IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO.

16, 15 AUGUST 2023 14307

Management and Orchestration of Edge Computing

for IoT: A Comprehensive Survey
Yao Chiang, Student Member, IEEE, Yi Zhang , Member, IEEE, Hao Luo , Tse-Yu Chen , Guan-Hao Chen ,
Huan-Ting Chen, Yan-Jhu Wang, Hung-Yu Wei , Senior Member, IEEE, and Chun-Ting Chou, Member, IEEE

Abstract—With the development of telecommunication tech- the emerging various applications and use cases in the 5G
nologies and the proliferation of network applications in the network, such as augmented reality (AR), interactive games,
past decades, the traditional cloud network architecture becomes and Internet of Things (IoT), the network services’ require-
unable to accommodate such demands due to the heavy burden
on the backhaul links and long latency. Therefore, edge com- ments for Quality of Service (QoS) have become more and
puting, which brings network functions close to end-users by more strict on latency and throughput. However, both the
providing caching, computing and communication resources at large latency incurred in the current cloud-based network
network edges, turns into a promising paradigm. Benefit from its architecture and the insufficient bandwidth capacity of back-
nature, edge computing enables emerging scenarios and use cases, haul networks may fail to fulfill the requirements of these
such as augmented reality (AR) and Internet of Things (IowT).
However, it also creates complexities to efficiently orchestrate innovative applications [1].
heterogeneous services and manage distributed resources in the Facing this, edge computing has recently become a promis-
edge network. In this survey, we make a comprehensive review of ing paradigm. Specifically, edge computing aims to offer cloud
the research efforts on service orchestration and resource man- computing capabilities physical proximity to the end-users by
agement for edge computing. We first give an overview of edge providing computing, caching, and communication resources
computing, including architectures, advantages, enabling tech-
nologies and standardization. Next, a comprehensive survey of at the network edge. In this way, the experienced service delay
state-of-the-art techniques in the management and orchestration and the congestion at the backhaul network can be significantly
of edge computing is presented. Subsequently, the state-of-the- reduced. Furthermore, combining the development of enabling
art research on the infrastructure of edge computing is discussed technologies, such as software-defined network (SDN) and
in various aspects. Finally, open research challenges and future network function virtualization (NFV), edge computing can
directions are presented as well.
offer various advantages for the network ecosystem from the
Index Terms—Computation offloading, content caching, edge aspects of scalability and cost reduction.
computing, resource management, virtual network embedding Meanwhile, as the growing number of IoT applications take
(VNE).
a more and more important part in our daily life, they also
bring more demanding requirements, such as low latency,
I. I NTRODUCTION security, traffic optimization, and location awareness [2].
Various applications, such as video streaming and AR are also
ITH the advancement in telecommunication technolo-
W gies, wireless mobile networks have evolved rapidly
in the last two decades. Meanwhile, mobile data traffic is
widely used in the IoT scenarios. By shifting the computing
capacity from the cloud data centers to the network edge near
the end devices, edge computing is well positioned to lever-
expected to undergo tremendous growth according to the latest
age these mentioned advantages and to provide cost-effective
Cisco Visual Networking Index Forecast. What is more, with
solutions. Not only does edge computing address the individ-
Manuscript received 7 August 2022; revised 29 December 2022; accepted ual issues of the traditional applications like video streaming
3 February 2023. Date of publication 16 February 2023; date of current but it can also be used to build a high-integrity system for the
version 8 August 2023. This work was supported by the National Science
and Technology Council (NSTC) of Taiwan under Grant 111-2221-E-002- novel use cases of smart factories or campuses [1]. These pos-
097-MY3, Grant 111-2622-8-002-030-, and Grant 109-2221-E-002-148-MY2. sibilities have contributed to the growing importance of edge
(Corresponding author: Hung-Yu Wei.) computing and have catalyzed many related studies.
Yao Chiang, Hao Luo, Tse-Yu Chen, Guan-Hao Chen, Huan-Ting Chen,
and Hung-Yu Wei are with the Department of Electrical Engineering, National To provide useful insights into the current status and
Taiwan University, Taipei 10617, Taiwan (e-mail: [email protected]; uncover potential research directions of edge computing, there
[email protected]; [email protected]; [email protected]; have been a number of survey papers over the last few years
[email protected]; [email protected]).
Yi Zhang is with the Department of Information and Communication addressing different aspects of the edge computing paradigm.
Engineering, Xiamen University, Xiamen 361005, China (e-mail: yizhang@ Specifically, both fundamental key enabling technologies, the
xmu.edu.tw). reference architecture, and deployment scenarios of the multi-
Yan-Jhu Wang and Chun-Ting Chou are with the Graduate Institute
of Communication Engineering, National Taiwan University, Taipei 10617, access edge computing (MEC) paradigm are presented in [1].
Taiwan (e-mail: [email protected]; [email protected]). The definition, advantages, architectures, and application areas
This article has supplementary material provided by the of MEC are presented in [3]. As a concept prompt from the
authors and color versions of one or more figures available at
https://doi.org/10.1109/JIOT.2023.3245611. requirements of IoT, an overview of fog computing (FC),
Digital Object Identifier 10.1109/JIOT.2023.3245611 including the definition and research trends, was given in [4].
2327-4662 
c 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14308 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

TABLE I
S UMMARY OF E XISTING S URVEYS ON M ULTIACCESS E DGE C OMPUTING

Finally, inclusive taxonomy for architectural, algorithmic, and 2) We present a survey of the state-of-the-art manage-
technologic aspects of FC is also provided in [5]. ment and orchestration techniques for edge computing,
On the other hand, some survey papers also addressed the including task offloading, resource allocation, content
state-of-the-art techniques involved in edge computing. Task caching, and VNE. Instead of just listing the related
offloading in edge computing, which helps the execution of research, we compare and summarize these related work
computation-intensive applications of mobile devices with the with a table.
computing resource of edge nodes, has been discussed in [6] in 3) We have surveyed the infrastructure of edge computing,
detail. The state-of-the-art MEC research with a focus on joint including their deployment scenarios and development
radio and computation resource management were summa- tools, to provide insights for achieving an efficient
rized in [7]. The survey of edge computing from the aspect of management and orchestration framework that is able to
caching, computing and communication was presented in [8]. work well with the common applications and use cases.
Furthermore, Mehrabi et al. [9] investigated mechanisms that 4) The open issues and future challenges concerning the
jointly utilize the resources of the end devices and the installed management and orchestration of edge computing are
MEC to provide services to end devices. Also, with an aim identified. We also provide a summary of the state-
to mitigate the escalation in resource congestion due to mil- of-the-art research works and describe potential future
lions of IoT devices, the edge computing for IoTs have been research directions.
deeply surveyed in [10]. Finally, fundamentals of deep learn- The remainder of this article is organized as shown in
ings (DLs), DLs for edge computing, and future challenges Fig. 1. We divide sections into three parts. The first part
in DL were outlined by Wang et al. [11]. The compari- presents an overview of edge computing. The second part
son between these related surveys are shown in Table I. The is the survey part that starts from Sections III–VI. In this
symbol “O” means that the survey contains the topic in the part, we focus on the infrastructure of edge computing first
corresponding column, while “X” is on the contrary. Our work in Section III, and we divide the main issue into subsections.
contains all the listed topics, including seldom-discussed ones, Based on the infrastructure, the resource management and
such as caching and virtual network embedding (VNE). orchestration issues are discussed in Section IV. In Section V,
However, these surveys focus on either state-of-the-art we describe and summarize the state-of-the-art techniques on
architectures, application use cases, enabling technologies, or management and orchestration in the edge. In Section VI,
reviews of attributes and technical aspects of edge computing. we present the emerging applications and related research.
Owing to the limited resource and fast-changing environment The third part is the navigation part. we discuss the technical
at the network edge, edge computing also creates complexities challenges and shed light on the future research directions of
for the network operators. Therefore, resource management edge computing in Section VII. Finally, this article concludes
and service orchestration cannot be overemphasized when it in Section VIII.
comes to realizing the potentials of edge computing. Different
from the existing surveys, we provide a comprehensive survey
of management and orchestration for edge computing, espe- II. E DGE C OMPUTING : OVERVIEW
cially the survey about network management like the VNE In the past two decades, various computing paradigms have
issue is also included. Specifically, the main contributions have been proposed to address the huge amount of computing and
been modified as follows. storage requirements from heterogeneous devices [15]. We
1) A comprehensive survey on the current development will introduce and analyze the related concepts in this sec-
of edge computing is presented, including a compar- tion. Firstly, the enabling technologies are discussed. Secondly,
ison of different edge architectures among MEC, FC, issues and existing solutions are illustrated. Due to space lim-
and Cloudlets, and the summary of their advantages and itations, we upload some fundamentals of Edge computing as
enabling technologies. supplementary material. It covers the architectures of different

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14309

Sec III. Sec IV. Sec V. Sec VI. Sec VII.

Sec II. Key Research/Techniques Resource Management State-of-the-art Research on Emerging Challenges,
Overview on Infrastructure and Service Orchestration Resource Management and applications and Open Issues, and Future
in the Edge in the Edge Orchestration in the Edge related research Research Directions

A. Architecture A. Management and A. Architecture

A. Enabling Technologies Implementation A. Performance Metric Orchestration of A. Application Type
Enhancement and
and Enhancement Task Offloading Integration
B. Issues and Solutions B. Resource Allocation B. Edge Server
B. Virtual Network in B. Management and Position B. Machine Learning
Edge Computing Orchestration of for Management
S. I. Edge Architectures C. Mobility Management Content Caching C. Cloud and Orchestration

S. II. Advantages of C. Security and

D. Cache Privacy in
Edge Computing Supplementary Material Edge Computing
S. III. Standardization of E. Awareness
Edge Computing Overview part Survey part Navigation part

Fig. 1. Structure and organization of this survey.

edge computing schemes and the advantages of edge comput- particular order. Since network functions are only supported
ing from both users’ and operators’ perspectives. Moreover, by dedicated devices and the placement of those devices are
the standards of edge computing, including the ETSI MEC fixed in a network, it might be inefficient when the demand
standard [28] and the P1935 [29] standard, are also introduced. for services changes over time. NFV separated the network
P1935 is a standardized design for resource and application function from the specific hardware, letting all applications
management in Edge/Fog computing. be able to run on the standardized computation, storage, and
networking nodes [33]. In this way, providing customized
A. Enabling Technologies services according to the needs of specific scenarios can then
To reach the potential of edge computing, resource at be achieved.
network edges should be utilized efficiently and flexibly. 3) Edge Network Using SDN and NFV: Recently, based on
Network virtualization is one of the key enablers for resource the basis of SDN and NFV, the concept of VN has emerged
management in edge environments. It allowed the operator with an aim to abstract the conventional hard-coded hardware
to create and revoke virtual machines dynamically [30]. In networks into programmable functions. VN technology pro-
this section, we will give brief descriptions of SDN and vided a unified platform for an edge environment to integrate
NFV, which are the two fundamental techniques for network and coordinate resources. The types of resources of each edge
virtualization. node may be different. It is difficult to handle these various
1) Software-Defined Network: SDN provides an intelli- resources at the same time for the network operators. In addi-
gent resource allocation platform from a global perspective, tion, not only do the types of resources increase the difficulty
enabling carriers to dynamically deploy services. Thanks to of management, but also the difference in the capacity of each
SDN technology, it decoupled the network control plane from node makes resource allocation harder. Because the resource
the data forwarding plane, resources and information among capacity in edge nodes is rather limited, sometimes it is nec-
each network device can be collected and coordinated, making essary to aggregate their resources to serve applications with
the centralized management and orchestration between edge higher resource demand. With virtualization technology, the
nodes be realized [1], [31], [32]. The infrastructure layer is integration and collaboration of resources can be easier to
composed of network equipment, such as routers, switches, achieve.
which forward packets based on their own rules. The SDN a) VN in edge computing: In the SDN/NFV-based envi-
controllers are in charge of virtualizing the resources, such ronment, all of the existing networks such as the CNs in mobile
as network status, network topology information, statistics communications, any mobile applications, can be virtualized
details collected from each device, and coordinating between into various kinds of VNs. In such a way, traditional Internet
those resources [32]. With the help of SDN controllers, the service provider (ISP) could be decoupled into two indepen-
information is shared among network equipment, increasing dent roles, including service provider (SP) and infrastructure
the traffic forwarding efficiency. Once receiving requests from provider (InP). The SP can be further split into a VN provider
the upper application layer, the SDN controller will inte- (VNP), a VN operator (VNO), and the SP [34], [35]. The VNP
grate the virtual resources and allocate them to each request. integrates the substrate resources provided by InPs and forms a
The application layer can then provide much more innovative virtualized substrate network (SN) topology to the VNO. The
applications for solving real-world network problems. VNO installs and runs the VNs asked by the SP on the virtu-
2) Network Function Virtualization: Apart from SDN alized SN provided by the VNP. The SP then offers services
which increases the network efficiency by providing central- to users through VNs.With this kind of method, the resource
ized management for the network equipment, NFV increased can be allocated more efficiently and effectively. Finally, the
the flexibility of network function deployment and enables the InP gains revenue from the SP, and monitors each VN request
edge network operator to provide customized services to users. (VNR) [34]. In the following sections, we first present one of
In traditional networks, network functions, such as firewalls, the key enablers—VNE for the management and orchestration
network address translators (NATs), load balancers, are imple- in SDN and NFV-based edge environments.
mented by specific hard-coded hardware. Data traffic from b) VNE: The procedure of VNE can be divided into two
the source node to a destination node must pass through a parts: virtual node mapping stage and virtual link mapping
series of equipment providing these network functions in a stage [35]. In the node mapping stage, the VNR nodes will
Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14310 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

