramzan, and Keshav Dahal. Remote patient mon-
itoring: a comprehensive study. Journal of Ambient Intelligence and Humanized
Computing, 10(1):57–76, 2019.
[73] Yuyi Mao, Changsheng You, Jun Zhang, Kaibin Huang, and Khaled B Letaief. A
survey on mobile edge computing: The communication perspective. IEEE Communi-
cations Surveys & Tutorials, 19(4):2322–2358, 2017.
[74] Nasir Abbas, Yan Zhang, Amir Taherkordi, and Tor Skeie. Mobile edge computing: A
survey. IEEE Internet of Things Journal, 5(1):450–465, 2017.
[75] ETSI MEC Leadership Team. ETSI MEC. https://www.etsi.org/images/files/

technologies/ETSI-MEC-Public-Overview.pdf. [Online; accessed 11/2021].
[76] Dario Sabella, Alessandro Vaillant, Pekka Kuure, Uwe Rauschenbach, and Fabio
Giust. Mobile-edge computing architecture: The role of MEC in the Internet of
Things. IEEE Consumer Electronics Magazine, 5(4):84–91, 2016.
[77] Emiliano Miluzzo, Ramón Cáceres, and Yih-Farn Chen. Vision: mClouds-computing
on clouds of mobile devices. pages 9–14. Proceedings of the third ACM workshop on
Mobile cloud computing and services, 2012.
[78] Afnan Fahim, Abderrahmen Mtibaa, and Khaled A Harras. Making the case for
computational offloading in mobile device clouds. pages 203–205. Proceedings of
the 19th annual international conference on Mobile computing & networking (ACM),
2013.
[79] Duc Van Le and Chen-Khong Tham. An optimization-based approach to offloading

in ad-hoc mobile clouds. pages 1–6. IEEE Global Communications Conference
(Globecom), 2017.
[80] Ragib Hasan, Mahmud Hossain, and Rasib Khan. Aura: An incentive-driven ad-hoc
IoT cloud framework for proximal mobile computation offloading. Future Generation
124
References
[81] Duc Van Le and Chen-Khong Tham. A deep reinforcement learning based offloading
scheme in ad-hoc mobile clouds. pages 760–765. IEEE Conference on Computer
Communications Workshops (INFOCOM WKSHPS), 2018.
[82] Duc Van Le and Chen-Khong Tham. Quality of service aware computation offloading
in an ad-hoc mobile cloud. IEEE Transactions on Vehicular Technology, 67(9):8890–
8904, 2018.
[83] Md Golam Rabiul Alam, Mohammad Mehedi Hassan, Md ZIa Uddin, Ahmad Almo-
gren, and Giancarlo Fortino. Autonomic computation offloading in mobile edge for
IoT applications. Future Generation Computer Systems, 90:149–157, 2019.
[84] Ke Zhang, Yongxu Zhu, Supeng Leng, Yejun He, Sabita Maharjan, and Yan Zhang.
Deep Learning Empowered Task Offloading for Mobile Edge Computing in Urban
Informatics. IEEE Internet of Things Journal, 2019.
[85] Zongyao Wang, Xue Feng, Hongbo Zhu, and Changping Liu. Optimal data offloading
via an ADMM algorithm in mobile ad hoc cloud with malicious resource providers.
Computer Communications, 2020.
[86] Xueling Lin, Jingjie Jiang, Calvin Hong Yi Li, Bo Li, and Baochun Li. Circa:
collaborative code offloading among multiple mobile devices. Wireless Networks,
26(2):823–841, 2020.
[87] Sajeeb Saha, Md Ahsan Habib, Tamal Adhikary, Md Abdur Razzaque, Md Mustafizur
Rahman, Meteb Altaf, and Mohammad Mehedi Hassan. Quality-of-Experience-Aware
Incentive Mechanism for Workers in Mobile Device Cloud. IEEE Access, 9:95162–
95179, 2021.
[88] Palash Roy, Sujan Sarker, Md Abdur Razzaque, Md Mamun-or Rashid, Mohm-
mad Mehedi Hassan, and Giancarlo Fortino. Distributed task allocation in Mobile
Device Cloud exploiting federated learning and subjective logic. Journal of Systems
Architecture, 113:101972, 2021.
[89] Oanh Tran Thi Kim, Nguyen Dang Tri, Nguyen H Tran, Choong Seon Hong, et al.
A shared parking model in vehicular network using fog and cloud environment. In
IEEE 17th Asia-Pacific Network Operations and Management Symposium (APNOMS),
pages 321–326, 2015.
125
References
[90] Ke Zhang, Yuming Mao, Supeng Leng, Sabita Maharjan, and Yan Zhang. Opti-
mal delay constrained offloading for vehicular edge computing networks. In IEEE
International Conference on Communications (ICC), pages 1–6, 2017.
[91] Chao Zhu, Giancarlo Pastor, Yu Xiao, Yong Li, and Antti Ylae-Jaeaeski. Fog following
me: Latency and quality balanced task allocation in vehicular fog computing. In IEEE
International Conference on Sensing, Communication, and Networking (SECON),
pages 1–9, 2018.
[92] Lee Gillam, Konstantinos Katsaros, Mehrdad Dianati, and Alexandres Mouzakitis.
Exploring edges for connected and autonomous driving. In IEEE Conference on
Computer Communications Workshops (INFOCOM WKSHPS), pages 148–153, 2018.
[93] Pedro Henrique Cruz Caminha, Felipe Ferreira da Silva, Roberto Gonçalves Pacheco,
Rodrigo de Souza Couto, Pedro Braconnot Velloso, Miguel Elias Mitre Campista, and
Luís Henrique MK Maciel Kosmalski Costa. Sensingbus: Using bus lines and fog
computing for smart sensing the city. IEEE Cloud Computing, 5(5):58–69, 2018.
[94] Arnav Thakur and Reza Malekian. Fog computing for detecting vehicular congestion,
an internet of vehicles based approach: A review. IEEE Intelligent Transportation
Systems Magazine, 11(2):8–16, 2019.
[95] Zhaolong Ning, Jun Huang, and Xiaojie Wang. Vehicular fog computing: En-
abling real-time traffic management for smart cities. IEEE Wireless Communications,
26(1):87–93, 2019.
[96] Abdallah Moubayed, Abdallah Shami, Parisa Heidari, Adel Larabi, and Richard
Brunner. Edge-enabled V2X service placement for intelligent transportation systems.
IEEE Transactions on Mobile Computing, 2020.
[97] Ibrahim Shaer, Anwar Haque, and Abdallah Shami. Multi-Component V2X Applica-
tions Placement in Edge Computing Environment. In IEEE International Conference
on Communications (ICC), pages 1–6, 2020.
[98] Jie Tang, Shaoshan Liu, Liangkai Liu, Bo Yu, and Weisong Shi. Lopecs: A low-power
edge computing system for real-time autonomous driving services. IEEE Access,
8:30467–30479, 2020.
[99] Ahmad Raza Hameed, Saif ul Islam, Ishfaq Ahmad, and Kashif Munir. Energy-
and performance-aware load-balancing in vehicular fog computing. Sustainable
Computing: Informatics and Systems, 30:100454, 2021.
126
References
[100] Fuhui Zhou, Yongpeng Wu, Haijian Sun, and Zheng Chu. UAV-enabled mobile edge
computing: Offloading optimization and trajectory design. In IEEE International
Conference on Communications (ICC), pages 1–6, 2018.
[101] Qiyu Hu, Yunlong Cai, Guanding Yu, Zhijin Qin, Minjian Zhao, and Geoffrey Ye
Li. Joint offloading and trajectory design for UAV-enabled mobile edge computing
systems. IEEE Internet of Things Journal, 6(2):1879–1892, 2018.
[102] Jingyu Xiong, Hongzhi Guo, and Jiajia Liu. Task offloading in UAV-aided edge
computing: Bit allocation and trajectory optimization. IEEE Communications Letters,
23(3):538–541, 2019.
[103] Cheng Zhan, Han Hu, Xiufeng Sui, Zhi Liu, and Dusit Niyato. Completion time and
energy optimization in the uav-enabled mobile-edge computing system. IEEE Internet
of Things Journal, 7(8):7808–7822, 2020.
[104] Gaoxiang Wu, Yiming Miao, Yu Zhang, and Ahmed Barnawi. Energy efficient
for UAV-enabled mobile edge computing networks: Intelligent task prediction and
offloading. Computer Communications, 150:556–562, 2020.
[105] Lei Yang, Haipeng Yao, Jingjing Wang, Chunxiao Jiang, Abderrahim Benslimane,
and Yunjie Liu. Multi-UAV-enabled load-balance mobile-edge computing for IoT
networks. IEEE Internet of Things Journal, 7(8):6898–6908, 2020.
[106] Xiaoheng Deng, Yajun Liu, Congxu Zhu, and Honggang Zhang. Air–Ground Surveil-
lance Sensor Network based on edge computing for target tracking. Computer Com-
munications, 166:254–261, 2021.
[107] Quan Chen, Hai Zhu, Lei Yang, Xiaoqian Chen, Sofie Pollin, and Evgenii Vinogradov.
Edge Computing Assisted Autonomous Flight for UAV: Synergies between Vision
and Communications. IEEE Communications Magazine, 59(1):28–33, 2021.
[108] Mobile Edge Computing (MEC), Deployment of Mobile Edge Computing in an NFV
environment. ETSI Group Report MEC 017, 2018.
[109] Cagatay Sonmez, Atay Ozgovde, and Cem Ersoy. Edgecloudsim: An environment
for performance evaluation of edge computing systems. Transactions on Emerging
Telecommunications Technologies, 29(11):e3493, 2018.
[110] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computa-
tion, 9(8):1735–1780, 1997.
127
References
[111] Sylvia Todorova Kouyoumdjieva, lafur Ragnar Helgason, and Gunnar Karlsson.
CRAWDAD dataset kth/walkers (v. 2014-05-05). Downloaded from https://crawdad.
org/kth/walkers/20140505/ostermalm_sparse_run2, May 2014. traceset: oster-
malm_sparse_run2.
[112] Daniel Dias and Luis Henrique Maciel Kosmalski Costa. CRAWDAD dataset coppe-
ufrj/riobuses (v. 2018-03-19). Downloaded from https://crawdad.org/coppe-ufrj/
RioBuses/20180319/RioBuses, March 2018. traceset: RioBuses.
[113] Lorenzo Bracciale, Marco Bonola, Pierpaolo Loreti, Giuseppe Bianchi, Raul Amici,
and Antonello Rabuffi. CRAWDAD dataset roma/taxi (v. 2014-07-17). Downloaded
from https://crawdad.org/roma/taxi/20140717/taxicabs, July 2014. traceset: taxicabs.
[114] Martin Pielot. CRAWDAD dataset telefonica/mobilephoneuse (v. 2019-04-29). Down-