be mapped onto substrate nodes with sufficient resources. In on [39]. Details for security and privacy are described more
the link mapping stage, each virtual link between two virtual in Section IV and VII.
nodes will be mapped onto the path with sufficient bandwidth
connecting the corresponding substrate nodes. Most of this III. K EY R ESEARCH /T ECHNIQUES ON
article adopted the shortest path here to minimize bandwidth I NFRASTRUCTURE IN THE E DGE
consumption. After mapping the VNR onto the SN, the cost
With the increasing research on edge computing systems,
can be calculated.
how to achieve an efficient management and orchestration
There are two perspectives to look at the VNE in managing
framework that is able to work well with the common appli-
the edge networks. The first perspective is from the perspective
cations becomes an important issue. The deployment details,
of SP and InP, it is much easier to deploy services and applica-
which greatly depends on the requirements and environment,
tions according to specific objectives [35]. And, since all of the
are multifarious. What is more, in a massive-machine sce-
services share the same physical hardware, the cost of upgrad-
nario like the IoT scenarios, these minor differences may add
ing and maintaining the infrastructure can be reduced [24],
up to a larger gap in experience performance. Therefore, in
[25], [26]. The scalability of the network is flexible [26], [27].
this section, we have divided into two sections to introduce
As the number of users increases, the InP can augment the
the infrastructure of edge computing and the various types of
network by providing more resources, and vice versa. With
resources that exist in edge environments.
the programmable virtual functions, the lifecycle management
In the first part, we introduced the edge computing infras-
of network service can be easily achieved by automatic mon-
tructure, including its deployment scenarios and development
itoring and parameter correction [36]. Another perspective is
tools. The development tools can be categorized into VIM,
from the users’ point of view, first, since all of the network
networking software, and Orchestrator. The popular choices
functions are software, users can subscribe to the services and
of VIM deployment includes OpenStack, Kubernetes, and
applications easily with the open interfaces provided by the
OpenShift. They are able to manage the underlying vir-
SP [37]. And the users’ specific requirements, such as secu-
tual machines or containers, simplifying the control flow
rity, and latency can be quantified and added as constraints
and networking settings. The development of the hyperscaler,
for VNRs. The InP can then be customized to satisfy each of
such as Amazon Web Services (AWS) and Google Cloud
their demands.
Platform, also offers more deployment options and improves
B. Issues and Solutions the scalability of edge computing systems. The networking
software, such as OpenAirInterface and OpenEPC are both
Although Edge computing is a promising scenario in today’s
common options. Based on the technique of NFV and SDN,
network architecture, several issues still exist to be solved.
the software reduced the difficulty of deployment and scaling
In this section, we will show these disadvantages of Edge
significantly. The advantages mentioned above are particularly
computing and list some recent works that aim to solve them.
evident in the IoT scenarios. The summary of Orchestrator
1) Limited Resources: Edge computing has enabled appli-
deployment options is listed in Table II. These options may
cations to be deployed on the platform. However, edge servers
have different system designs and are suitable for different
are usually limited in their physical size and need more com-
underlying infrastructure. The second part introduces the exis-
puting and storage resources. Intelligent and efficient resource
tence of various types of resources in the edge environment
allocation and service lifecycle management are critical fac-
and groups them from the perspective of InP and SP.
tors. Approaches and state-of-art of the management and
Aside from the tools used for the practical implementation,
orchestration works are introduced in Sections IV and V.
there are numerous different deployment scenarios, involving
2) Data Management: For a business or organization to
the type of contributions, the architecture, and the applications.
stay on track, data is king in every manner. Nevertheless, the
We compare and classify these papers based on these differ-
edge computing environment is prone to data loss because the
ences in the following subsections. Below, we go through the
edge servers may misjudge data, and the networks may be
papers focusing on architecture enhancement.
unstable. Wang et al. [38] mentioned that blockchain technol-
ogy ensures the stability and dependability of data caching and
transactions. A. Architecture Implementation and Enhancement
3) Security Concern: As Edge computing enables a large Architecture enhancement can be considered as the propo-
amount of data processing at the network edge, the secu- sition of a new or improved architecture of an edge computing
rity and privacy issues cannot be overemphasized. For some system, which means that it tends to focus on the overall
sensitive applications, users need to provide their private system structure design rather than the specific application.
information to the edge server, resulting in the risk of hack- Most of them gave a novel design of the edge computing
ing and cybercriminals. Moreover, since there are different system which does not only match the ETSI MEC standard
kinds of applications and services on the edge platform, each but also has advantages over the regular designs. In this sec-
of them requires different security levels. At this point, it is tion, we will go through the papers in each category, describe
essential to provide a suitable security mechanism for dif- their system design, and explain their differences.
ferent applications. Solutions exist to enhance security, such Cloud SPs (CSPs) have provided several orchestration
as countermeasures for malicious hardware/software injection, services on their cloud data centers, and they can be applied
formulating reliable routing protocols, authorization, and so to the edge computing management scenario. Kubernetes [60],

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14311

TABLE II
Q UALITATIVE C OMPARISON OF D IFFERENT O RCHESTRATOR D EPLOYMENT O PTIONS

TABLE III
C OMPARISON OF THE L ITERATURE ON H IGH -L EVEL A RCHITECTURE

[47]
[48]
[49]
[52]
[50]
[51]
[53]
[54]
[55]
[56]
[57]
[58]
[59]

first developed by Google in 2014, is responsible for manag- The general architecture is composed of the most basic
ing containers in cluster-based Platform-as-a-Service (PaaS) components, that is, the edge computing system. In this case,
systems [61]. Amazon Elastic Container Service [62] allows the location of the edge computing system is just generally
developers to run and scale containers under AWS with- described in an ambiguous approach. In [49], the edge com-
out worrying about underlying infrastructure such as the puting system is not connected to the radio network but only
configuration of servers, networks and storage. Amazon one network gateway. The proposed architecture is composed
CloudFront [63] acts as a Web service for content deliv- of a firewall, an SDN controller, an orchestration entity, and an
ery. It provides routing of user requests to the nearest application server. The SDN controller manages the devices
edge location so it can optimize the performance with other based on the OpenFlow protocols, as well as the orchestra-
services in AWS. However, these services are usually resource- tion entity can be implemented based on Open MANO and
intensive and, thus, not all suitable for the edge scenario. RIFT.io.
To address the condition, several types of research have In the MEC architecture, the edge computing system are
been done to adjust the original architecture. For example, usually connected to the RAN, or more clearly, the BS [50],
Jeffery et al. [64] showed that etcd was a bottleneck in [51], [52]. Another MEC platform for the 5G scenario named
Kubernetes and rearchitected Kubernetes for edge environ- ADRENALINE was proposed by Muñoz et al. in [51]. In their
ments with increased performance, availability and scalability. architecture, the orchestration system is composed of a cloud
Kaur et al. [65] proposed KEIDS, a scalable and comprehen- orchestrator, a transport SDN orchestrator, an NFV orchestra-
sive controller for Kubernetes, to solve an interference and tor (NFVO), a VNF manager (VNFM), and a Global Service
energy minimization problem in edge-cloud environments. Orchestrator. Hu et al. designed image recognition algorithms
Aside from the regular designs that are aligned with the for both the camera sensor and MEC server to achieve high
common architecture, such as ETSI MEC architecture and Fog accuracy and low latency in [52]. The authors applied the
architecture mentioned above, there are also some papers that action-value methods to train the system by jointly optimizing
either introduced novel system designs or changed the loca- offloading decision and image compression parameters.
tion of edge servers. The comparison of these papers from the FC architecture is usually used in IoT use cases. Hou et al.
aspect of high-level architecture, the location of edge servers, investigated a MEC platform in the Industry IoT in [54].
and edge system architecture are shown in Table III. The architecture is composed of a cloud computing layer, an
1) High-Level Architecture: The high-level architecture IoT-MEC layer, and a device layer. With the virtualization
can be divided into five classes: general, Cloudlet, ETSI MEC, technologies, the resources can be divided into the resource
Fog, and satellite. virtualization layer and the resource management layer. A new
Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14312 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

architecture with MEC assistance for the C-V2X system is

proposed by Wang et al. in [56]. In their design, the system
architecture includes four layers: central cloud layer, MEC
layer, roadside equipment layer, and user equipment layer,
with the CN between the central cloud layer and MEC layer
as well as the access network between the MEC layer and
the roadside equipment layer. Shekhar et al. [55] presented a
dynamic resource management middleware, which meets the
mobility, latency, and power consumption requirements. They
designed a real-time object detection application for the visu-
ally impaired, in which the image classification optimization
Fig. 2. Concept of VN.
problem is solved on the edge/fog side.
Cloudlet architecture is similar to the MEC one. side, are closer to the user equipment and, thus, have lower
Rimal et al. [57] proposed an architecture that provides MEC in latency.
the fiber-wireless network with cloudlet-aware resource man- c) Edge between RAN and CN: Sarrigiannis et al. [53]
agement. The system is designed in two time-division multiple proposed a MEC-based architecture dedicated to the IoT use
access (TDMA) layers to enhance the network performance. case. Their high-level architecture is a variant of the regular
Compared to the other choices mentioned above, satellite one, composed of a Core tier, an Edge tier, a device tier, as
architecture is more special as it takes the satellite links into well as an NVFO and a VIM connected with the above tiers
account. Politis et al. [59] presented a Satellite-enabled 5G via the control plane.
testbed design with Satellite elements connected to the MEC-
d) Edge on CN side: Placing the edge servers on the
enabled Edge nodes and the backhaul between RAN and CN.
CN makes the system more centralized. Reference [57] is an
Cao et al. [58] proposed a space-based fog architecture which
example of it.
includes many satellite nodes with computing and storage
capabilities. These satellites can be classified into two types:
1) the cloud nodes and 2) the fog nodes. The cloud nodes B. Virtual Network in Edge Computing
are responsible for task analysis and resource scheduling. The There are various types of resources in an edge envi-
fog nodes handle the computing tasks received from the cloud ronment, such as computing resources, bandwidth, storage,
nodes. energy, end-to-end service time. We group these resources
2) Location of Edge Servers: Depending on the place that from the perspective of InP and SP. From the viewpoint of
edge servers are deployed, the existing papers can be classified the InP, the resources they concern may be CPU, bandwidth,
into the following categories. energy. From the SP’s perspective, latency might be the most
a) Edge-only: Edge-only means that the system only take important resource since it will strongly affect the users’ QoE.
the MEC system into accounts, which is the simplest case. Here, we categorize some of the VNE papers based on the
Some of the papers provide a design that matches the ETSI perspective of the InP and the SP, and further classify them
MEC standard and maps the proper software to the compo- according to their objectives, solutions, and the discussed VNR
nents [47]. The deployment of a virtualized MEC testbed is formats, as shown in Table IV.
presented in [47] based on open-source software with several 1) Various Roles in Virtual Network: In the SDN/NFV-
different options for each component. Bolivar et al. investi- based environment, all of the existing networks such as the
gated the mapping of each software to the components in CNs in mobile communications, and any mobile applications,
ETSI MEC standard, including OpenStack, Kubernetes, OSM, can be virtualized into various kinds of VNs. In such a way,
ManageIQ, etc. Their architecture has a special advantage of traditional ISP could be decoupled into two independent roles,
the support to both VM-based and container-based systems. including SP and InP. The SP can be further split into a VNP, a
The edge-cloud architecture takes the cloud into account VNO, and the SP. The VNP integrates the substrate resources
additionally. Compared to the edge-only architecture, it needs provided by InPs and forms a virtualized SN topology to the
to consider the usage of cloud computing as well and, thus, is VNO. The VNO installs and runs the VNs asked by the SP on
more complicated. Scotece et al. [48] proposed a distributed the virtualized SN provided by the VNP. The SP then offers
file system with consistency among the mobile devices, edge, services to users through VNs. The overall procedure of man-
and cloud entities. In their architecture, the offloading mid- aging and orchestrating resources in edge with VNE can be
dleware is deployed on mobile devices, edge nodes, and summarized as follows. The SP first creates customized VNs
the cloud. The architecture also performs mobility manage- by virtualizing services and applications, and hands these VNs
ment and the fault-tolerance with the help of the cloud to the VNO in VNR formats. The VNO then embeds these
controller. VNRs on a SN with different VNE algorithms according to the
b) Edge at RAN side: If the RAN and the CN are in objective, such as maximizing revenue and minimizing latency.
the scope of the system, the papers can be classified by the Finally, the InP gains revenue from the SP, and monitors each
location of their edge servers. As we have mentioned above, VNR. For example, when substrate nodes crash, the InP must
the edge servers can be on the RAN side, on the CN side, or remap the VNRs for service continuity. The procedure can be
between the RAN and the CN. The edge servers on the RAN summarized as follows and is shown in Fig. 2.

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14313

TABLE IV
VNE PAPER C LASSIFICATION

[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]
[76]
[74]
[75]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]

2) Goals and Objectives of Optimization: Here, we list service and decrease the users’ QoE. As for batch embedding,
some of the possible objectives that VNE algorithms might since the operator embeds the VNRs arriving in a time win-
pursue, mainly based on the resources they care about, such dow, each VNR may not obtain resources instantly. These are
as CPU and bandwidth resources, energy, and latency. critical and are worth considering when discussing the VNE
a) Minimize CPU and bandwidth cost: In an edge envi- problem especially with ultralow latency requirements in an
ronment, CPU resources are deployed in a distributed manner edge environment.
for handling computing tasks. Take the AR service for example b) Minimize energy consumption: In addition to the CPU
again. Each process, such as compressing, decompressing, and and bandwidth resources, some of the research considered
video rendering, needs computing resources. Besides, band- energy resources [72], [73], [74], [75], [76] which is another
width is necessary for data transmission between the two type of resource important in the edge. Since most of the
processes. In general, most of this article considers the CPU devices in the edge may be smartphones or tablets in our daily
and bandwidth as their main resources, and their objective is life, which are battery limited, how to satisfy the requirements
to minimize the CPU and bandwidth cost in each embedding of each VNR with less energy consumption becomes another
stage to maximize the long-term revenue of the InP [66], [67], issue in the edge environments.
[68], [69], [70], [71], [76]. In order to reduce energy consumption, the most common
Some common methods used for increasing network uti- method is to hibernate or power down some devices without
lization, such as path splitting [70], [85], batch embed- compromising network performance [75]. Take AR service for
ding [86], remapping [70] mechanisms. Haeri et al. [85] example again, the demand for AR service may be different
allowed the SN to split a virtual link over substrate links, from our daily routine. Therefore, if the edge operator can take
and formed the link mapping subproblem as a multicom- the VNR for the AR service off when there is low demand,
modity flow problem. Their simulation results showed that energy can be saved. In [73], a heuristic algorithm is proposed
by path splitting, the network performance can be signifi- to reduce power consumption by placing virtual nodes and
cantly enhanced. Chang et al. [86] considered the scenario links on the same substrate component so that InP can shut
that VNRs arrive simultaneously, so the InP has a global view down as many network resources as possible.
for the entire network, and is able to select VNRs from the c) Minimize latency: Latency is another key requirement,
batch for the optimal embedding. Yu et al. [70] presented especially in the future 5G networks, where many delay-
the concept of virtual link migration. By reoptimizing the sensitive applications will emerge, such as AR, real-time
embedded paths periodically, the utilization of the SN can be cognitive assistance, and facial recognition on a mobile
increased. device [87], [88], [89]. By embedding VNR nodes in edge
The above-mentioned mechanisms can enhance network environments, the low latency requirements can be satisfied
performance undoubtedly. However, they might face some due to the physical proximity of the end-users and the edge
problems, such as service interruption or failure to meet the resources. However, sometimes it is necessary to map VNR
latency constraints when applied to an edge environment. For nodes to multiple edge environments or even in a cloud
example, when it comes to latency-sensitive applications such network due to the limitation of resources in a single edge. In
as AR, reoptimizing the paths may affect the continuity of the this way, the resource requirements can be satisfied but might