loaded from https://crawdad.org/telefonica/mobilephoneuse/20190429, April 2019.
[115] Andreas Reinhardt, Paul Baumann, Daniel Burgstahler, Matthias Hollick, Hristo
Chonov, Marc Werner, and Ralf Steinmetz. On the accuracy of appliance identification
based on distributed load metering data. pages 1–9. IEEE Sustainable Internet and
ICT for Sustainability (SustainIT), 2012.
[116] Changho Shin, Eunjung Lee, Jeongyun Han, Jaeryun Yim, Wonjong Rhee, and
Hyoseop Lee. The ENERTALK dataset, 15 Hz electricity consumption data from 22
houses in Korea. Scientific data, 6(1):1–13, 2019.
[117] Richard M Karp. Reducibility among combinatorial problems. In Complexity of

computer computations, pages 85–103. Springer, 1972.
[118] Bai-cun Zhou, Cong-ying Han, and Tian-de Guo. Convergence of stochastic gradient
descent in deep neural network. Acta Mathematicae Applicatae Sinica, English Series,
37(1):126–136, 2021.
[119] Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage, and Anil Anthony Bharath.
Deep reinforcement learning: A brief survey. IEEE Signal Processing Magazine,
34(6):26–38, 2017.
[120] Michael R Garey and David S Johnson. Approximation algorithms for bin pack-
ing problems: A survey. In Analysis and design of algorithms in combinatorial
optimization, pages 147–172. Springer, 1981.
128
References
[121] Panpan Jin, Xincai Fei, Qixia Zhang, Fangming Liu, and Bo Li. Latency-aware
VNF Chain Deployment with Efficient Resource Reuse at Network Edge. In IEEE
International Conference on Computer Communications (INFOCOM), pages 267–276,
2020.
[122] Fatema Akhter. A heuristic approach for minimum set cover problem. Int. J. Adv. Res.
Artif. Intell, 4:40–45, 2015.
129
Thesis Summary
Mr . Mohammed Laroui
Distributed Edge Computing For

Enhanced IoT Devices and New
Generation Network Efficiency
Thesis Committee :
Pr. Hassine Moungla University of Paris Cite Thesis Director
Dr. Zohra Slama UDL SBA Supervisor
Pr. Hind Castel Telecom SudParis Reviewer
Pr. Ken Chen Sorbonne University Reviewer
Pr. Hossam Afifi Telecom SudParis Examiner
Pr. Badr Benmammar University of Tlemcen Examiner
Dr. Nassim Dennouni University of Chlef Examiner
Year : 2022
Université Paris Cité
En cotutelle avec L’Université Djillali Liabes
de Sidi Bel Abbes
École Doctorale Informatique, Télécommunication et Électronique «Edite de Paris : ED130»
Laboratoire d'Informatique Paris Descartes (LIPADE)
Résumé
Distributed Edge Computing For Enhanced IoT

Devices and New Generation Network
Efficiency
Par Mohammed Laroui
Thèse de doctorat en Informatique (Réseaux)
Dirigée par Hassine Moungla

Et par Zohra Slama
Présentée et soutenue publiquement le 21 Juillet 2022
Devant un jury composé de :
Hind Castel, Professeur, Telecom SudParis, Rapporteur
Ken Chen, Professeur, Sorbonne Université, Rapporteur
Hossam Afifi, Professeur, Telecom SudParis, Examinateur
Badr Benmammar, Professeur, Université de Tlemcen, Examinateur
Nassim Dennouni, Docteur, Université de Chlef, Examinateur

Travaux de Recherche Publié
1. Mohammed Laroui, Hatem Ibn-Khedher, Moussa Ali Cherif, Hassine Moungla, Hos-
nications Technologies (ETT), e4211, 2021. [1]
2. Mohammed Laroui, Boubakr Nour, Hassine Moungla, Moussa Ali Cherif, Hossam
Afifi, and Mohsen Guizani. Edge and Fog Computing for IoT: A Survey on Current
Research Activities & Future Directions. Computer Communications (ComCom), 2021.
[2]
3. Mohammed Laroui, Hatem Ibn-Khedher, Hassine Moungla, and Hossam Afifi. Au-
Communications Conference (Globecom), 7-11 December 2021, Madrid, Spain. [3]
4. Mohammed Laroui, Hatem Ibn-Khedher, Hassine Moungla, and Hossam Afifi. Artifi-
Edge. In IEEE International Conference on Communications (ICC), (pp.1-6), 14-23
June 2021, Montreal, QC, Canada. [4]
5. Hatem Ibn-Khedher, Mohammed Laroui, Mouna Ben Mabrouk, Hassine Moungla,

Communications and Mobile Computing (IWCMC), (pp.254-259), 28 June-2 July 2021,
Harbin City, China. [5]
6. Mohammed Laroui, Hatem Ibn Khedher, Hassine Moungla, Hossam Afifi, and Ahmed
E Kamal. Virtual Mobile Edge Computing Based on IoT Devices Resources in Smart
Cities. In IEEE International Conference on Communications (ICC), (pp.1-6), 7-11
June 2020, Dublin, Ireland. [6]
7. Mohammed Laroui, Moussa Ali Cherif, Hatem Ibn Khedher, Hassine Moungla, and
i
Mobile Computing (IWCMC), (pp.946-951), 15-19 June 2020, Limassol, Cyprus. [7]
8. Mohammed Laroui, Boubakr Nour, Hassine Moungla, Hossam Afifi, and Moussa Ali
Cherif. Mobile Vehicular Edge Computing Architecture using Rideshare Taxis as a
Mobile Edge Server. In IEEE 17th Annual Consumer Communications & Networking
Conference (CCNC), (pp.1-2), 10-13 January 2020, Las Vegas, NV, USA. [8]
9. Mohammed Laroui, Aicha Dridi, Hossam Afifi, Hassine Moungla, Michel Marot, and
Moussa Ali Cherif. Energy Management For Electric Vehicles in Smart Cities: A Deep
Learning Approach. In IEEE International Wireless Communications and Mobile
Computing (IWCMC), (pp.2080-2085), 24-28 June 2019, Tangier, Morocco. [9]
10. Mohammed Laroui, Akrem Sellami, Boubakr Nour, Hassine Moungla, Hossam Afifi,
Sofiane Boukli Hacene. Driving Path Stability in VANETs. In IEEE Global Communi-
cations Conference (GLOBECOM), (pp.1-6), 9-13 December 2018, Abu Dhabi, United
Arab Emirates. [10]
ii
1. Introduction générale
1.1 Motivation
L’Internet des objets (IoT) a été un facteur déterminant dans le développement de l’industrie
de la haute technologie. Avec le grand nombre d’appareils et de profits à réaliser au cours
des prochaines années, L’IoT aura un impact profond et dominant sur l’industrie de la
haute technologie en général et sur l’industrie des semi-conducteurs en particulier. Les
appareils IoT peuvent en fait être n’importe quel type de capteurs et de puces avec différentes
capacités fabriqués par différents fabricants, et il existe de nombreuses applications qui
peuvent être conçues pour permettre les villes intelligentes, les transports intelligents, les
maisons intelligentes et les soins de santé intelligents.
Dans l’IoT, les appareils connectés peuvent générer une énorme quantité de données à très
haut débit et certaines applications peuvent nécessiter une très faible latence. Les données
sont directement envoyées dans le cloud pour être stockées et traitées dans les centres de
données. L’infrastructure cloud traditionnelle sera confrontée à une série de défis en raison
de la centralisation du stockage, calcule, et de la longue distance entre les appareils connectés
et les centres de données.
Pour relever ce défi, edge computing semble être une technologie prometteuse qui fournit
des ressources de calcul plus proches des appareils IoT.
Dans cette thèse, nous sommes motivés pour identifier et résoudre les problèmes d’architecture
edge computing et l’intégration de cette architecture avec les applications IoT. De plus, la
littérature manque d’architectures générales edge-IoT, ce qui nous amène à concevoir des
architectures récentes, en particulier pour le service offloading et vehicular edge computing.
1
De plus, malgré l’importance des tâches d’optimisation, de telles fonctions efficaces
manquent dans les architectures globales edge-IoT.
1.2 Contribution
Notre contribution présente de nouveaux modèles d’optimisation différents (algorithmes
exacts et à grande échelle) au niveau de la couche edge qui résolvent à la fois les problèmes
de service offloading liés au placement dans le edge de réseau et les problèmes de vehicular
edge computing qui utilisent les véhicules comme des serveurs edge mobile pour fournir des
services aux utilisateurs tels que le calcule et le online streaming.
Tout d’abord, pour le service offloading, nous proposons des modèles de placement
optimaux dans le edge du réseau pour les tâches des utilisateurs. Ensuite, pour vehicular edge
computing, nous proposons des modèles exact pour la couverture des vehicules utilisant des
serveurs edge mobile (Taxis et UAV), et la conduite autonome basé sur le edge computing.
De plus, pour faire face aux problèmes de haute complexité des algorithmes optimaux
dans les réseaux denses, nous proposons des algorithmes à grande échelle pour les cas
d’utilisation du service offloading et vehicular edge computing.
Pour évaluer l’efficacité des algorithmes proposés pour les réseaux à petite et grande
échelle, nous utilisons différents outils, notamment, des plates-formes pour les modèles
d’intelligence artificielle et des simulateurs pour les réseaux edge.
2
2. Edge Computing pour Internet des
Objets: Definitions et État de l’art
2.1 Definitions
2.1.1 Internet des Objets (IoT)