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14314 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

sacrifice the latency due to the propagation delays between becomes one of the most important issues to VNE in the
edges or clouds. The problem of how to allocate resources edge.
to each request while satisfying the latency constraints within 4) VNR Format Considerations: Most of the papers con-
multiple domains becomes even harder. sider VNRs in arbitrary topology formats [66], [68], [70], [71],
We list some of the VNE papers considering minimizing [73], [74], [75], [76], [77], [78], [79], [80], [82], [83] since
latency as their objectives [66], [67], [81], [82], [83], [84], all of the networks and applications can be abstracted into
which are worth referencing when applying VNE in edge VNs with network virtualization technology. However, some
networks. works, such as [67], [69], [72], [81], and [84] only take spe-
Chiang et al. [67] optimized the end-to-end service time cific VNR formats such as chain format into account as they
and revenue in a MEC network by proposing a novel embed- believe that chain network VNRs are more practical to the real
ding algorithm named particle swarm optimization (PSO)- world. Take face detection application as an example, decom-
CSNR. The latest processor technology is adopted on the press the photograph first, and then carry out face recognition.
base node of the CPU with asymmetric core frequency. The If the identity is verified, it will proceed to the next step [67].
service times they consider include processing delay, transmis- All of these functions are processed one after another, and
sion delay, propagation delay, and scheduling delay (queuing chain topologies are the best formats to represent them.
delay), which are the four types of delay that affect network The edge network provides a low latency environment
speed [90]. This work focuses its scenarios in the edge for services and applications by shifting resources from the
networks, providing a research direction toward the VNE Internet cloud to the edge devices. In the SDN/NFV-based
problem in edge environments. environments, the management and orchestration of resources
3) Solving Approaches for VNE: Some of the works pro- in edge can be realized with VNE which has attracted much
vide integer linear programming (ILP) formulation to get the attention in recent years. However, most of the works focused
exact solutions [72], [73], [75], [78], [79], [80], [81]. However, on cloud environments where abundant resources are equipped
the VNE problem is known to be NP-hard [91]. Even given with. Unlike the cloud environments, edge servers usually have
node mapping results, optimal allocation of virtual links on limited resources, and often deployed in a distributed way. If
a single substrate path is still NP-hard [92], [93]. Therefore, no edge server is available in the region close to the end-user,
most of the research develops heuristic or metaheuristic algo- the virtual node will be mapped to a more distant substrate
rithms to achieve suboptimal solutions in a reasonable time in node which may incur a higher response time. As a result,
a large-scale network [66], [67], [69], [70], [71], [72], [73], how to assign the physical resources in edge servers economi-
[74], [76], [77], [78], [79], [81], [82], [84]. Fajjari et al. [71] cally becomes much more important. In addition, resources in
propounded an embedding algorithm named VNE-AC based edges should not only be restricted to CPU and bandwidth but
on the ant colony. also other resources such as storage [94]. Therefore, various
In addition, with the rapid development in machine learning types of substrate nodes should be taken into account when
recently, some research starts to apply this technique to solv- considering VNE in edge networks.
ing the VNE problems. Yao et al. [68] proposed a pretrained
reinforcement learning (RL) model to maximize long-term rev- IV. R ESOURCE M ANAGEMENT AND S ERVICE
enue. In this way, the changes in the network resources can O RCHESTRATION IN THE E DGE
be taken into consideration, and better decisions can be made.
The major complexity of edge computing is to support
There are many algorithms for solving VNE problems.
various emerging applications in the dynamic wireless edge
However, most of the existing works considered VNE prob-
network with scarce resources. For example, an edge system
lems in a cloud environment, and their algorithms may not
deployed in a smart factory use case may be responsible for
be suitable when applying in edge environments for the fol-
monitoring and recognition applications, information security
lowing reasons. First, we believe that some of the VNRs in
firewall, production line management, as well as the resource
edge query instant resources. Namely, they might have a strict
allocation among the above services. Hence, it is important
end-to-end service time constraint. Thus, a simple embedding
to investigate how to efficiently manage all the distributed
algorithm that can find an approximate optimal solution in a
resources and orchestrate services in the edge network. In
reasonable time may be a better choice for them rather than
this section, we present the performance evaluation of edge
a method that can get the global optimal but takes too much
computing, resource allocation issues, as well as mobility man-
time. Second, for the scenario that the edge orchestrator itself
agement. Due to the page limitation, we decided to move
is within one of the edge devices, the embedding algorithm
the mobility management part to supplementary material. The
itself will also consume the edge resources, such as comput-
related taxonomy is presented in Fig. 3.
ing resources and energy. A complicated algorithm that may
waste too much time and energy for calculating results does
not benefit the resource utilization in the edge. A. Performance Metrics
Nevertheless, since the resources on edge may not be as To guide long-term planning for research in management
abundant as the cloud, it is critical to utilize them eco- and orchestration of the edge system, it is necessary to prop-
nomically. Therefore, an efficient embedding algorithm is erly evaluate the system performance. Specifically, energy
still undoubtedly necessary. How to balance the comput- consumption, latency, revenue, security and privacy, and fault
ing complexity and the performance of an algorithm then tolerance are the main criteria used in the management and

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14315

monitoring, more and more data is collected and managed

by IoT edge servers [39], [113], [114], [115]. This results in
security threats, such as malicious hardware/software injec-
tion, jamming attacks, Distributed Denial-of-Service (DDoS)
attacks, eavesdropping, etc. Some algorithms and mechanisms
must be developed. However, the development of the security
service also consumes resources on Edge so it may reduce
the user satisfaction. According to this, designers have to
pay attention to the balance of the security level and the
QoS. Solutions to enhance security are included below and
Supplementary in Section VII. Zhang et al. [116] aimed to find a protection
Materials method for commercially confidential information centered on
data security technology. Xu et al. [117] proposed BeCome,
a blockchain-enabled computation offloading method, to guar-
antee data integrity.
4) Privacy: Privacy is an essential metric for the end-
users in Edge computing. It refers to an individual’s
Fig. 3. Taxonomy of issues in management and orchestration of edge
computing.
right to be independent of records and activities to which
they do not consent [115]. Although many algorithms are
designed to increase user satisfaction, the edge environ-
ment should also simultaneously provide users with data
orchestration of the edge system. Thus, in this section, we privacy, location privacy and identity privacy. For exam-
will present the management for the edge computing systems ple, Jia et al. [118] proposed identity-based anonymous
based on these performance metrics. Note that the objectives of authenticated key agreement protocol for the MEC environ-
management can be one or a combination of several of them, ment to assure both user anonymity and untraceability.
and these criteria can be considered as system constraints as 5) Revenue: Aside from the end-users, the SPs also play
well. an important and irreplaceable role in the ecosystems of edge
1) Energy Consumption: Due to the limited battery sup- computing. Their primary concerns include how to reduce the
ply of edge devices, energy consumption would be one of the operational cost and boost their commercial interest. Although
evaluation metric for edge computing [17], [95], [96], [97], deploying more infrastructure resources can enhance QoS, it
[98], [99]. The energy consumption in edge computing can be will increase SP’s operational cost at the same time. Therefore,
caused by both network equipment and UEs. For the network to optimize the social welfare of the whole edge system,
side, the energy is severely consumed during the process of the management strategies should be considered from the
handling services for the end-users. This may include the perspective of SP’s revenue as well [119], [120].
transmission in frontend/backend network and providing the 6) Fault Tolerance: Fault tolerance is the ability of a com-
corresponding resources for execution. As for the user side, the puting system to handle unexpected faults and failures so that
energy consumption is mainly caused by the high-frequency the system can maintain its reliability and QoS. Traditional
communication process with the BSs. fault tolerance approaches include reactive methods, such as
2) Latency: Latency is another important metric for the checkpointing, replication, and task resubmission and proac-
management of edge computing system. Since the latency tive methods, such as self-healing, preemptive migration, and
requirements of next-generation 5G networks is 1 ms, which is load balancing [121]. Proactive methods are designed to
about ten times reduction compared to 4G [100], [101], [102], detect and prevent faults and errors in advance before they
[103], [104], [105], [106], [107], [108], [109], [110], [111], occur. Thus, it can maximize system availability and reduce
[112]. Generally, the latency of a network application com- interruption. On the other hand, reactive methods are mainly
poses of the transmission delay and processing latency. Even developed to decrease the failures after their occurrence and,
though the implementation of the edge computing enables hence, are more suitable to eliminate the effect of infrequent
data to be processed near end devices, reducing up the failures.
data transmission delay for network applications, there are Because an edge server has limited storage and computing
still some challenges needed to be overcome to ensure the resources but the services deployed are usually time-sensitive,
service latency from the viewpoint of management: 1) the the design for efficient and real-time fault tolerance is required.
network transmission latency is heavily depended on the An architecture for multicluster IoT applications was proposed
location of corresponding edge nodes and 2) the workload in [122] and applied with a unified data replication solution.
assignment will also affect the processing latency of the A single management interface for the edge-cloud scenario is
applications. also deployed with Kubernetes to enable automatic data recon-
3) Security: Security consists of information security, figuration. Tuli et al. [123] proposed a proactive mechanism,
cyber-security, forensic security, and network security. With combining generative adversarial network (GAN), to predict
the emergence of applications that bring convenience to preemptive migration decisions in the edge environment and
end-users, such as the smart city, smart factory, and health reduce energy consumption.

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14316 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

B. Resources Allocation vice versa. Therefore, with the integration of the distributed
To optimize the performance of network services in edge computing and storage resources in the edge network, the cor-
networks under different requirements, such as computation responding fronthaul network resource allocation among edge
intensity, delay sensitivity, and ultrareliability, the importance of nodes should be properly optimized to reduce the transmission
resource allocation cannot be overemphasized. However, since overhead and explore the potential of the edge systems.
the manageable resources in wireless communication systems
are evolving from a centralized manner to being distributed V. S TATE - OF - THE -A RT R ESEARCH ON R ESOURCE
in heterogeneous infrastructures, resource allocation problems M ANAGEMENT AND O RCHESTRATION IN THE E DGE
have become increasingly challenging. This section will discuss Although edge computing can bring many advantages to
the resource orchestration and related issues in the edge system mobile networks, it also faces many technical challenges
from three perspectives, inclusive of computational capability, when it comes to the scarce resource at edge nodes and the
storage availability, and networking resource. dynamic nature of the wireless environment. Thus, in order
1) Computing Resource: Here, we study the orchestra- to enhance the network performance of the edge system, a
tion of computing resources distributed in the edge network. lot of efforts from both the academia and the industry have
Generally, the computing resource available in the edge nodes been made to investigate the issues of managing resources
(e.g., Fog node and MEC server) can be utilized to support of the edge system and orchestrating users’ services. Among
the workload required for realizing the novel applications. For those research, task offloading and content caching are the
example, the edge node can perform video transcoding with two critical techniques that had been mostly discussed in
its computation capacity to enhance real-time streaming ser- recent years, which becomes the promising solutions to realize
vice. Also, the projection and rendering process of the VR the appealing applications envisioned in the edge computing
services can be executed in the proximate edge nodes to meet systems. Thus, in this section, we will focus on the state-of-
the latency constraint. However, since the computing capacity art researches on the management and orchestration of both
of each edge node is rather limited, how to efficiently allocate task offloading and content caching in the edge computing
computing capacity to each service under resource constraint paradigm.
will be the main issue [96], [107], [109], [110], [124], [125].
2) Storage Resource: The storage resource in edge nodes A. Management and Orchestration of Task Offloading
can be used to cache some popular content. While most stor- Task offloading technique is an efficient way to break the
age resources are distributed in the BSs, including MBSs, limitations of mobile devices where users can impart their
SBSs, and PBSs, the storage capacity of infrastructures such compute-intensive tasks to proximate edge servers for exe-
as wireless relays and baseband unit (BBU) pools can also cution. Recently, enormous resource-demanding applications,
be exploited to provide better network performance. However, such as AR and interactive games emerge in the edge comput-
since the storage capacity of each edge node is constrained ing paradigm. However, since both the computation capability
compared with the remote cloud server, such heterogeneous and battery life of current mobile devices are rather con-
nature would make the resource allocation problems more strained, the development of these innovative applications may
challenging [126]. be hindered accordingly.
3) Networking Resource: The two main concerns regard- The main objective of the task offloading is to minimize
ing networking resource allocation in the edge system lie in the completion time of the tasks while the predefined latency
the management of wireless resources within RAN and the constraint is satisfied. However, since the computing resource
orchestration of fronthaul network resources. Conventionally, at edge nodes is limited compared with cloud servers, manage-
the allocation of wireless resources, including power, spec- ment of the offloading decision and the corresponding resource
trum, and time, is intertwined with interference management in allocation policy become critical issues. In addition, due to
a complicated way, where the multidomain interference should the users’ mobility and the uncertainty of wireless connection,
be managed properly during wireless multiplexing. However, how to orchestrate users’ services to avoid the interruption of
in the edge computing scenario, where the ultradense deploy- task offloading has become another urgent issue. Thus, to ful-
ment of edge nodes shortens the distances among the users, fill the benefits of task offloading, we should address these
the stronger interference existing between users should also be issues properly. In this section, we will investigate the state-
addressed carefully in the management of wireless resources. of-the-art research efforts on task offloading technique from
What is more, numerous edge users will compete fiercely the viewpoint of management and orchestration. A summary
for the scarce wireless resource in the edge computing sce- of the literature on task offloading is shown in Table V.
nario, making the allocation of wireless resources even more 1) Offloading Scenario: We first focus on the common
challenging [35], [69]. offloading scenarios of edge systems and reviews a set of
On the other hand, the resource allocation problem of recent research efforts for each case. Generally, the scenario
fronthaul network is significantly more complex than the of task offloading can be classified into three types, inclusive
legacy networks due to the increased number of end-users and of single edge, multiedge, and edge-cloud.
the relatively limited transmission capacity. Also, the alloca- In the single edge scenario [95], [124], [127], there is a
tion of computing and storage resources will further affect local edge server providing task offloading service to process
the scheduling of fronthaul communication resources, and users’ computation tasks. Since the computing resource on a

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14317

TABLE V
C OMPARISON OF THE L ITERATURE ON C OMPUTATION O FFLOADING

[127]
[124]
[100]
[95]
[101]
[128]
[17]
[125]
[23]
[129]
[96]
[102]
[103]
[130]
[104]
[119]
[105]
[106]
[107]
[97]
[131]
[98]
[132]
[108]
[20]
[110]
[111]
[99]
[112]
[133]
[120]

single edge server is limited, they focus on optimizing the RL-based offloading scheme to select the offloading rate of
offloading decisions from multiple users and the correspond- IoT devices based on battery level and radio transmission rate.
ing resource allocation problem. Specifically, Yang et al. [127] Furthermore, Shang et al. [98] studied a DL-assisted energy-
proposed a DL-based method to solve the joint optimization efficient computation offloading algorithm for vehicular edge
problem of offloading decision and computation resource allo- computing (VEC) system, which find a nonoptimal solution
cation. Li et al. [124] proposed a deep Q-network-based with low complexity.
computation offloading scheme in the multiuser wireless MEC Finally, in the edge-cloud scenario [99], [109], [110], [111],
system to minimize the long-term system delay and energy aside from the proximate edge server with limited computing
consumption. Hsu et al. [95] proposed a green task dis- resources, the edge users’ computation task can be offloaded to
tribution algorithm for task offloading problem in a MEC a remote cloud server where the computing resources are gen-
scenario with the aim of minimizing the system power erally assumed to be infinite and the round trip time (RTT)
consumption. may be relatively long. Although such a scheme makes the
The multiedge scenario considers computation offloading offloading decision problem more complex and exerts more
among multiple edge servers [98], [107], [131], which can challenges on the service orchestration, the edge-cloud sce-
be regarded as an extension of the single edge scenario. nario is actually more realistic and can further leverage the
However, different from the single edge scenario, the comput- network performance of the whole system.
ing resources are rather distributed in the multiedge scenario, The related works are given as in the following.
which makes it more challenging to orchestrate the optimal Kiran et al. [110] investigated the task offloading and resource
offloading decision and computing resource allocation. Thus, allocation problem in wireless MEC aiming to minimize the
the literature focuses more on the collaboration between delay while saving the battery power of the UE simultaneously.
numerous edge nodes to maximize the network performance Pan et al. [111] studied the problem of dependency-aware task
of the whole edge region, which is the mixed-integer nonlin- offloading decisions in the MEC system. They aimed to min-
ear programming (MINLP) problem in general. For instance, imize the execution time of tasks with constraints on energy
Zhou and Zhang [107] studied a task offloading and resource consumption. A risk-aware cloud-edge computing framework
allocation problem for wireless metropolitan area network in [99] was proposed for the delay-sensitive inspections of
with the awareness of fairness. Min et al. [131] proposed an autonomous manufacturing in a smart factory. Specifically, the