Est une nouvelle technologie envisagée comme un réseau d’appareils et de machines qui
communiquent entre eux et avec Internet. L’IoT est connu comme l’un des catalyseurs
importants des technologies futures. Il suscite également un grand intérêt auprès des en-
treprises. Dans un sens plus large, l’IoT vise à créer des systèmes basés sur l’interconnexion
d’objets intelligents. Ces objets échangent des informations entre eux en utilisant différents
protocoles, tels que Wi-Fi, Bluetooth, ZigBee, etc. La principale caractéristique de l’IoT est
l’intégration de différentes technologies de communication (par exemple, capteurs filaires et
sans fil, etc.) pour améliorer la coopération et l’interaction entre les différentes technologies.
La figure 1 montre les domaines d’application de l’IoT.
L’augmentation rapide des appareils connectés/intelligents dans le monde et l’évolution
des besoins des utilisateurs et des applications avec une grande quantité de traitement de
données ont conduit à un ensemble de nouvelles technologies permettant un traitement rapide
des données et des services fiables. Le cloud computing fait partie de ces technologies.
2.1.2 Cloud Computing

Le cloud computing [11] vise à fournir divers services aux utilisateurs dans le cloud. Dif-
férents types de cloud peuvent être déployés, notamment privé, public, hybride et commu-
3
Smart City
Security & Surveillance Smart Home
Internet of things
FOOD CHAIN...
application domains
Industrial Automation Smart Healthcare
Smart Vehicle
Fig. 1 Les Domaines d’application de l’IoT.
nautaire. Le cloud public [12] fournit des services à un grand nombre d’utilisateurs sur
Internet. Le cloud privé [13] offre des services spécifiques aux organisations privées. Le
cloud communautaire [14] vise à fournir des services à un groupe d’organisations. Enfin,
le cloud hybride [15] permet aux organisations d’équilibrer les coûts et les problèmes de
contrôle.
La figure 2 affiche l’architecture du cloud computing.
Dans le cloud computing, les utilisateurs se connectent directement au cloud via Internet
et commencent à échanger des données sur le réseau, ce qui entraîne une quantité massive de
données en peu de temps. Cependant, l’architecture centralisée du cloud n’est pas efficace
pour traiter la quantité massive de données générées par les appareils IoT qui nécessitent
un temps de réponse court. Pour surmonter un tel problème, une technologie alternatif, à
savoir l’edge computing, a été utilisé qui traite les données dans les appareils connectés ou
les passerelles locales.
2.1.3 Edge/Fog Computing

L’edge computing vise à être la future solution IoT qui résout différents problèmes, y compris
les applications limitées dans le temps et basées sur le calcul. Les avantages du traitement
des données dans le edge du réseau sont les suivants : réduire la charge du réseau et la latence
4
Cloud Computing
Cloud
Controller
Cloud
Controller
Data Center
Data Center
Internet

BTS
Fig. 2 Architecture du Cloud Computing.
des communications, donner aux petits et moyens inventeurs toutes les chances d’aider à
soutenir les innovations futures, réduire la consommation d’énergie des nœuds mobiles et
éliminer la congestion au sein du réseau central, tout en offrant plus de fiabilité, de sécurité
et de protection de la vie privée. Dans la section suivante, nous présentons l’état de l’art pour
les deux axes de recherche étudiés, service offloading et vehicular edge computing.
La figure 3 montre l’architecture edge computing.
2.2 État de l’art

L’intégration de l’edge computing avec l’Internet des objets (Edge-IoT) fait face à de nom-
breux problèmes et défis liés à la QoS. Pour cela, différentes axes de recherche liées à
Edge-IoT ont été étudiées pour couvrir tous les modèles de communication afin de satisfaire
5
Cloud Computing
Cloud
Controller
Cloud
Controller
Data Center
Data Center
Internet
Edge Layer
Ed Ed Ed
ge ge geS
Se Se erv
rve rve er
r r

BTS
Connected Devices
Fig. 3 Architecture du Edge Computing.
les exigences des applications et fournir des services efficaces aux utilisateurs. Cette partie
présente l’état de l’art pour deux axes de recherche pour Edge-IoT; service offloading et
vehicular edge computing.
2.2.1 Service Offloading

Service offloading fait référence au déplacé certains services à forte intensité de calcul
vers les serveurs edge pour traitement. L’allocation de ressources pour cette procédure fait
6
référence à l’allocation de certaines ressources de calcul sur des serveurs edge en fonction
des exigences du service. Le déplacement des services vers le edge du réseau facilite le
stockage, la prestation de services, la mise en cache du contenu et la gestion de l’IoT, ce
qui permet d’améliorer les temps de réponse et les taux de transfert, garantissant ainsi aux
utilisateurs le service le plus rapide et le meilleur possible.
Le tableau 1 affiche le résumé des travaux de recherche existants pour le service offloading
dans l’environment edge computing.
Article de Idée princi- Formulation Position Avantages Limites

recherche pale du prob- du edge
lème
Miluzzo et Exécution Quelques Mobile Exécuter des Temps de

al. [16] des tâches équations tâches pour réponse
dans des ap- pour le d’autres util- long.
pareils mobiles paiement isateurs.
intelligents. des ser-
vices,
etc.
Fahim et al. Task offload- Aucune Mobile Consommation L’impact de

[17] ing à l’aide formulation d’énergie la mobilité
d’appareils mathéma- et temps de n’est pas
mobiles tique n’est traitement étudié.
comme nœuds utilisée. efficaces.
de calcul.
7
Van et al. Traitement des Processus Mobile Améliorer Les dé-
[18] tâches dans les décisionnel le temps tails de
cloudlets mode Markov et la con- l’implémenta-
biles. (MDP). sommation tion de la
d’énergie. solution sont
manquants.
Hasan et al. Aura, task Équations Mobile Améliorer L’impact de

[19] offloading pour la le coût et la mobilité
dans une archi- mesure le temps n’est pas
tecture cloud des coûts, d’exécution. étudié.
ad-hoc mobile. du temps
et des
ressources.
Van et al. Tasks of- - Processus Mobile Améliorer L’évaluation

[20] floading dans décisionnel la consom- dans les
un environ- de Markov mation réseaux
nement cloud (MDP). d’énergie, le à grande
mobile ad hoc. - Deep-Q- paiement et échelle n’est
Network les délais. pas étudiée.
(DQN).
Van et al. Technique de Processus Mobile Améliorer le La solution

[21] ’task offload- décisionnel taux de perte n’est pas
ing’ dans le de Markov de tâches, le étudiée dans
cloud mobile (MDP). délai moyen un envi-
ad-hoc. et la con- ronnement
sommation réel.
d’énergie.
8
Alam et al. Offloading - Processus Mobile Minimiser la L’impact de
[22] des calculs décisionnel latence. la mobilité
dans un en- de Markov n’est pas
vironnement (MDP). étudié.
mobile edge - Deep
computing. Q-learning.
Zhang et al. Task offload- Deep Fixed Task offload- La complex-