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14318 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

proposed model considers the uncertainty of channel access the task dependency and the device computing capacities.
delay in wireless communication and optimizes the offload- A modified genetic algorithm (GA) was developed to get
ing service deployment with the minimum operation cost and the optimized task offloading result in [133], which greatly
acceptable latency. improved the computing ability and execution efficiency. The
2) Computation Task: After giving a brief description of decomposable intelligence in [120] not only supports decom-
scenarios that task offloading takes place, we then turn to posable inference pipelines in the cloud-edge architecture but
illustrate the essence of the computation task involved in the also enables a pipeline-sharing mechanism to reduce MEC
offloading process. The discussion is divided according to resource usage, where a decomposable pipeline could parti-
three commonly used task models, inclusive of binary, partial, tion the entire DL model into multiple submodels in terms of
and multisubtasks. either layer-level or model-level.
Among the binary task model [104], [105], [119], the tasks 3) Resource Management of Task Offloading: Here, we
can only be offloaded in a binary manner. That is, the tasks focus on discussing the resource allocation problem after a set
can be executed locally on the UEs or be offloaded to the edge of offloading decisions is taken. Specifically, since the task will
servers for execution. Also, we should take the different execu- first be transmitted to the edge server and then be executed by
tion costs (e.g., the required data to transmit and the required the edge server, the main resources involved in the manage-
CPU cycle for execution, and requirements (e.g., latency or ment of task offloading is the radio and computing resources.
energy constraint) of each task into account when managing Generally, the radio resource in the literature is considered
the computing resources of edge nodes so as to maximize to be allocated in either a discrete or continuous way. For
the network performance of the edge systems. There are sev- example, Zhang et al. [120] allocated continuous bandwidth
eral works that had addressed the binary offloading scenario. to the offloading users while Hsu et al. [95], [132] allocated
Fan et al. [104] studied the multitype task offloading problem different number of subcarrier/subchannel to each user. On
and proposed a game-based scheduling scheme. To overcome the other hand, most of the previous works [98], [105], [107],
the limited computing resources of the fog nodes, a novel [110], [129], [133] assumed computing resources to be con-
vehicular FC (VFC) system was investigated in [119] to satisfy tinuous, e.g., the number of CPU cycles. In some cases [23],
the time-sensitive and computation-intensive service requests [124], [127], the computing resources were also considered to
by exploiting the fog capability of parked vehicles. Moreover, be discrete, e.g., the number of CPU core.
Chen et al. [105] studied the computation offloading problem To address the tradeoff between different kinds of resources,
in an ultradense sliced RAN. They modeled the problem as a the allocation of radio and computing resources for the tasks
Markov decision process (MDP) and adopted a double deep that are going to be offloaded to the edge servers should be
Q-network to maximize the system utility. jointly optimized [17], [20], [96], [97], [101], [102], [103],
The tasks in the partial task model can be partially offloaded [124], [125]. In [101], they jointly considered computation
to the edge servers while the rest of the computation is done offloading, content caching and radio resource allocation
on the UEs [23], [129], [132]. For example, a user can offload in the fog-enabled IoT scenario. The joint decision-making
50% of the required CPU cycle of a task to the edge server for problem was formulated as the MDP in order to minimize
execution. In this way, the computation load for edge servers the end-to-end delay. Chen et al. [128] showed that the task
to execute a task can be relieved. Accordingly, the edge servers offloading problem is NP-hard and proposed a distributed
can serve more users with the same amount of computation game to efficiently find the offloading decision. In [124],
resource, ending up enhancing the network capacity of the edge they jointly optimized the offloading decision and computa-
system. In [23], they considered the computation offloading tional resource allocation to minimize the sum cost of delay
issue in a single-edge network architecture with the objec- and energy consumption for UEs. To minimize the over-
tive of minimizing the overall latency and power consumption. all users’ energy consumption with latency constraints in
Wang et al. [129] investigated the partial computation offloading a multiple-input–multiple-output (MIMO) multicell system,
problem to minimizes energy consumption and latency of appli- Sardellitti et al. [17] studied the joint allocation problem
cation on UEs. Hsu et al. [132] studied the allocation problem of radio and computational resources. In [125], they aimed
of communication and computation resources for task offload- to promote QoS by offloading the computational tasks to
ing in a 2-tier MEC-cloud framework in order to minimize multiple MEC servers by considering the allocation of both
energy consumption and offloading failure probability. radio resources and computation resources of the MEC servers.
In the multisubtasks task model [112], [120], [133], the Zhao et al. [96] addressed the optimization problem of offload-
computation tasks are divided into several subtasks and each ing ratio, transmission power, computational resources, and
of them can be offloaded to the edge servers. In some cases, subcarrier allocation with the aim of minimizing the overall
the task dependency between subtasks, a logical relation- energy consumption under the latency constraints. In [102],
ship in which a subtask relies on the execution results of the joint optimization of the radio resources and computa-
other subtasks before it can be successfully performed, is tion resources in Femto access points (FAPs) was obtained.
also considered. This would make it more challenging to Lyu et al. [103] introduced the Lagrange multiplier method
obtain the optimal offloading policy in multiple subtasks sce- to allocate computation resources to the offloaded tasks and
nario. The related articles are elaborated in the following. maximize the utility. To minimize the total energy consump-
Abdenacer et al. [112] minimized the tasks communica- tion of users in small cell networks with MEC systems,
tion cost using a greedy optimization approach according to Guo et al. [97] investigated offloading decision, spectrum,

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14319

power, and computation resource allocation problems and above. In this section, we will survey the previous efforts hav-
propose a hierarchical GA and PSO-based computation algo- ing been made to the management and orchestration of content
rithm. Kan et al. [20] optimized offloading decision and caching, including caching place, content popularity, content
resource allocation of radio and computational resources in types, resource allocation, and user mobility. A comparison of
a multiserver MEC system to enhance the QoS of edge the literature on content caching is shown in Table VI.
services. 1) Caching Place: Generally, edge caching considers store
4) Mobility Awareness of Task Offloading: Since the wire- contents at nodes closed to the RAN. In more detail, there
less channel capacity between BS and UE will vary dynami- are two main caching places discussed in the existing cache
cally owing to UE’s mobility, the offloading processes may fail scheme, which are BS and UE, respectively. The related
if UEs move outside BS’s coverage. Also, to ensure the success work on designing caching policy in these two caching
of task offloading, sufficient resources must be allocated to places and the corresponding issues will be elaborated in the
each UE, which makes the orchestration of task offloading following.
more challenging due to the limitation of the computation As mobile users require higher throughput and lower
and wireless resources in the edge server. Thus, there are network latency, caching at BSs becomes a promising solu-
several works that concern user mobility in their offloading tion where the contents can be delivered directly with high
strategies, where the BS obtains the optimal offloading solu- bandwidth of RAN. Thus, many works have studied the man-
tions while considering the constraints of limited resources agement of constrained cache space for caching at BSs. The
and UEs’ mobility. Zhan et al. [100] studied the problem optimization of edge caching in fog RANs (F-RANs) was
of offloading decisions and resource allocation over multiple studied in [137], where a distributed caching scheme was
BSs with the awareness of user mobility during the process proposed to minimize the request service delay and fron-
of task offloading. The computation offloading problem in thaul traffic load. Sun et al. [138] studied the cache placement
VEC scenario was investigated in [130]. They formulated the and bandwidth allocation problem in a multiuser system of
above problem as an MDP and utilized the proximal policy a heterogeneous network with the aim of minimizing outage
optimization (PPO) algorithm to address the enormous state probability. In [140], the content caching placement problem
space. In [106], the mobility of UEs was characterized by was studied aiming to maximize the energy efficiency of
the sojourn time in each coverage of FC nodes. They jointly heterogeneous networks.
optimized offloading decisions and computation resource allo- Caching at UEs is another potential scheme where mobile
cation to reduce the probability of migration. In [108], they users can store content locally and share it with other UEs
studied the efficient task offloading schemes in VEC networks in proximity directly via Device-to-Device (D2D) commu-
where the mobility of vehicles and the maximum latency of nication. In this way, the caching resources in UEs can
tasks were considered. be exploited and users’ QoE can be significantly enhanced.
However, caching at devices faces several technical challenges,
including limited storage space of devices, short transmission
coverage, the uncertainty of wireless connections, and intra-
B. Management and Orchestration of Content Caching cell interference. Thus, the caching strategy, including what
With the proliferation of mobile devices, the network traffic to cache and who to cache, for UEs requires being carefully
has grown exponentially in recent years, exerting a heavy load orchestrated to ensure the success of the D2D transmission.
on the backhaul network. Facing this, edge content caching, From the viewpoint of management, the UEs should cache
which utilizes the cache capacity in the edge area to store the contents that are popular in the edge area to increase the
popular content beforehand, has become a promising solution. likelihood of content sharing.
Once the content is cached, the requests for the same content Some caching design for D2D communication is presented
can be satisfied by fetching the requested content from the in the following. The multiwinner auction-based caching
proximate edge server directly. In this way, the duplicate traffic placement in D2D-enabled cellular networks was studied
in the backhaul network will be significantly alleviated. in [148]. Zhao et al. [149] proposed a D2D caching
The benefits of content caching can be various. From the scheme for small-cell networks aiming to maximize the
users’ perspective, content caching can improve users’ QoE overall offloading probability based on the known content
since they can enjoy the cached network services with less popularity.
latency. On the other hand, operators can save their cost in 2) Content Popularity: To improve the performance of
providing the same network service. In spite of the benefits mobile-edge caching, the cache hit ratio is basic criteria that
of content caching, it also incurs technical issues in the man- should be followed, which refers to the ratio of the number of
agement and orchestration of the edge system. Specifically, cached files requested by the users over the total number of
owing to the scarce storage resources of edge nodes and the requests. Intuitively, the edge nodes should cache the content
dynamic nature of the wireless network, the performance of with high popularity to enhance the hit ratio since the more
content caching will be severely influenced by the fluctuation popular a cached content is, the higher possibility it may be
of content popularity, different content types, and user mobil- requested in the future. Thus, the popularity of content is the
ity. Consequently, how to properly manage the distributed key factor that should be considered when it comes to man-
storage resource and orchestrate the caching decision of edge aging and orchestrating the scarce storage resources in edge
nodes becomes the key factor to fulfill the benefits mentioned nodes.

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14320 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

TABLE VI
C OMPARISON OF THE L ITERATURE ON C ONTENT C ACHING

[134]
[135]
[136]
[137]
[138]
[139]
[140]
[141]
[142]
[143]
[144]
[145]
[146]
[147]
[148]
[149]
[150]
[151]
[152]
[153]
[154]
[155]
[156]
[157]

To begin with, most of the existing literature on edge We start from the caching design for the video streaming
caching assume that the popularity of contents is static over a service, which accounts for 80% of Internet traffic. Compared
time period [137], [138], [140], [148], [149]. The commonly to traditional file downloading services, video streaming has
used popularity model is the Zipf model, which describes the more complex features. For instance, when designing the
phenomenon that most people are interested in a few popular caching policy for the video file, we need to determine which
contents and these few contents account for major traffic load. video segment to cache and their corresponding resolution ver-
For example, 10% of the online videos on the Internet account sion so as to optimize users’ QoE. Hence, the way we design
for nearly 80% of the views. the caching policy for video should be different accordingly.
However, since the content popularity may vary with time, The caching design of video files was studied in [136],
may differ from regions and cannot be known to the cache [141], [147], [154], and [156]. In order to track the dynam-
system beforehand, the caching policy adopting the Zipf model ics of the environment, the authors in [136] jointly considered
may be less effective in the real world. Thus, there are edge caching and video transcoding in the constrained MDP
several works proposed to design the online cache method (CMDP). They adopted asynchronous advantage actor-critic
that did not require the knowledge of content popularity (A3C) algorithm to solve the corresponding MDP issues.
beforehand. To consider the dynamic variation of content Han et al. [?] studied the proactive edge caching problem in
information and guarantee the freshness of user-received con- order to maximize the weighted average QoE of VoD services,
tents, Zhang et al. [134] proposed a refreshing scheme for where the caching fraction and the encoding bitrate of each
edge caching that update the version of the cached content video were optimized. A novel mechanism to jointly pro-
items according to user requests. Zhang et al. [135] proposed vide proactive caching, bandwidth provisioning, and adaptive
a α-fair utility-oriented vehicular edge caching scheme in order video streaming was designed for software-defined wireless
to achieve higher hit ratio and avoid the interruption of caching networks (SDWNs) in [141]. Song et al. [147] proposed a
services. Muller et al. [145] proposed a novel context-aware QoE-driven caching method in vehicular network by lever-
proactive caching method, which can learn context-specific aging class-based user interest model and deep deterministic
popularity online by observing context information of users policy gradient (DDPG) algorithm. To investigate the impact
and updating the cached content. In [146], an online edge of collaborative caching on QoE and backhaul data traffic,
caching and delivery schemes with time-varying content pop- Thar et al. [154] proposed a self-tuned bitrate selection algo-
ularity were studied. Zhang et al. [155] proposed a digital- rithm and an efficient cache replacement strategy. A novel col-
twin-empowered edge caching framework that schedules the laborative social-aware and QoE-driven video caching scheme
caching capability of roadside units (RSUs) and smart vehicles was proposed in [156]. By modeling the information dissemi-
based on user preference similarity and service availability. nation process among users in the social community, they are
3) Caching of Different File Types: We continue to address the first one to embed the social impact of online users into
the issues regarding how to cache different types of content. the video caching mechanism in the network layer.

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14321

On the other hand, the caching for the IoT devices becomes significantly decline due to the interruption of content deliv-
more and more important as billions of IoT devices generate ery. Thus, the impact of user mobility on the content caching
a large amount of monitoring, measurement, and automa- at different cells should be quantified, and the challenges
tion data. Nonetheless, IoT data are quite different from of what and where to place the contents among the cache
multimedia data because it has much shorter lifetimes. Thus, nodes should be addressed. The related work focusing on
different caching policies are also in need. Vural et al. [143] mobility-aware content caching strategy is presented in the
proposed a communication cost and IoT data freshness-aware following. In [139], they studied the issue of mobility-aware
caching model in order to reduce the network traffic by content caching, which was formulated into an optimization
storing the highly requested data. Yao and Ansari [144] investi- problem to maximize the caching utility. In [142], they inves-
gated the cache-enabled energy harvesting aided IoT network tigated both user mobility and video segment popularity
and formulated a Stackelberg game to observe the interac- for a cache-enabled C-RAN and proposed two correspond-
tions between caching and energy harvesting. To minimize ing proactive caching strategies. In [150], they proposed a
the microservice retrieval delay and improve the cache hit blockchain empowered architecture to ensure secure vehicular
ratio, Tian et al. [157] formulated the cooperative microser- content caching. DDPG was adopted to learn dynamic network
vice caching problem as an MDP and proposed a distributed topology and time-variant wireless channel.
deep RL (DRL)-based caching algorithm.
4) Resource Management of Content Caching: After elab- VI. E MERGING A PPLICATIONS AND R ELATED R ESEARCH
orating on the different aspects of the content caching tech- Combining edge computing systems with novel applications
nique, we now turn to address the resource management is a popular topic. These papers often focus on how to design
issues in content caching. As mentioned before, the key an edge-based or edge-assisted system for specific applications
elements involved in the content caching process are the that have rarely been combined with edge computing before,
storage resource. Thus, here we will survey how to effi- such as volumetric video, panoramic video, e-health, mixed
ciently orchestrate the distributed storage resource among edge reality (MR), and smart factory. There are several characteristics
nodes. of the emerging applications: the application type, the edge
Traditionally, the caching policies of edge systems are server position as we categorized in Section III-A2, use of
designed in a distributed manner, where each BS is only aware cloud servers, the caching mechanism, and network and context
of neighboring BSs’ information and then update local caches. awareness. The overall research works are listed in Table VII.
However, since the caching capacity of each BS is limited,
independently designing a caching policy for each BS would A. Application Type
result in low utilization of caching resources. This happens
There are a lot of applications for the edge computing
when the cached content in some BSs is duplicate. To this
system, leveraging its advantages of lower latency and less
end, collaborative caching policies have been proposed, where
backhaul traffic. Because of taking high percentage of global
BSs are able to share their cached contents with each other to
traffic, video applications, such as video streaming, video
avoid duplicates, ending up improving the resource utilization
delivering, and surveillance system (video processing) are
of the edge servers and caching efficiency.
especially popular. Some researchers start to investigate in
However, caching status information exchanges among
VR, AR, volumetric video, panoramic video, and 4K video
edge nodes may induce the signaling overheads in collabo-
streaming. Aside from the video applications, V2X and smart
rative caching. Also, the delay of searching and retrieving
healthcare are other two popular use cases that suit the
contents from other caches need to be taken into account
edge computing environment. Different applications tend to
when designing collaborative caching strategies. Hence, an
have different management structure as they have different
efficient collaborative caching policy should be designed
requirements and specialization.
with minimum overhead [151], [152], [153]. In [151], an
The normal video delivering and video streaming is the basis
efficient collaborative edge caching strategy was proposed
of all video applications we discuss in this section. The con-
by leveraging horizontal collaborative between the BSs.
cept of the combination of Edge computing system with the
Wu et al. [152] considered both fixed and mobile caching
video delivering application is based on using the Edge servers
nodes and proposed a hybrid edge caching scheme in order
to fully or partially replace the traditional content delivery
to maximize the profits of participants with BS–vehicle coop-
network (CDN), as we can see in [158], [159], and [160]. For
eration. Jiang et al. [153] proposed an improved brain storm
example, an edge-assisted system in [159] introduced to pro-
optimization (BSO) approach to solve the hierarchical coop-
vide CDN functionality as a service. Taleb et al. introduced the
erative caching optimization problem.
CDN slice, a CDN service instance created by the request of
5) Mobility Awareness of Content Caching: Owing to user content providers and can be managed automatically. There are
mobility, UEs will trigger the service migration and han- two components designed in their scheme: 1) CDNaaS orches-
dover when moving out of an edge node’s coverage. This trator (CDN-O) and 2) CDNaaS slice coordinator (CDN-SC),
process would cause additional latency and energy consump- which both interface with multiple MEOs for slice deployment
tion. Furthermore, the migration failure may occur due to late and management. The CDN-O is responsible for resource
migration or early migration, which increases the waste of allocation. The CDN-SC monitors and manages each slice
storage space. As a result, the QoS of edge services may autonomously. Martin et al. proposed a MEC system in [160]

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14322 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

TABLE VII
C OMPARISON OF THE L ITERATURE ON A PPLICATIONS AND O PTIMIZATION

[158]
[159]
[160]
[18]
[161]
[16]
[162]
[163]
[164]
[165]
[166]
[167]
[168]
[169]
[170]
[19]
[171]
[172]
[173]
[174]
[175]
[176]
[177]
[178]
[179]
[180]
[181]
[182]
[183]
[184]
[185]
[186]
[99]
[187]
[188]

with three video delivery mechanisms. In their system, the put a VR mobile application in a 5G testbed. As a basic
MEC server located in the OAI eNodeB connected to the experiment, the authors focused on testing the MEC-based
EPC gateway communicated with the CDN servers through VR application under several scenarios with different quality
the Internet and was able to get knowledge of the radio link requirements. More particularly, Kumar et al. [173] explored
in order to influence the client-side decisions. the characteristics of satellite and terrestrial backhaul links,
Unlike the “general” video applications above, special video and selected the backhaul links based on the requirements
applications may combine the VR, AR, and video/image pro- and the specific contexts. In the process, the clients send the
cessing. A volumetric video streaming system with Edge video requests to the MEC server, which is responsible for the
offloading and the optimization algorithm was introduced dynamic link selection.
in [167]. Qian et al. proposed a scheme where the Edge Video processing also takes an important part in video
transcodes the Point Cloud streaming to the regular pixel-based applications. Contrary to the video delivering systems that
video and then sends the video to the mobile device for decod- focus more on the downlink traffic, video processing focuses
ing. Schneider et al. [171] introduced an Edge-assisted AR more on the uplink traffic [174], [175], [176]. Pietrini et al.
remote support application. The operator’s mobile AR device gave an interesting idea that implemented the object detection
camera captures frames at a specific frame rate and sends to the service in the IoT scenario in [174]. In their design, the cam-
Edge server. The server finishes the tracking process and for- eras send the photographs to the fog nodes for processing
wards the video stream to a remote expert. Cho et al. proposed and then send the results to the cloud for further analysis.
ACACIA, a service abstraction framework that enables contin- Al-Saleh et al. [175] designed session management proto-
uous interactive applications on mobile edge in [172]. Through cols, where a method of key distribution was developed to
the proximity service discovery, it was possible to optimize the secure channel security in the whole surveillance system.
network based on the user context information. Datta [170] Bellavista et al. proposed an ML-assisted infrastructure to

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14323

learn a shared model with face recognition as the appli- for many use cases, reduce the transmission delay, enhance
cation in [177]. In the system, the cloud acts as a global the reliability and improve its management efficiency. MEC-
supervisor, while the local knowledge is stored at the edge. based management and orchestration for the smart factory
Pasandi and Nadeem introduced a intelligent cross-camera is introduced in [99], [186], [187], and [188]. Coutinho and
video analytics at the edge in [176]. In their design, the col- Boukerche [186] proposed a shared edge caching framework
laboration mechanism reduces energy consumption, bandwidth considering different applications, content catalogs and con-
overheads and latency, and provides better accuracy while tent request characteristics from devices and vehicles in the
ensuring privacy. smart factory. As mentioned in Section V, Zhang and Wei [99]
V2X is also a popular application that can integrate with edge aimed at solving the uncertainty problem of channel access
computing system [178], [179], [180], [181]. Avino et al. [178] delay in IEEE 802.11ax (WiFi 6) protocol and optimizing
proposed a MEC-based architecture system that leverages the the deployment of decomposable tasks. The 802.11 protocol
data collected by the network infrastructure to enforce the is responsible for providing connectivity to the sensors and
sensor measurement on the vehicles and further improve the is adopted in the smart factory scenario. To tackle with
security. In their design, the MEC host is connected to the the need of delivering offloaded workloads to several edge
OAI EPC. Mämmelä et al. [179] studied the possibilities and computing servers simultaneously, a nonorthogonal multiple
advantages of local mobile network edge application supporting access (NOMA) assisted MEC framework was investigated
vehicles in different scenarios. In the architecture, the data pro- in [187] to minimize the energy consumption under a joint
cessing tasks are offloaded to the Edge servers. Ojanperä et al. optimization of the multiaccess multitask offloading, trans-
proposed a generic system architecture for the 5G-enabled mission and resource allocation. A solution to the problem
road safety services in [180]. The architecture is composed of balancing cloud-edge collaboration, resource consumption
of four layers: 1) core (Cloud); 2) edge; 3) roadside; and and QoSs is introduced in [188]. The authors designed a deep-
4) vehicle, which is similar to [56]. The authors also proposed Q-learning-based SFC mapping algorithm (DQL-SFCM) for
a layered information broker approach, in which each entity the optimal solution under several service requirements and
has a particular local manager for data processing and man- network environments in the smart factory.
agement. Lu et al. [181] introduced a Vehicle-Edge-Cloud
closed-loop framework for Compressive Imaging processing on
connected and autonomous vehicles. In this system, both band-
width, memory footprint, and energy consumption are saved B. Edge Server Position
effectively. On the other hand, a point cloud feature-based As the papers enhancing their architectures, the position of
cooperative perception framework (F-Cooper) was proposed edge servers reflects the focus of the system. Placing the edge
in [189] for connected autonomous vehicles to achieve a better at the BS is a common choice for all types of applications.
object detection precision. Chen et al. [190] introduced a low- The video services are able to leverage the edge servers closer
latency, high-level (L3) data sharing protocol for connected and to their users to significantly reduce the delay. The intelli-
autonomous vehicular networks. However, the point cloud data gent health services are more flexible as they need to tailor
contains private information, such as vehicle identity, location, the system to various requirements, and the RAN-side ones
and trajectory, directly uploading the raw point cloud to the usually require real-time experience.
edge nodes or other vehicle will lead to serious privacy leak- Ghazal et al. introduced a MEC network architecture that
age [191]. Bi et al. [191] proposed a privacy-preserving object can be used to collect and transport the crowd-sourced and
detection framework over random point cloud shares for CAVs geo-mapped panoramic videos via their customized mobile
(referred to SecPCV), aiming to guarantee the privacy of both applications in [168]. The user-generated videos are sent from
point cloud and object detection results. a vehicle to the CN and then delivered to the users through the
MEC-based Intelligent healthcare system is also an emerg- MEC servers on the BSs. Islam and Shin proposed an e-Health
ing topic, which is investigated in [182], [183], [184], application in [183]. In their system, they combined unmanned
and [185]. The MR application also plays an important role aerial vehicles (UAVs) and IoT, with MEC, to provide a
in this case. Akter et al. introduced a MEC-based MR appli- real-time health monitoring service. They also implemented a
cation in [182] that helps visually impaired people by telling Blockchain-based method to address the data integrity issue.
them the items around them via the object detection technique. Placing the Edge on the CN side, Makris et al. [18]
Most of the computing tasks are offloaded to the MEC located developed a controller interface to select the optimal radio
in the BS to reduce energy consumption on UE. access technology for each user by integrating heterogeneous
Intelligent manufacturing is also a promising topic for radio resource in the cell.
industrialized countries, leading to the proposal of the smart Compared to the cases that place the Edge system on
factory [192], [193]. Smart factory enables advanced automa- the RAN side, placing it behind the CN side is more cen-
tion based on the interconnected network. By connecting tralized [184], [185]. Li et al. proposed a typical e-Health
devices, machinery and production systems, the smart fac- application for video-based heart rate detection with various
tory can collect and share large manufacturing data, and meet offloading models, which also focuses on real-time health data
the related task requirements. Edge computing supports its processing in [184]. Another e-Health application in [185]
need by providing computing and storage resources closer to runs in the Fog-assisted distributed analytic system. In their
the IIoT devices. Therefore, it can integrate a large system system, the fall application (U-Fall) does sensing and data

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14324 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

capturing on smartphones, and is connected to the back- video streaming service. The MEC server in the eNodeB is
end via the network. Collaborative detection on front-end responsible for selecting the bitrate according to the LTE chan-
and back-end can improve the accuracy and reduce the false nel conditions. Radio Network Information Service collects
alarms. A new algorithm based on the acceleration and information and delivers it to the MEC server based on the
nonlinear time series analysis was proposed to process fall context in the DASH-based framework. Another DASH-based
detection. video delivery system with MEC assistance was proposed by
Liu et al. [169] placed its edge servers connected between Li et al [163]. As above, their solution collects the information
the CN and RAN as an external entity. Compared to [167], of available bandwidth to improve the performance in the
this article combined the Edge-assisted panoramic VR video multiaccess heterogeneous network environment with the
streaming system with the modern multiconnectivity-based advantage of high resilience.
mmWave cellular network. In their architecture, mmWave can What is more, some of the works use machine learning to
be used to support the high bandwidth needed in the system. track the traffic trend in order to improve the user QoEs [164],
In rare cases, the Edge servers can be placed behind the [165], [166]. Yang et al. implemented a MEC-assisted mobile
CN, where the edge system behaves like the cloud. video streaming system in [164]. In their proposed scheme,
the machine learning technology was used to predict popular
C. Cloud videos and radio channel qualities. The authors also mod-
According to the different needs, some papers may include eled and investigated the efficiency of the MEC architecture
the cloud environment into their overall scope. In this case, to improve users’ QoE. Another example of a MEC-assisted
the management and orchestration become more complicated streaming system with machine learning was proposed by
and flexible because of the possibilities of computing task Zhou et al. in [165]. In their system, the Q-learning algo-
offloading. rithm was implemented to help determine the video quality or
Quadri et al. [161] investigated the video streaming applica- bitrate that is most suitable for the user’s network conditions.
tion from another angle. The authors modeled the network con- Chang et al. [166] presented an Edge-assisted adaptive video
dition by Mininet and managed the resources by OpenVolcano streaming scheme with Q-learning techniques and bandwidth
to simulate the network infrastructure of a mobile operator sharing policies, with video adaption based on the real-time
in Milano city. In [19], a MEC-VR application was proposed network information.
with a novel scheme to provide high-quality VR video without In summary, we introduce several novel applications that are
the frame prerendering or viewpoint detection. In the proposed integrated with edge computing system, including the video
scheme, the Edge servers are placed at the RAN and behind the processing, AR, VR, V2X, and edge-based healthcare. We cat-
CN, respectively, where the “MEC-VR” servers at the RAN egorize these emerging application with their application type,
are connected via the user plane function (UPF). edge server position, cloud server usage, the caching mecha-
nism, and the awareness of the network and context. Among
D. Cache these papers, we can see the edge-based or edge-assisted video
applications as the mainstream, and most of them are dedi-
For the video applications, especially the ones that make the cated to reduce the end-to-end latency by various approaches.
Edge servers play the role of CDN, caching is an important On the other hand, there is less discussion of edge-cloud col-
topic that causes significant performance differences. A system laborations, which is often an important issue in a realistic,
with a high cache hit rate will save more computing and commercial system.
network resources and, thus, is more efficient and powerful.
Due to the above reasons, some papers optimize their designs
by some well-developed caching mechanism [158], [159]. VII. C HALLENGES , O PEN I SSUES , AND
Viola et al. [158] presented a proxy that caches the contents F UTURE R ESEARCH D IRECTIONS
on the network edge to reduce the capital expenditure of the So far, we have presented a thorough and comprehen-
CDN and is able to handle the CDN malfunction by switching sive survey of the literature surrounding the management and
the sessions to other CDNs. orchestration of edge computing. The emergence of edge
computing brings many advantages to mobile networks and
E. Awareness provides new opportunities for future network services. A
lot of efforts from both academia and industry have been
To further improve the performance of the system, some
made to enhance the performance of the edge computing
systems are aware to the network or context information,
system, however, there are still some scopes, which are not
which helps Edge servers to make smart decisions accord-
completely developed yet. This section discusses the cur-
ing to the different conditions [16], [159], [162], [164], [165],
rent technical challenges and meanwhile identifies the future
[166].
research directions.
Nikaein et al. designed a flexible, programmable MEC plat-
form with a real-time content caching application in [16].
The authors implemented the RAN-aware content optimization A. Architecture Enhancement and Integration
by redirecting the App to a relay server, and further mod- The architectures of the edge computing system have been
ified the streaming bitrate based on the RAN conditions. carefully introduced in the literature, but we find out that the
Wang et al. [162] applied the MEC system to the adaptive complete architecture under heterogeneous networks has not