[23] ing dans un en- Q-learning. ing fiable et ité n’est pas
vironnement optimal. étudiée.
mobile.
Wang et al. Task offload- Algorithme Mobile Identifier les Le détail de

[24] ing optimal d’optimisati- fournisseurs l’implément-
dans un en- on dis- de ressources ation de la
vironnement tribué. malveillants. solution est
cloud Adhoc manquant.
mobile.
Lin et al. Circa, offload- Algorithmes Mobile Améliorer Risque élevé

[25] ing framework d’allocation le temps d’échec de la
basé sur les des tâches. d’exécution. communica-
appareils mo- tion.
biles.
9
Saha et al. Cloud frame- Modèle Mixed Faible coût. La solution
[26] work pour les optimal et n’est pas
appareils mo- algorithme étudiée dans
biles. ’greedy’ un envi-
pour le ronnement
provision- réel.
nement des
tâches.
Roy et al. Task allocation Algorithmes Mobile Efficace à La solution

[27] framework pour task la fois pour n’est pas
pour les allocation. l’utilité de étudiée dans
appareils l’acheteur et un envi-
mobile. la QoE. ronnement
à grande
échelle.
Table 1 Résumé des recherches existant sur le service offloading dans le edge computing.
Les algorithmes de service offloading dans la littérature sont basés sur la formulation
de modèles exacts pour mesurer le comportement optimal d’une telle solution. Cependant,
dans la pratique, les opérateurs de réseau conçoivent toujours des heuristiques pour traiter
les problèmes d’évolutivité dans les réseaux à grande échelle avec d’énormes quantités de
données. Actuellement, ces heuristiques sont des solutions quasi-optimales et sont parfois
très loin d’être optimales. De plus, les heuristiques ne sont pas utilisées lorsque l’entrée
du système change et donne des erreurs fatales lorsqu’une entrée inattendue arrive dans le
système.
2.2.2 Vehicular Edge Computing

Les véhicules intelligents sont considérés comme des appareils mobiles équipés de capteurs,
ayant la capacité de collecter, de calculer et de communiquer des données. Les informations
10
sont collectées à la fois à partir de capteurs embarqués dans le véhicule et de l’environnement
extérieur. Edge computing peut fournir une architecture efficace et évolutive pour les réseaux
de véhicules en améliorant le traitement des données et le trafic en temps réel. Nous
présentons dans cette partie les solutions existantes pour les réseaux véhiculaires basées sur
le edge computing (vehicular edge computing).
Le tableau 2 présente le résumé des travaux existants de vehicular edge computing.
Article de Idée princi- Formulation Position Avantages Limites

recherche pale du prob- du edge
lème
Kim et al. Gestion du ’Many- Fixed Fiabilité L’évolutivité

[28] stationnement to-one pour trouver du système
dans un en- match- une place fog n’est pas
vironnement ing game de parking étudiée.
fog-cloud. model’. disponible.
Zhang et al. Task offload- ’Stackelberg Fixed Délai opti- Le risque de

[29] ing dans un en- game’. mal pour le coupure de
vironnement calcul des liaison n’est
mobile. tâches. pas étudié.
Zhu et al. Allocation Un modèle Mixed Améliorer la Nécessite

[30] des taches exact pour latence. une
dans un en- le problème ressource
vironnement d’allocation de calcul
mobile. des tâches. élevée.
Gillam et al. Architecture N/A Fixed Réduire la la- Aucune éval-

[31] de calcul tence. uation n’est
distribuée pour présentée.
les véhicules.
11
Caminha et SensingBus, N/A Mobile Améliore Les bus ne
al. [32] un système l’utilisation peuvent pas
de collecte de du pro- assurer une
données à par- cesseur et de large couver-
tir de capteurs la mémoire. ture.
sur les bus
urbains.
Thakur et Techniques N/A Fixed Le système La solution

al. [33] d’identification peut détecter nécessite une
de la conges- la conges- évaluation
tion basées sur tion pour dans des
les véhicules atteindre les réseaux
connectés. objectifs de à grande
gestion du échelle.
trafic.
Ning et al. Modèle de ges- Modèle Fixed Réduire le L’impact

[34] tion du trafic exact pour temps de de la mo-
dans un en- le ’offload- réponse. bilité des
vironnement ing’. véhicules
vehiculaire- n’est pas
edge. étudié.
Moubayed Placement de Modèle de Fixed Améliorer L’impact

et al. [35] service V2X placement le délai et de la mo-
dans un en- de ser- l’utilisation bilité des
vironnement vice V2X des véhicules
edge-mobile. optimal. ressources. n’est pas
étudié.
12
Shaer et al. Placement Modèle op- Fixed Minimiser le La métrique
[36] de service timal pour délai. d’utilisation
V2X dans le place- des
un environment des ressources
nement edge services. n’est pas
computing. étudiée.
Tang et al. Système de Modèle Fixed Utilisation L’utilisation

[37] conduite de ’Task efficace des des
autonome offloading’. ressources ressources et
en temps et consom- les métriques
réel dans mation de QoS ne
l’environnement d’énergie. sont pas
edge comput- étudiées.
ing.
Hameed et Approche de Modèle op- Mobile Améliorer Un seul

al. [38] ’clustering’ timal pour la qualité modèle de
dynamique l’utilisation de service mobilité est
dans un en- du réseau. et la con- utilisé qui
vironnement sommation représente
’vehicular edge d’énergie. juste un
computing’. cas de
l’environnem-
ent réel.
Table 2 Résumé des recherches existant sur le vehicular edge computing.
Les convergences actuelles entre la technologie edge computing et les réseaux vehiculaire
devraient prendre en compte les métriques de sécurité et les réactions en temps réel dans les
situations critiques. Par conséquent, malgré les travaux existants dans les communications
véhiculaires assistées par le Edge computing, les applications de nouvelle génération telles
13
que la sécurité routière et la conduite autonome nécessitent un déploiement efficace pour
atteindre une latence ultra-faible dans les réseaux de véhicules denses.
14
3 Nos Solutions Proposées Pour Les
Réseaux Edge Computing
La procédure d’optimisation du réseau reste une tâche importante qui doit être soigneusement
étudiée en raison des évolutions rapides du réseau et des demandes des utilisateurs. Dans
cette partie, nous présentons nos différents solutions pour le service offloading et vehicular
edge computing.
3.1 Service Offloading dans Edge Computing