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14325

been well addressed yet. For example, the top-down com- B. Machine Learning for Management and Orchestration
munication protocols among various units of a heterogeneous For the task offloading with the assist of edge computing,
network are not fully investigated, i.e., from UAVs and satel- the joint problem of offloading decision and resource alloca-
lites to sensors and cameras. Therefore, it is necessary to figure tion is generally NP-hardness and, therefore, is intractable to
out some advanced and flexible architectures to make sure that be solved. The difficulty of the optimization problem is fur-
everything could be operated properly under a single system ther exacerbated due to the dynamic changes of the network
manager. environments. Content caching also faces the challenge of
When trying to exploit SDN and NFV techniques to edge fast-changing content request patterns when determining to
networks, we may face two main challenges as follows. The capture those popular contents. Recently, machine learning
first is the integration of different types of resources. It (ML) techniques [22] have become a promising tool for
should be noted that the SDN and NFV techniques trans- the orchestration of task offloading and content caching.
form the conventional network from hardware-oriented to Specifically, DRL, which learns the characteristics of time-
software-oriented and, therefore, lead to a big difference to the varying systems by observing the system state, actions, and
network deployment and management. The integration of var- rewards, can provide fast and near-optimal solutions to the
ious resources becomes more important especially when there optimization problem in edge networks [197].
are lots of infrastructures scattered in the edge network. In this Many research works have studied the problems behind
way, the critical issue is how to dynamically allocate various the edge-assisted applications by some ML-based techniques,
resources and meanwhile offer end-to-end services according however, the central implementation of an ML algorithm still
to the user demands [21], [194]. faces serious challenges. The critical challenge is that the con-
The second is the integration of SDN and NFV tech- ventional ML approaches could not meet the requirements for
niques. As we know that both SDN and NFV techniques the large-scale system because they require a large complex-
have a virtualization platform and an orchestrator/controller ity for the training process and, therefore, have high demands
to manage the resources. The difference is that the SDN on storage, computation, and network resources. A poten-
helps the network operators to optimize the network traffic tial solution is the distributed ML approach, which divides
via centralized management of large amounts of resources the computation task of a learning algorithm into multiple
distributed in the edge environments while the NFV bene- subtasks and then allocates them to some distributed MEC
fits the edge network operators by providing various kinds servers. When a distributed ML approach is applied, a series
of services. Note that the SDN and NFV are implemented of questions need to be exhaustively answered: what kind of
independently at present, the network performance could computation tasks can be divided? How to divide an original
be improved if they are integrated together. It is encour- computation task into multiple subtasks? How to synchronize
aged that the telecom operators, which have developed their the output information generated from different MEC servers?
own white papers based on the framework of NFV manage- And how to integrate those outputs into a complete output for
ment and orchestration, will further consider the integrated the master learning model? Besides, either the conventional
framework of SDN and NFV and achieve a unified envi- ML approaches or the distributed ML approaches may raise
ronment with automated intelligent control capability to the privacy and security issues to the training process, which will
operator [195], [196]. be discussed later.
More and more applications and use cases combine with
edge computing scenarios. The video application is the most
popular choice due to the high requirements of the users
as well as the industry maturity. The V2X, IoT, and smart C. Security and Privacy in Edge Computing
healthcare are difficult to be implemented and, therefore, few Though the edge computing architecture brings some benefits
research works involve the related topics. What is more, those to the network applications, at the same time it raises unforeseen
highly integrated applications, such as smart cities and smart security and privacy issues. As we know that those network
factories, are seldom discussed practically. Currently, the edge virtualization technologies, i.e., SDN and NFV, enable different
system has been applied to serve in multiple small cells with services to co-exist on the same physical hardware devices [198].
limited capacities so that dynamic network slicing technique The services and applications in the edge networks are different,
can play a key role in it. As the traffic grows, the QoE which means that their privacy and security requirements are
of users as well as the system performance will be signifi- also diverse. How to ensure the satisfaction of each user’s
cantly reduced if without any diversion mechanism. Moreover, request with its security and privacy concern in co-existing
the combination of computing and caching resources will environments will be one of the future research directions [35],
change the design philosophy of traditional communication [199], [200]. Besides, edge computing is centered around the
systems to a large extent. To fully exploit the potential of complex interweaving of multiple and varied technologies,
task offloading and content caching, the efficient integra- such as peer-to-peer systems, wireless networks, and network
tion of these resources for all users and applications has not virtualization, and so. Some unique and unstudied scenarios in
been concluded. In this way, a more comprehensive resource the orchestration of task offloading and content caching create
allocation scheme is urgently needed and the research on the potential for original channels of malicious behaviors, e.g.,
this direction is continuously developed with the evolving of the interplay of heterogeneous edge nodes and the migration
the network. of services across global and local scales.

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14326 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

Those ML-based techniques adopted to realize edge intel- [2] Y. Zhao, W. Wang, Y. Li, C. C. Meixner, M. Tornatore, and J. Zhang,
ligence may cause severe privacy concerns. Take the com- “Edge computing and networking: A survey on infrastructures and
applications,” IEEE Access, vol. 7, pp. 101213–101230, 2019.
putation offloading in the edge network as an example. To [3] N. Abbas, Y. Zhang, A. Taherkordi, and T. Skeie, “Mobile edge com-
determine the optimal offloading decisions, users are required puting: A survey,” IEEE Internet Things J., vol. 5, no. 1, pp. 450–465,
to report their private information to the edge server, such Feb. 2018.
as channel gain, current battery level, and computation tasks. [4] R. K. Naha et al., “Fog computing: Survey of trends, architec-
tures, requirements, and research directions,” IEEE Access, vol. 6,
Such behavior will encounter the risk of data leakage by pp. 47980–48009, 2018.
eavesdroppers, which could be further illegally utilized to [5] P. Habibi, M. Farhoudi, S. Kazemian, S. Khorsandi, and A. LeonGarcia,
predict the user location. “Fog computing: A comprehensive architectural survey,” IEEE Access,
vol. 8, pp. 69105–69133, 2020.
For this situation, some solutions have been raised.
[6] P. Mach and Z. Becvar, “Mobile edge computing: A survey on archi-
The following and [116], [117], [118] are some exam- tecture and computation offloading,” IEEE Commun. Surveys Tuts.,
ples. Xiong et al. [201] proposed an edge-assisted privacy- vol. 19, no. 3, pp. 1628–1656, 3rd Quart., 2017.
preserving raw data sharing framework for connected [7] Y. Mao, C. You, J. Zhang, K. Huang, and K. Letaief, “A survey
on mobile edge computing: The communication perspective,” IEEE
autonomous vehicles. By encrypting raw data into two cipher- Commun. Surveys Tuts., vol. 19, no. 4, pp. 2322–2358, 4th Quart.,
texts and, respectively, resolved by two privacy-preserving 2017.
convolutional neural networks (P-CNN), the receiving vehicle [8] S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, and W. Wang, “A
survey on mobile edge networks: Convergence of computing, caching
can get extracted data with high accuracy without leak- and communications,” IEEE Access, vol. 5, pp. 6757–6779, 2017.
ing private information. Xiong et al. [202] proposed an [9] M. Mehrabi, D. You, V. Latzko, H. Salah, M. Reisslein, and F. Fitzek,
artificial-intelligence-enabled three-party game (ATG) frame- “Device-enhanced MEC: Multi-access edge computing aided by end
work, which includes a sensing data classification-anonymity device computation and caching: A survey,” IEEE Access, vol. 7,
pp. 166079–166108, 2019.
(CA) model and a three-party game (TG) model to protect [10] W. Yu et al., “A survey on the edge computing for the Internet of
user privacy on sensitive data and analyze data privacy leakage Things,” IEEE Access, vol. 6, pp. 6900–6919, 2018.
in different phases in the mobile-edge crowdsensing (MECS) [11] X. Wang, Y. Han, V. Leung, D. Niyato, X. Yan, and X. Chen,
framework. Bi et al. proposed a dynamic privacy measurement “Convergence of edge computing and deep learning: A comprehensive
survey,” IEEE Commun. Surveys Tuts., vol. 22, no. 2, pp. 869–904, 2nd
and protection (DPMP) framework [203] to maintain balance Quart., 2020.
between users’ data benefits and privacy threats for MECS. [12] C.-H. Hong and B. Varghese, “Resource management in fog/edge com-
DPMP includes a DPM model for quantifying the amount of puting: A survey on architectures, infrastructure, and algorithms,” ACM
Comput. Surveys, vol. 52, no. 5, p. 97, Sep. 2019. [Online]. Available:
data privacy and calculating privacy threshold for individual https://doi.org/10.1145/3326066
users and a Dynamic Private sensing data Selection (DPS) [13] B. Sonkoly, J. Czentye, M. Szalay, B. Németh, and L. Toka, “Survey on
algorithm for maximizing user’s data benefits according to placement methods in the edge and beyond,” IEEE Commun. Surveys
their privacy thresholds. Tuts., vol. 23, no. 4, pp. 2590–2629, 4th Quart., 2021.
[14] Q. Luo, S. Hu, C. Li, G. Li, and W. Shi, “Resource scheduling in edge
For us, we have considered federated learning (FL) [204] computing: A survey,” IEEE Commun. Surveys Tuts., vol. 23, no. 4,
can be a potential direction because it enables multiple users to pp. 2131–2165, 4th Quart., 2021.
collaboratively learn a shared model while keeping their own [15] Y. Zhang, J. Ren, J. Liu, C. Xu, H. Guo, and Y. Liu, “A survey
on emerging computing paradigms for big data,” Chin. J. Electron.,
data locally. More specifically, each user learns the offloading vol. 26, no. 1, pp. 1–12, 2017. [Online]. Available: https://ietresearch.
decision individually based on its local information and the onlinelibrary.wiley.com/doi/abs/10.1049/cje.2016.11.016
edge server is merely responsible for the global model update. [16] N. Nikaein, X. Vasilakos, and A. Huang, “LL-MEC: Enabling low
latency edge applications,” in Proc. IEEE 7th Int. Conf. Cloud Netw.
(CloudNet), 2018, pp. 1–7.
VIII. C ONCLUSION [17] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of
This article surveys and summarizes the research efforts radio and computational resources for multicell MobileEdge comput-
ing,” IEEE Trans. Signal Inf. Process. Netw., vol. 1, no. 2, pp. 89–103,
made on service orchestration and resource management of Jun. 2015.
edge computing. We first presented an overview of edge [18] N. Makris, V. Passas, C. Nanis, and T. Korakis, “On minimizing service
computing, including architectures, advantages, enabling tech- access latency: Employing MEC on the fronthaul of heterogeneous
5G architectures,” in Proc. IEEE Int. Symp. Local Metrop. Area Netw.
nologies and standardizations. Then, the state-of-the-art tech- (LANMAN), 2019, pp. 1–6.
niques in management and orchestration of edge computing [19] S. Shi, V. Gupta, M. Hwang, and R. Jana, “Mobile VR on edge cloud:
are presented, including task offloading, content caching and A latency-driven design,” in Proc. MMSys, Jun. 2019, pp. 222–231.
VNE. Third, the state-of-the-art research on the infrastructure [20] T. Kan, Y. Chiang, and H. Wei, “QoS-aware mobile edge computing
system: Multi-server multi-user scenario,” in Proc. IEEE Globecom
of edge computing is discussed from the aspect of architec- Workshops (GC Wkshps), 2018, pp. 1–6.
ture enhancement and optimization for emerging applications. [21] M. Bagaa, T. Taleb, J. B. Bernabe, and A. Skarmeta, “QoS
Finally, we point out the future research directions of these hot and resource-aware security orchestration and life cycle manage-
ment,” IEEE Trans. Mobile Comput., vol. 21, no. 8, pp. 2978–2993,
topics, such as IoT applications, machine learning techniques, Aug. 2022.
and security issues. [22] C.-H. Hsu, Y. Chiang, Y. Zhang, and H.-Y. Wei, “Mobility-aware QoS
promotion and load balancing in MEC-based vehicular networks: A
deep learning approach,” in Proc. IEEE 93rd Veh. Technol. Conf. (VTC-
R EFERENCES Spring), 2021, pp. 1–6.
[1] T. Taleb, K. Samdanis, B. Mada, H. Flinck, S. Dutta, and D. Sabella, [23] H. Zhang, W. Wu, C. Wang, M. Li, and R. Yang, “Deep reinforcement
“On multi-access edge computing: A survey of the emerging 5G learning-based offloading decision optimization in mobile edge com-
network edge cloud architecture and orchestration,” IEEE Commun. puting,” in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), 2019,
Surveys Tuts., vol. 19, no. 3, pp. 1657–1681, 3rd Quart., 2017. pp. 1–7.