Avec l’évolution rapide des générations de réseaux et les exigences des applications, le service
offloading est considéré comme l’une des tâches réseau les plus importantes qui permettent
de ’offload’ les services gourmands en ressources vers un centre de calcul optimal dans les
serveurs edge et le cloud central pour le traitement. Cette tâche complexe nécessite une
communication efficace, des stratégies d’optimisation et des technologies de virtualisation
pour gérer différents services et fournir une qualité de service efficace aux utilisateurs.
Nous avons proposé des modèles optimaux [7, 6, 1, 4] d’orchestration des tâches, des
VNF et des SFC dans le edge de réseau (Optimal Virtual Edge nodes Placement Algorithm
(VEnPA), Service Offloading in Virtual Mobile Edge Computing (SO-VMEC), Optimal SFC
Placement in IoT-VMEC (OSPV)) en tenant compte des contraintes de ressources de calcul.
Pour gérer les réseaux denses, nous avons proposé des algorithmes à grande échelle
[7, 6, 1, 4] qui prennent en charge une énorme quantité d’appareils, de données et de
demandes d’utilisateurs. Les algorithmes d’intelligence artificielle d’apprentissage par
renforcement et d’apprentissage par renforcement profond sont utilisés pour le placement des
tâches des utilisateurs et les tranches VNF (Q-learning et Deep Q-learning (DQN)), tandis
15
qu’un algorithme Efficient SFC Placement (ESPV) basé sur le modèle Bin Packing [39] est
implémenté pour le placement des SFC dans les serveurs edge virtuel.
Dans ce qui suit, nous présentons nos solutions pour le vehicular edge computing.

Avec le développement récent des réseaux urbains, les véhicules intelligents sont considérés
comme des appareils mobiles équipés de différents capteurs embarqués. De nos jours, les
véhicules commencent à avoir la capacité de communiquer entre eux (véhicule à véhicule) et
avec l’infrastructure urbaine (véhicule à infrastructure), de collecter des données et d’effectuer
des calculs et des traitements dans le véhicule. Aujourd’hui, généralement, les applications
de calcul intensif sont exécutées sur le cloud. Cependant, les applications sensibles aux
retards peuvent souffrir de la réponse longue durée et peuvent souffrir d’une mauvaise qualité
d’expérience.
Les nouvelles générations de réseaux véhiculaires permettent d’utiliser les véhicules
comme serveur edge dans le contexte des systèmes vehicular edge computing (VEC) pour
le traitement des données, etc. Il permet de fournir des informations à d’autres appareils
connectés, y compris les appareils des utilisateurs, d’autres véhicules, etc.
Nous avons proposé des modèles optimaux [8, 3, 5] pour vehicular edge computing, y
compris une couverture maximale (MVEC) pour les applications de streaming vidéo en ligne
en utilisant des taxis comme des serveurs edge mobiles à l’intérieur de la ville, le offloading
optimal des VNF de la conduite autonome (OVEAP) dans les serveurs edge, et le modèle
UAV-edge (OFMU) pour la diffusion de video en ligne en utilisant des UAV autonomes
comme serveurs edge mobile.
De plus, pour couvrir les fortes exigences des réseaux denses. nous avons proposé des
algorithmes à grande échelle [8, 3, 5]. Pour le premier cas d’utilisation, afin de maximiser
la couverture des véhicules clients à l’aide du serveur edge mobile (Taxis), nous proposons
un algorithme heuristique qui sélectionne uniquement les taxis qui partagent le(s) même(s)
segment(s) de chemin avec le véhicule client afin de réduire la complexité de l’algorithme.
Pour le deuxième cas d’utilisation, nous proposons un algorithme heuristique pour FMU
(AFMU) basé sur le modèle Weighted Set Cover (WSC) pour augmenter la disponibilité du
16
drone pendant le service. Enfin, pour le dernier cas d’utilisation, nous proposons Efficient
Edge Autopilot Placement (DVEAP) basé sur l’apprentissage par renforcement profond pour
placer efficacement les VNF de conduite autonome dans le serveur edge.
17
4 Conclusion et perspectives
Cette thèse présente edge computing pour l’Internet des objets. La première partie présente
l’introduction générale. La deuxieme partie presente les technologies étudiées (l’Internet des
objets, le cloud computing, l’edge computing) et les travaux proposés par la communauté de
recherche pour deux grands domaines de edge computing, service offloading et vehicular
edge computing.
La troisième partie présente nos solutions proposé (optimal et grand echelle) pour service
offloading et vehicular edge computing.
les modèles optimaux proposés pour optimiser l’utilisation de edge pour différentes
métriques telles que le temps et l’utilisation des ressources. Nous avons proposé des modèles
optimaux d’orchestration des tâches, des VNF et des SFC dans le edge (VEnPA, SO-VMEC,
OSPV). De plus, nous avons proposé d’autres modèles optimaux pour vehicular edge com-
puting, y compris la couverture maximale (MVEC), la conduite autonome (OVEAP), et le
modèle UAV-edge (OFMU) utilisant des UAV autonomes comme serveurs edge mobile. Les
résultats numériques montrent que l’algorithme proposé donne de bons résultats en termes
d’utilisation des ressources dans le edge pour les réseaux à petite échelle.
De plus, pour gérer les réseaux denses, nous avons proposé des algorithmes à grande
échelle qui prennent en charge une énorme quantité d’appareils, de données et de demandes
d’utilisateurs pour service offloading et vehicular edge computing. Des algorithmes heuris-
tiques sont proposés pour l’orchestration SFC (ESPV), maximiser la couverture des véhicules
(AFMU) par les serveurs edge mobile (Taxis et UAV). De plus, les algorithmes d’intelligence
artificielle (Q-learning et Deep Q-learning (DQN)) sont utilisés pour le placement des VNF
de conduite autonome (DVEAP) et la navigation autonome des drones. Les résultats obtenus
prouvent l’efficacité et la faisabilité des solutions proposées par rapport aux algorithmes
optimaux.
19
Malgré les solutions existantes décrites ci-dessus, plusieurs préoccupations et défis
doivent être étudiés afin d’intégrer pleinement edge computing aux applications IoT. Les
défis suivants examinent les principaux enjeux, ainsi que les nouveaux concepts que la
communauté de la recherche devrait prendre au sérieux dans les travaux futurs.
• La disponibilité des services doit être assurée pour la prochaine génération d’applications
de edge computing grâce à l’utilisation de divers mécanismes tels que la surveillance, la
prédiction et les sauvegardes du système.
• La mobilité est un facteur important dans le système Edge-IoT, car la plupart des
appareils connectés, tels que les véhicules, et les drones, sont mobiles, ce qui entraîne de
fréquentes dégradations de service et des pannes de liaison entre les serveurs et les appareils,
ce qui diminue la qualité de service de système Edge-IoT. Pour cela, des modèles efficaces
et développés pour la détection des obstacles et la prévision du trafic, doivent être proposés
pour surmonter le problème de mobilité.
• L’architecture edge computing est composée de plusieurs systèmes distribués, la
consommation d’énergie devrait être considérable, ce qui augmentera les dépenses. En
conséquence, de nombreux efforts doivent être faits pour résoudre ce problème en dévelop-
pant des modèles énergétiques efficaces dans les systèmes edge, tels que l’optimisation des
ressources, le recours à une énergie plus respectueuse de l’environnement, etc.
20
References
[1] Mohammed Laroui, Hatem Ibn-Khedher, Moussa Ali Cherif, Hassine Moungla, Hos-
nications Technologies, page e4211, 2021.
[2] Mohammed Laroui, Boubakr Nour, Hassine Moungla, Moussa A Cherif, Hossam Afifi,
and Mohsen Guizani. Edge and fog computing for IoT: A survey on current research
activities & future directions. Computer Communications, 180:210–231, 2021.
[3] Mohammed Laroui, Hatem Ibn-Khedher, Hassine Moungla, and Hossam Afifi. Au-
Communications Conference (GLOBECOM), pages 1–6, 2021.
[4] Mohammed Laroui, Hatem Ibn-Khedher, Hassine Moungla, and Hossam Afifi. Artificial
Intelligence Approach for Service Function Chains Orchestration at The Network Edge.
In IEEE International Conference on Communications (ICC), pages 1–6, 2021.
[5] Hatem Ibn-Khedher, Mohammed Laroui, Mouna Ben Mabrouk, Hassine Moungla,
Communications and Mobile Computing (IWCMC), pages 254–259, 2021.
[6] Mohammed Laroui, Hatem Ibn Khedher, Hassine Moungla, Hossam Afifi, and Ahmed E
Kamal. Virtual Mobile Edge Computing based on IoT devices resources in smart cities.
In IEEE International Conference on Communications (ICC), pages 1–6, 2020.
[7] Mohammed Laroui, Moussa Ali Cherif, Hatem Ibn Khedher, Hassine Moungla, and
Hossam Afifi. Scalable and Cost Efficient Resource Allocation Algorithms Using Deep
Reinforcement Learning. In IEEE International Wireless Communications and Mobile
Computing (IWCMC), pages 946–951, 2020.
21
References
[8] Mohammed Laroui, Boubakr Nour, Hassine Moungla, Hossam Afifi, and Moussa Ali
Cherif. Mobile vehicular edge computing architecture using rideshare taxis as a
mobile edge server. In IEEE 17th Annual Consumer Communications & Networking
Conference (CCNC), pages 1–2, 2020.
[9] Mohammed Laroui, Aicha Dridi, Hossam Afifi, Hassine Moungla, Michel Marot,
and Moussa Ali Cherif. Energy management for electric vehicles in smart cities: a
deep learning approach. In IEEE International Wireless Communications & Mobile
Computing Conference (IWCMC), pages 2080–2085, 2019.
[10] Mohammed Laroui, Akrem Sellami, Boubakr Nour, Hassine Moungla, Hossam Afifi,
and Sofiane B Hacene. Driving path stability in VANETs. In IEEE Global Communica-
tions Conference (GLOBECOM), pages 1–6, 2018.
[11] Prince Kwame Senyo, Erasmus Addae, and Richard Boateng. Cloud computing
research: A review of research themes, frameworks, methods and future research
directions. International Journal of Information Management (IJIM), 38(1):128–139,
2018.
[12] Peter Mell, Tim Grance, et al. The NIST definition of cloud computing. 2011.
[13] Dimitrios Zissis and Dimitrios Lekkas. Addressing cloud computing security issues.
Future generation computer systems (FGCS), 28(3):583–592, 2012.
[14] Sean Marston, Zhi Li, Subhajyoti Bandyopadhyay, Juheng Zhang, and Anand Ghal-
sasi. Cloud computing—The business perspective. Decision support systems (DSS),
51(1):176–189, 2011.
[15] Gabriel Mateescu, Wolfgang Gentzsch, and Calvin J Ribbens. Hybrid comput-
ing—where HPC meets grid and cloud computing. Future generation computer systems
(FGCS), 27(5):440–453, 2011.
[16] Emiliano Miluzzo, Ramón Cáceres, and Yih-Farn Chen. Vision: mClouds-computing
on clouds of mobile devices. pages 9–14. Proceedings of the third ACM workshop on
Mobile cloud computing and services, 2012.
[17] Afnan Fahim, Abderrahmen Mtibaa, and Khaled A Harras. Making the case for
computational offloading in mobile device clouds. pages 203–205. Proceedings of the
19th annual international conference on Mobile computing & networking (ACM), 2013.
22
References
[18] Duc Van Le and Chen-Khong Tham. An optimization-based approach to offload-