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14327

[24] A. Khan, W. Kellerer, K. Kozu, and M. Yabusaki, “Network sharing [49] S. Ghosh, “BlueArch—An implementation of 5G Testbed,” J.
in the next mobile network: TCO reduction, management flexibility, Commun., vol. 14, pp. 1110–1118, Dec. 2019.
and operational independence,” IEEE Commun. Mag., vol. 49, no. 10, [50] A. van Kempen, T. Crivat, B. Trubert, D. Roy, and G. Pierre, “MEC-
pp. 134–142, Oct. 2011. ConPaaS: An experimental single-board based mobile edge cloud,”
[25] M. Hoffmann and M. Staufer, “Network virtualization for future mobile in Proc. 5th IEEE Int. Conf. Mobile Cloud Comput. Services Eng.
networks: General architecture and applications,” in Proc. IEEE Int. (MobileCloud), 2017, pp. 17–24.
Conf. Commun. Workshops (ICC), 2011, pp. 1–5. [51] R. Muñoz et al., “The ADRENALINE testbed: An SDN/NFV
[26] J. G. Herrera and J. F. Botero, “Resource allocation in NFV: A com- packet/optical transport network and edge/core cloud platform for end-
prehensive survey,” IEEE Trans. Netw. Service Manag., vol. 13, no. 3, to-end 5G and IoT services,” in Proc. Eur. Conf. Netw. Commun.
pp. 518–532, Sep. 2016. (EuCNC), 2017, pp. 1–5.
[27] K. Hejja and X. Hesselbach, “Online power aware coordinated virtual [52] H. Hu et al., “Video surveillance on mobile edge networks—
network embedding with 5G delay constraint,” J. Netw. Comput. Appl., A reinforcement-learning-based approach,” IEEE Internet Things J.,
vol. 124, pp. 121–136, Dec. 2018. vol. 7, no. 6, pp. 4746–4760, Jun. 2020.
[28] “Multi-Access Edge Computing—Standards for MEC–ETSI.” [53] I. Sarrigiannis, K. Ramantas, E. Kartsakli, P. Mekikis,
Accessed: Jul. 20, 2020. [Online]. Available: https://www.etsi.org/ A. Antonopoulos, and C. Verikoukis, “Online VNF lifecycle
technologies/multi-access-edge-computing management in an MEC-enabled 5G IoT architecture,” IEEE Internet
[29] “P1935—Standard for edge/fog manageability and orchestration.” Things J., vol. 7, no. 5, pp. 4183–4194, May 2020.
Accessed: Dec. 6, 2017. [Online]. Available: https://standards.ieee.org/ [54] X. Hou, Z. Ren, K. Yang, C. Chen, H. Zhang, and Y. Xiao, “IIoT-MEC:
project/1935.html A novel mobile edge computing framework for 5G-enabled IIoT,” in
[30] S. Goundar, S. B. Bhushan, and P. K. Rayani, Architecture and Security Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), 2019, pp. 1–7.
Issues in Fog Computing Applications. New York, NY, USA: IGI [55] S. Shekhar, A. Chhokra, H. Sun, A. Gokhale, A. Dubey, and
Global, 2019. X. Koutsokos, “Supporting fog/edge-based cognitive assistance IoT
[31] T. H. Luan, L. Gao, Z. Li, Y. Xiang, G. Wei, and L. Sun. “Fog services for the visually impaired: Poster abstract,” in Proc. Int.
Computing: Focusing on Mobile Users at the Edge.” 2015. [Online]. Conf. Internet Things Design Implement. (IoTDI), 2019, pp. 275–276.
Available: https://arxiv.org/abs/1502.01815 [Online]. Available: https://doi.org/10.1145/3302505.3312592
[32] SDN Architecture, Issue 1.1, TR-521, ONF, Menlo Park, CA, USA, [56] Y. Wang, J. Wang, Y. Ge, B. Yu, C. Li, and L. Li, “MEC support for
2016. C-V2X system architecture,” in Proc. IEEE 19th Int. Conf. Commun.
[33] Network Functions Virtualisation (NFV); Management and Technol. (ICCT), 2019, pp. 1375–1379.
Orchestration, document GS NFV-MAN 001, ETSI, Sophia Antipolis, [57] B. P. Rimal, D. P. Van, and M. Maier, “Cloudlet enhanced fiber-wireless
France, 2014. access networks for mobile-edge computing,” IEEE Trans. Wireless
[34] G. Schaffrath et al., “Network virtualization architecture: Proposal and Commun., vol. 16, no. 6, pp. 3601–3618, Jun. 2017.
initial prototype,” in Proc. 1st ACM Workshop Virtualized Infrastruct. [58] S. Cao, H. Han, J. Wei, Y. Zhao, S. Yang, and L. Yan, “Space cloud-
Syst. Archit., 2009, pp. 63–72. fog computing: Architecture, application and challenge,” in Proc. 3rd
Int. Conf. Comput. Sci. Appl. Eng. (CSAE), 2019, pp. 1–7. [Online].
[35] A. Fischer, J. F. Botero, M. T. Beck, H. De Meer, and X. Hesselbach,
Available: https://doi.org/10.1145/3331453.3361637
“Virtual network embedding: A survey,” IEEE Commun. Surveys Tuts.,
[59] C. Politis, K. Liolis, M. Corici, E. Troudt, Z. Szabó, and J. Cahill,
vol. 15, no. 4, pp. 1888–1906, 4th Quart., 2013.
“Design of moving experimentation facility to showcase satellite inte-
[36] Network Functions Virtualisation (NFV); Resiliency Requiremetns,
gration into 5G,” in Proc. Eur. Conf. Netw. Commun. (EuCNC), 2019,
document GS NFV-REL 001, ETSI, Sophia Antipolis, France, 2015.
pp. 177–181.
[37] Mobile Edge Computing (MEC); Framework and Reference
[60] D. Bernstein, “Containers and cloud: From LXC to docker to kuber-
Architecture, document GS MEC 003, ETSI, Sophia Antipolis,
netes,” IEEE Cloud Comput., vol. 1, no. 3, pp. 81–84, Sep. 2014.
France, 2019.
[61] V. Medel, O. Rana, J. N. Bañares, and U. Arronategui, “Modelling
[38] G. Wang, C. Li, Y. Huang, X. Wang, and Y. Luo, “Smart contract- performance & resource management in kubernetes,” in Proc.
based caching and data transaction optimization in mobile edge IEEE/ACM 9th Int. Conf. Utility Cloud Comput. (UCC), 2016,
computing,” Knowl. Based Syst., vol. 252, Sep. 2022, Art. no. 109344. pp. 257–262.
[Online]. Available: https://www.sciencedirect.com/science/article/pii/ [62] S. Ifrah, Deploying Containerized Applications With Amazon ECS.
S0950705122006748 Berkeley, CA, USA: Apress, 2019, pp. 83–133. [Online]. Available:
[39] A. Alwarafy, K. A. Al-Thelaya, M. Abdallah, J. Schneider, and https://doi.org/10.1007/978-1-4842-5101-0_3
M. Hamdi, “A survey on security and privacy issues in edge-computing- [63] S. Mathew and J. Varia, Overview of Amazon Web services, Amazon,
assisted Internet of Things,” IEEE Internet Things J., vol. 8, no. 6, Washington, DC, USA, 2014.
pp. 4004–4022, Mar. 2021. [64] A. Jeffery, H. Howard, and R. Mortier, “Rearchitecting kubernetes for
[40] D. Krishnaswamy, R. Krishnan, D. Lopez, P. Willis, and A. Qamar, “An the edge,” in Proc. 4th Int. Workshop Edge Syst. Anal. Netw. (EdgeSys),
open NFV and cloud architectural framework for managing application 2021, pp. 7–12. [Online]. Available: https://doi.org/10.1145/3434770.
virality behaviour,” in Proc. 12th Annu. IEEE Consum. Commun. Netw. 3459730
Conf. (CCNC), 2015, pp. 746–754. [65] K. Kaur, S. Garg, G. Kaddoum, S. H. Ahmed, and M. Atiquzzaman,
[41] C. Pham, H. Dai Tran, S. I. Moon, K. Thar, and C. S. Hong, “A general “KEIDS: Kubernetes-based energy and interference driven scheduler
and practical consolidation framework in CloudNFV,” in Proc. ICOIN, for Industrial IoT in edge-cloud ecosystem,” IEEE Internet Things J.,
2015, pp. 295–300. vol. 7, no. 5, pp. 4228–4237, May 2020.
[42] A. Shrivastwa, Hybrid Cloud for Architects: Build Robust Hybrid Cloud [66] Z. Yan, N. Wei, Q. Jin, and X. Zhou, “Latency-aware resource-efficient
Solutions Using AWS and OpenStack. Birmingham, U.K.: Packt, 2018. virtual network embedding in software defined networking,” in Proc.
[43] Open Source Mano, ETSI, Sophia Antipolis, France, 2016. 28th Wireless Opt. Commun. Conf. (WOCC), 2019, pp. 1–5.
[44] “Intel.com—RIFT.” Accessed: Jun. 23, 2022. [Online]. Available: [67] Y. Chiang, Y.-H. Chao, C.-H. Hsu, C.-T. Chou, and H.-Y. Wei, “Virtual
https://networkbuilders.intel.com/ecosystem/rift network embedding with dynamic speed switching orchestration in
[45] R. Bruschi, P. Lago, G. Lamanna, C. Lombardo, and S. Mangialardi, fog/edge network,” IEEE Access, vol. 8, pp. 84753–84768, 2020.
“Openvolcano: An open-source software platform for fog computing,” [68] H. Yao, S. Ma, J. Wang, P. Zhang, C. Jiang, and S. Guo, “A
in Proc. PROCON@ITC, vol. 2, 2016, pp. 22–27. continuous-decision virtual network embedding scheme relying on rein-
[46] G. A. Carella and T. Magedanz, “Open baton: A framework for virtual forcement learning,” IEEE Trans. Netw. Service Manag., vol. 17, no. 2,
network function management and orchestration for emerging software- pp. 864–875, Jun. 2020.
based 5G networks,” Newsletter, vol. 2016, p. 190, Jul. 2016. [69] J. Li, N. Zhang, Q. Ye, W. Shi, W. Zhuang, and X. Shen, “Joint resource
[47] L. T. Bolivar, C. Tselios, D. M. Area, and G. Tsolis, “On the deploy- allocation and online virtual network embedding for 5G networks,” in
ment of an open-source, 5G-aware evaluation testbed,” in Proc. 6th Proc. IEEE Global Commun. Conf. (GLOBECOM), 2017, pp. 1–6.
IEEE Int. Conf. Mobile Cloud Comput. Services Eng. (MobileCloud), [70] M. Yu, Y. Yi, J. Rexford, and M. Chiang, “Rethinking virtual network
2018, pp. 51–58. embedding: Substrate support for path splitting and migration,” ACM
[48] D. Scotece, N. R. Paiker, L. Foschini, P. Bellavista, X. Ding, and SIGCOMM Comput. Commun. Rev., vol. 38, no. 2, pp. 17–29, 2008.
C. Borcea, “MEFS: Mobile edge file system for edge-assisted mobile [71] I. Fajjari, N. Aitsaadi, G. Pujolle, and H. Zimmermann, “VNE-AC:
apps,” in Proc. IEEE 20th Int. Symp. World Wireless Mobile Multimedia Virtual network embedding algorithm based on ant colony metaheuris-
Netw. (WoWMoM), 2019, pp. 1–9. tic,” in Proc. IEEE Int. Conf. Commun. (ICC), 2011, pp. 1–6.

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14328 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