ing in ad-hoc mobile clouds. pages 1–6. IEEE Global Communications Conference
(Globecom), 2017.
[19] Ragib Hasan, Mahmud Hossain, and Rasib Khan. Aura: An incentive-driven ad-hoc
IoT cloud framework for proximal mobile computation offloading. Future Generation
[20] Duc Van Le and Chen-Khong Tham. A deep reinforcement learning based offloading
scheme in ad-hoc mobile clouds. pages 760–765. IEEE Conference on Computer
Communications Workshops (INFOCOM WKSHPS), 2018.
[21] Duc Van Le and Chen-Khong Tham. Quality of service aware computation offloading in
an ad-hoc mobile cloud. IEEE Transactions on Vehicular Technology, 67(9):8890–8904,
2018.
[22] Md Golam Rabiul Alam, Mohammad Mehedi Hassan, Md ZIa Uddin, Ahmad Almo-
gren, and Giancarlo Fortino. Autonomic computation offloading in mobile edge for IoT
applications. Future Generation Computer Systems, 90:149–157, 2019.
[23] Ke Zhang, Yongxu Zhu, Supeng Leng, Yejun He, Sabita Maharjan, and Yan Zhang.
Deep Learning Empowered Task Offloading for Mobile Edge Computing in Urban
Informatics. IEEE Internet of Things Journal, 2019.
[24] Zongyao Wang, Xue Feng, Hongbo Zhu, and Changping Liu. Optimal data offloading
via an ADMM algorithm in mobile ad hoc cloud with malicious resource providers.
Computer Communications, 2020.
[25] Xueling Lin, Jingjie Jiang, Calvin Hong Yi Li, Bo Li, and Baochun Li. Circa: collabora-
tive code offloading among multiple mobile devices. Wireless Networks, 26(2):823–841,
2020.
[26] Sajeeb Saha, Md Ahsan Habib, Tamal Adhikary, Md Abdur Razzaque, Md Mustafizur
Rahman, Meteb Altaf, and Mohammad Mehedi Hassan. Quality-of-Experience-Aware
Incentive Mechanism for Workers in Mobile Device Cloud. IEEE Access, 9:95162–
95179, 2021.
[27] Palash Roy, Sujan Sarker, Md Abdur Razzaque, Md Mamun-or Rashid, Mohm-
mad Mehedi Hassan, and Giancarlo Fortino. Distributed task allocation in Mobile
Device Cloud exploiting federated learning and subjective logic. Journal of Systems
Architecture, 113:101972, 2021.
23
References
[28] Oanh Tran Thi Kim, Nguyen Dang Tri, Nguyen H Tran, Choong Seon Hong, et al. A
shared parking model in vehicular network using fog and cloud environment. In IEEE
17th Asia-Pacific Network Operations and Management Symposium (APNOMS), pages
321–326, 2015.
[29] Ke Zhang, Yuming Mao, Supeng Leng, Sabita Maharjan, and Yan Zhang. Optimal delay
constrained offloading for vehicular edge computing networks. In IEEE International
Conference on Communications (ICC), pages 1–6, 2017.
[30] Chao Zhu, Giancarlo Pastor, Yu Xiao, Yong Li, and Antti Ylae-Jaeaeski. Fog following
me: Latency and quality balanced task allocation in vehicular fog computing. In
IEEE International Conference on Sensing, Communication, and Networking (SECON),
pages 1–9, 2018.
[31] Lee Gillam, Konstantinos Katsaros, Mehrdad Dianati, and Alexandres Mouzakitis.
Exploring edges for connected and autonomous driving. In IEEE Conference on
Computer Communications Workshops (INFOCOM WKSHPS), pages 148–153, 2018.
[32] Pedro Henrique Cruz Caminha, Felipe Ferreira da Silva, Roberto Gonçalves Pacheco,
Rodrigo de Souza Couto, Pedro Braconnot Velloso, Miguel Elias Mitre Campista, and
Luís Henrique MK Maciel Kosmalski Costa. Sensingbus: Using bus lines and fog
computing for smart sensing the city. IEEE Cloud Computing, 5(5):58–69, 2018.
[33] Arnav Thakur and Reza Malekian. Fog computing for detecting vehicular congestion,
an internet of vehicles based approach: A review. IEEE Intelligent Transportation
Systems Magazine, 11(2):8–16, 2019.
[34] Zhaolong Ning, Jun Huang, and Xiaojie Wang. Vehicular fog computing: Enabling real-
time traffic management for smart cities. IEEE Wireless Communications, 26(1):87–93,
2019.
[35] Abdallah Moubayed, Abdallah Shami, Parisa Heidari, Adel Larabi, and Richard Brun-
ner. Edge-enabled V2X service placement for intelligent transportation systems. IEEE
Transactions on Mobile Computing, 2020.
[36] Ibrahim Shaer, Anwar Haque, and Abdallah Shami. Multi-Component V2X Applica-
tions Placement in Edge Computing Environment. In IEEE International Conference
on Communications (ICC), pages 1–6, 2020.
24
References
[37] Jie Tang, Shaoshan Liu, Liangkai Liu, Bo Yu, and Weisong Shi. Lopecs: A low-
power edge computing system for real-time autonomous driving services. IEEE Access,
8:30467–30479, 2020.
[38] Ahmad Raza Hameed, Saif ul Islam, Ishfaq Ahmad, and Kashif Munir. Energy-and
performance-aware load-balancing in vehicular fog computing. Sustainable Computing:
Informatics and Systems, 30:100454, 2021.
[39] Michael R Garey and David S Johnson. Approximation algorithms for bin packing
problems: A survey. In Analysis and design of algorithms in combinatorial optimization,
pages 147–172. Springer, 1981.
25

Distributed Edge Computing - IoT - PHD - Thesis

Uploaded by

Copyright:

Available Formats

Distributed Edge Computing - IoT - PHD - Thesis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Distributed Edge Computing - IoT - PHD - Thesis

Uploaded by

Copyright:

Available Formats

Distributed edge computing for enhanced IoT devices

and new generation network eﬀiciency

To cite this version:

HAL Id: tel-04216998

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

Distributed Edge Computing For

Distributed Edge Computing For Enhanced IoT

Par Mohammed Laroui

Thèse de doctorat en Informatique (Réseaux)

Dirigée par Hassine Moungla

Présentée et soutenue publiquement le 21 Juillet 2022

Devant un jury composé de :

Hind Castel, Professeur, Telecom SudParis, Rapporteur

Ken Chen, Professeur, Sorbonne Université, Rapporteur

Hossam Afifi, Professeur, Telecom SudParis, Examinateur

Badr Benmammar, Professeur, Université de Tlemcen, Examinateur

Nassim Dennouni, Docteur, Université de Chlef, Examinateur

I SINCERELY THANK MY PARENTS , MY BROTHER , MY SISTERS AND

I WOULD LIKE TO THANK THE RESEARCHERS WHO HAVE CONTRIBUTED

F IRST, I WOULD LIKE TO THANK MY SUPERVISORS , THEIR EXPERTISE

I N ADDITION , I WOULD LIKE TO THANK THE JURY MEMBERS , REVIEW-

F INALY, I WOULD LIKE TO THANK ALL MY FRIENDS AND PERSONS

3. Hatem Ibn-Khedher, Mohammed Laroui, Mouna Ben Mabrouk, Hassine Moungla,

List of figures xiii

2 Edge Computing For IoT: Overview & Related Works 5

3 Optimal Solutions For Edge Computing Networks 47

3.1.3 Use Case 3: Service Offloading in Virtual Mobile Edge Computing

4 Large Scale Solutions For Edge Computing Networks 99

5 Conclusion and Perspectives 115

2.1 IoT Application Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3.1 VNF Slices Placement in Edge Computing Over 5G Network [7]. . . . . . 48

4.1 RL application to VNF placement in Edge Computing [7]. . . . . . . . . . 100

2.1 CC & MCC vs EC & MEC . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4.1 Servers Configuration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

Edge Computing For IoT: Overview &

2.1 Edge Computing for IoT: Overview

2.1.1 Internet of Things (IoT)

In agriculture, the IoT is revolutionizing the world of agriculture on several levels,

Security & Surveillance Smart Home

Industrial Automation Smart Healthcare

Fig. 2.1 IoT Application Domains

2.1.2 Cloud Computing (CC)

Application Layer Smart Smart Smart

Fig. 2.2 3-Tiers IoT Architecture.

Cloud Computing Deployment Models

Cloud Computing Service Models

CC for Internet of Things

Private Cloud Infrastructure as Platform as a Software as a

Storage Storage Storage Storage

Networking Networking Networking Networking

Fig. 2.3 Categories of Cloud Computing Services.

2.1.3 Mobile Cloud Computing (MCC)

Satellite Access Point

Fig. 2.4 Cloud Computing Architecture for IoT.

Mobile Network Services

Agent Cloud Computing

Mobile Network B Data Center

Mobile Network Services

Mobile Users Network Operators