[72] M. Zhu, Q. Sun, S. Zhang, P. Gao, B. Chen, and J. Gu, “Energy-aware [97] F. Guo, H. Zhang, H. Ji, X. Li, and V. C. M. Leung, “An efficient
virtual optical network embedding in sliceable-transponder-enabled computation offloading management scheme in the densely deployed
elastic optical networks,” IEEE Access, vol. 7, pp. 41897–41912, 2019. small cell networks with mobile edge computing,” IEEE Trans. Netw.,
[73] Y. Zong et al., “Location-aware energy efficient virtual network embed- vol. 26, no. 6, pp. 2651–2664, Dec. 2018.
ding in software-defined optical data center networks,” IEEE/OSA J. [98] B. Shang, L. Liu, and Z. Tian, “Deep learning-assisted energy-efficient
Opt. Commun. Netw., vol. 10, no. 7, pp. 58–70, Jun. 2018. task offloading in vehicular edge computing systems,” IEEE Trans. Veh.
[74] A. Fischer, M. T. Beck, and H. De Meer, “An approach to energy- Technol., vol. 70, no. 9, pp. 9619–9624, Sep. 2021.
efficient virtual network embeddings,” in Proc. IFIP/IEEE Int. Symp. [99] Y. Zhang and H.-Y. Wei, “Risk-aware cloud-edge computing frame-
Integr. Netw. Manage. (IM), 2013, pp. 1142–1147. work for delay-sensitive industrial IoTs,” IEEE Trans. Netw. Service
[75] J. F. Botero, X. Hesselbach, M. Duelli, D. Schlosser, A. Fischer, Manag., vol. 18, no. 3, pp. 2659–2671, Sep. 2021.
and H. De Meer, “Energy efficient virtual network embedding,” IEEE [100] W. Zhan, C. Luo, G. Min, C. Wang, Q. Zhu, and H. Duan, “Mobility-
Commun. Lett., vol. 16, no. 5, pp. 756–759, May 2012. aware multi-user offloading optimization for mobile edge computing,”
[76] P. Zhang, “Incorporating energy and load balance into virtual IEEE Trans. Veh. Technol., vol. 69, no. 3, pp. 3341–3356, Mar. 2020.
network embedding process,” Comput. Commun., vol. 129, pp. 80–88, [101] Y. Wei, F. R. Yu, M. Song, and Z. Han, “Joint optimization of caching,
Sep. 2018. computing, and radio resources for fog-enabled IoT using natural actor–
[77] M. R. Rahman and R. Boutaba, “SVNE: Survivable virtual network critic deep reinforcement learning,” IEEE Internet Things J., vol. 6,
embedding algorithms for network virtualization,” IEEE Trans. Netw. no. 2, pp. 2061–2073, Apr. 2019.
Service Manag., vol. 10, no. 2, pp. 105–118, Jun. 2013. [102] O. Munoz, A. Pascual-Iserte, and J. Vidal, “Optimization of radio and
[78] N. Shahriar, R. Ahmed, S. R. Chowdhury, A. Khan, R. Boutaba, computational resources for energy efficiency in latency constrained
and J. Mitra, “Generalized recovery from node failure in virtual application offloading,” IEEE Trans. Veh. Technol., vol. 64, no. 10,
network embedding,” IEEE Trans. Netw. Service Manag., vol. 14, no. 2, pp. 4738–4755, Oct. 2015.
pp. 261–274, Jun. 2017. [103] X. Lyu, H. Tian, C. Sengul, and P. Zhang, “Multiuser joint task offload-
[79] M. M. A. Khan, N. Shahriar, R. Ahmed, and R. Boutaba, “Multi- ing and resource optimization in proximate clouds,” IEEE Trans. Veh.
path link embedding for survivability in virtual networks,” IEEE Trans. Technol., vol. 66, no. 4, pp. 3435–3447, Apr. 2017.
Netw. Service Manag., vol. 13, no. 2, pp. 253–266, Jun. 2016. [104] W. Fan, L. Yao, J. Han, F. Wu, and Y. Liu, “Game-based multi-type
[80] S. Gong, J. Chen, X. Yin, and Q. Zhu, “Survivable virtual network task offloading among mobile-edge-computing-enabled base stations,”
embedding across multiple domains,” in Proc. 2nd IEEE Int. Conf. IEEE Internet Things J., vol. 8, no. 24, pp. 17691–17704, Dec. 2021.
Comput. Commun. (ICCC), 2016, pp. 2391–2396. [105] X. Chen, H. Zhang, C. Wu, S. Mao, Y. Ji, and M. Bennis, “Optimized
[81] S. Taeb et al., “Virtual network embedding with path-based latency computation offloading performance in virtual edge computing systems
guarantees in elastic optical networks,” in Proc. IEEE 27th Int. Conf. via deep reinforcement learning,” IEEE Internet Things J., vol. 6, no. 3,
Netw. Protocols (ICNP), 2019, pp. 1–12. pp. 4005–4018, Jun. 2019.
[82] G. Chochlidakis and V. Friderikos, “Low latency virtual network [106] D. Wang, Z. Liu, X. Wang, and Y. Lan, “Mobility-aware task offload-
embedding for mobile networks,” in Proc. IEEE Int. Conf. Commun. ing and migration schemes in fog computing networks,” IEEE Access,
(ICC), 2016, pp. 1–6. vol. 7, pp. 43356–43368, 2019.
[107] J. Zhou and X. Zhang, “Fairness-aware task offloading and resource
[83] M. T. Beck and C. Linnhoff-Popien, “On delay-aware embedding of
allocation in cooperative mobile edge computing,” IEEE Internet
virtual networks,” in Proc. 6th Int. Conf. Adv. Future Internet (AFIN),
Things J., vol. 9, no. 5, pp. 3812–3824, Mar. 2022.
2014, pp. 1–9.
[108] C. Yang, Y. Liu, X. Chen, W. Zhong, and S. Xie, “Efficient mobility-
[84] L. Qu, C. Assi, and K. Shaban, “Delay-aware scheduling and
aware task offloading for vehicular edge computing networks,” IEEE
resource optimization with network function virtualization,” IEEE
Access, vol. 7, pp. 26652–26664, 2019.
Trans. Commun., vol. 64, no. 9, pp. 3746–3758, Sep. 2016.
[109] M. M. et al., “Task data offloading and resource allocation in fog
[85] S. Haeri, Q. Ding, Z. Li, and L. Trajković, “Global resource capacity
computing with multi-task delay guarantee,” IEEE Access, vol. 7,
algorithm with path splitting for virtual network embedding,” in Proc.
pp. 152911–152918, 2019.
IEEE Int. Symp. Circuits Syst. (ISCAS), 2016, pp. 666–669.
[110] N. Kiran, C. Pan, S. Wang, and C. Yin, “Joint resource allocation
[86] X. Chang, B. Wang, and J. Liu, “Network state aware virtual and computation offloading in mobile edge computing for SDN based
network parallel embedding,” in Proc. IEEE 31st Int. Perform. Comput. wireless networks,” J. Commun. Netw., vol. 22, no. 1, pp. 1–11, 2020.
Commun. Conf. (IPCCC), 2012, pp. 193–194.
[111] S. Pan, Z. Zhang, Z. Zhang, and D. Zeng, “Dependency-aware compu-
[87] M. Maier, M. Chowdhury, B. P. Rimal, and D. P. Van, “The tactile tation offloading in mobile edge computing: A reinforcement learning
Internet: Vision, recent progress, and open challenges,” IEEE Commun. approach,” IEEE Access, vol. 7, pp. 134742–134753, 2019.
Mag., vol. 54, no. 5, pp. 138–145, May 2016.
[112] N. Abdenacer, H. Wu, N. N. Abdelkader, S. Dhelim, and H. Ning,
[88] P. Zhang, X. Pang, Y. Bi, H. Yao, H. Pan, and N. Kumar, “DSCD: Delay “A novel framework for mobile edge computing by optimizing task
sensitive cross-domain virtual network embedding algorithm,” IEEE offloading,” IEEE Internet Things J., vol. 8, no. 16, pp. 13065–13076,
Trans. Netw. Sci. Eng., vol. 7, no. 4, pp. 2913–2925, Oct./Dec. 2020. Aug. 2021.
[89] V. M. Raee, A. Ebrahimzadeh, M. Rayani, R. H. Glitho, M. E. Barachi, [113] A. M. Alwakeel, “An overview of fog computing and edge computing
and F. Belqasmi, “Energy efficient virtual network embedding in virtu- security and privacy issues,” Sensors, vol. 21, no. 24, p. 8226, 2021.
alized wireless sensor networks,” in Proc. CNCC, 2022, pp. 187–192. [Online]. Available: https://www.mdpi.com/1424-8220/21/24/8226
[90] J. Kurose and K. Ross, Computer Networks: A Top Down Approach [114] Y. Xiao, Y. Jia, C. Liu, X. Cheng, J. Yu, and W. Lv, “Edge computing
Featuring the Internet. London, U.K.: Peorsoim Addison Wesley, 2010. security: State of the art and challenges,” Proc. IEEE, vol. 107, no. 8,
[91] D. G. Andersen, “Theoretical approaches to node assignment,” in Proc. pp. 1608–1631, Aug. 2019.
Comput. Sci., 2002, pp. 1–13. [115] P. Ranaweera, A. D. Jurcut, and M. Liyanage, “Survey on multi-access
[92] J. M. Kleinberg, “Approximation algorithms for disjoint paths prob- edge computing security and privacy,” IEEE Commun. Surveys Tuts.,
lems,” Ph.D. dissertation, Dept. Comput. Sci., Massachusetts Inst. vol. 23, no. 2, pp. 1078–1124, 2nd Quart., 2021.
Technol., Cambridge, MA, USA, 1996. [116] X. Zhang, J. Lu, and D. Li, “Confidential information protection
[93] S. G. Kolliopoulos and C. Stein, “Improved approximation algorithms method of commercial information physical system based on edge
for unsplittable flow problems,” in Proc. 38th Annu. Symp. Found. computing,” Neural Comput. Appl., vol. 33, no. 3, pp. 897–907, 2021.
Comput. Sci., 1997, pp. 426–436. [117] X. Xu, X. Zhang, H. Gao, Y. Xue, L. Qi, and W. Dou, “BeCome:
[94] P. Zhang, H. Yao, and Y. Liu, “Virtual network embedding based on Blockchain-enabled computation offloading for IoT in mobile edge
computing, network, and storage resource constraints,” IEEE Internet computing,” IEEE Trans. Ind. Informat., vol. 16, no. 6, pp. 4187–4195,
Things J., vol. 5, no. 5, pp. 3298–3304, Oct. 2018. Jun. 2020.
[95] Y.-L. Hsu, H.-Y. Wei, and M. Bennis, “Green fog offloading strategy [118] X. Jia, D. He, N. Kumar, and K.-K. R. Choo, “A provably secure and
for heterogeneous wireless edge networks,” in Proc. IEEE Globecom efficient identity-based anonymous authentication scheme for mobile
Workshops (GC Wkshps), 2018, pp. 1–6. edge computing,” IEEE Syst. J., vol. 14, no. 1, pp. 560–571, Mar. 2020.
[96] M. Zhao et al., “Energy-aware task offloading and resource allocation [119] Y. Zhang, C.-Y. Wang, and H.-Y. Wei, “Parking reservation auction for
for time-sensitive services in mobile edge computing systems,” IEEE parked vehicle assistance in vehicular fog computing,” IEEE Trans.
Trans. Veh. Technol., vol. 70, no. 10, pp. 10925–10940, Oct. 2021. Veh. Technol., vol. 68, no. 4, pp. 3126–3139, Apr. 2019.

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14329

[120] Y. Zhang, J.-H. Liu, C.-Y. Wang, and H.-Y. Wei, “Decomposable intel- [144] J. Yao and N. Ansari, “Caching in energy harvesting aided Internet of
ligence on cloud-edge IoT framework for live video analytics,” IEEE Things: A game-theoretic approach,” IEEE Internet Things J., vol. 6,
Internet Things J., vol. 7, no. 9, pp. 8860–8873, Sep. 2020. no. 2, pp. 3194–3201, Apr. 2019.
[121] A. U. Rehman, R. L. Aguiar, and J. P. Barraca, “Fault-tolerance in the [145] S. Muller, O. Atan, M. van der Schaar, and A. Klein, “Context-
scope of cloud computing,” IEEE Access, vol. 10, pp. 63422–63441, aware proactive content caching with service differentiation in wire-
2022. less networks,” IEEE Trans. Wireless Commun., vol. 16, no. 2,
[122] A. Javed, J. Robert, K. Heljanko, and K. Främling, “IoTEF: A fed- pp. 1024–1036, Feb. 2017.
erated edge-cloud architecture for fault-tolerant IoT applications,” J. [146] S. M. Azimi, O. Simeone, A. Sengupta, and R. Tandon, “Online edge
Grid Comput., vol. 18, no. 1, pp. 57–80, 2020. caching and wireless delivery in fog-aided networks with dynamic
[123] S. Tuli, G. Casale, and N. R. Jennings, “PreGAN: Preemptive migra- content popularity,” IEEE J. Sel. Areas Commun., vol. 36, no. 6,
tion prediction network for proactive fault-tolerant edge computing,” in pp. 1189–1202, Jun. 2018.
Proc. IEEE INFOCOM Conf. Comput. Commun., 2022, pp. 670–679. [147] C. Song, W. Xu, T. Wu, S. Yu, P. Zeng, and N. Zhang, “QoE-
[124] J. Li, H. Gao, T. Lv, and Y. Lu, “Deep reinforcement learning based driven edge caching in vehicle networks based on deep reinforcement
computation offloading and resource allocation for MEC,” in Proc. learning,” IEEE Trans. Veh. Technol., vol. 70, no. 6, pp. 5286–5295,
IEEE Wireless Commun. Netw. Conf. (WCNC), 2018, pp. 1–6. Jun. 2021.
[125] T. Kan, Y. Chiang, and H. Wei, “Task offloading and resource allo- [148] T. Zhang, X. Fang, Y. Liu, G. Y. Li, and W. Xu, “D2D-enabled mobile
cation in mobile-edge computing system,” in Proc. 27th Wireless Opt. user edge caching: A multi-winner auction approach,” IEEE Trans. Veh.
Commun. Conf. (WOCC), 2018, pp. 1–4. Technol., vol. 68, no. 12, pp. 12314–12328, Dec. 2019.
[126] W. Li, Q. Li, L. Chen, F. Wu, and J. Ren, “A storage resource col- [149] N. Zhao et al., “Caching D2D connections in small-cell networks,”
laboration model among edge nodes in edge federation service,” IEEE IEEE Trans. Veh. Technol., vol. 67, no. 12, pp. 12326–12338,
Trans. Veh. Technol., vol. 71, no. 9, pp. 9212–9224, Sep. 2022. Dec. 2018.
[127] B. Yang, X. Cao, J. Bassey, X. Li, T. Kroecker, and L. Qian, [150] Y. Dai, D. Xu, K. Zhang, S. Maharjan, and Y. Zhang, “Deep rein-
“Computation offloading in multi-access edge computing networks: forcement learning and permissioned blockchain for content caching in
A multi-task learning approach,” in Proc. IEEE Int. Conf. Commun. vehicular edge computing and networks,” IEEE Trans. Veh. Technol.,
(ICC), 2019, pp. 1–6. vol. 69, no. 4, pp. 4312–4324, Apr. 2020.
[128] X. Chen, L. Jiao, W. Li, and X. Fu, “Efficient multi-user computation [151] Q. Li, W. Shi, X. Ge, and Z. Niu, “Cooperative edge caching
offloading for mobile-edge cloud computing,” IEEE/ACM Trans. Netw., in software-defined hyper-cellular networks,” IEEE J. Sel. Areas
vol. 24, no. 5, pp. 2795–2808, Sep./Oct. 2016. Commun., vol. 35, no. 11, pp. 2596–2605, Nov. 2017.
[129] Y. Wang, M. Sheng, L. W. X. Wang, and J. Li, “Mobile-edge com- [152] R. Wu, G. Tang, T. Chen, D. Guo, L. Luo, and W. Kang, “A profit-
puting: Partial computation offloading using dynamic voltage scaling,” aware coalition game for cooperative content caching at the network
IEEE Trans. Commun., vol. 64, no. 10, pp. 4268–4282, Oct. 2016. edge,” IEEE Internet Things J., vol. 9, no. 2, pp. 1361–1373, Jan. 2022.
[130] W. Zhan et al., “Deep reinforcement learning-based offloading schedul- [153] Y. Jiang, X. Chen, F.-C. Zheng, D. Niyato, and X. You, “Brain storm
ing for vehicular edge computing,” IEEE Internet Things J., vol. 7, no. optimization-based edge caching in fog radio access networks,” IEEE
6, pp. 5449–5465, Jun. 2020. Trans. Veh. Technol., vol. 70, no. 2, pp. 1807–1820, Feb. 2021.
[154] K. Thar, N. H. Tran, T. Z. Oo, and C. S. Hong, “QoE-traffic
[131] M. Min, L. Xiao, Y. Chen, P. Cheng, D. Wu, and W. Zhuang,
optimization through collaborative dge caching in adaptive mobile
“Learning-based computation offloading for IoT devices with energy
video streaming,” IEEE Access, vol. 6, pp. 52261–52276, 2018.
harvesting,” IEEE Trans. Veh. Technol., vol. 68, no. 2, pp. 1930–1941,
[155] K. Zhang, J. Cao, S. Maharjan, and Y. Zhang, “Digital twin empowered
Feb. 2019.
content caching in social-aware vehicular edge networks,” IEEE Trans.
[132] C.-W. Hsu, Y.-L. Hsu, and H.-Y. Wei, “Energy-efficient and reliable
Comput. Social Syst., vol. 9, no. 1, pp. 239–251, Feb. 2022.
MEC offloading for heterogeneous Industrial IoT networks,” in Proc.
[156] Y. Chiang, C.-H. Hsu, and H.-Y. Wei, “Collaborative social-aware
Eur. Conf. Netw. Commun. (EuCNC), 2019, pp. 384–388.
and QoE-driven video caching and adaptation in edge network,” IEEE
[133] Y. Guo, Z. Mi, Y. Yang, and M. S. Obaidat, “An energy sensitive Trans. Multimedia, vol. 23, pp. 4311–4325, 2020.
computation offloading strategy in cloud robotic network 12 based on [157] H. Tian et al., “DIMA: Distributed cooperative microservice caching
GA,” IEEE Syst. J., vol. 13, no. 3, pp. 3513–3523, Sep. 2019. for Internet of Things in edge computing by deep reinforcement
[134] S. Zhang, L. Wang, H. Luo, X. Ma, and S. Zhou, “AoI-delay learning,” World Wide Web, vol. 25, pp. 1769–1792, Aug. 2022.
tradeoff in mobile edge caching with freshness-aware content refresh- [158] R. Viola, A. Martin, M. Zorrilla, and J. Montalbán, “MEC proxy for
ing,” IEEE Trans. Wireless Commun., vol. 20, no. 8, pp. 5329–5342, efficient cache and reliable multi-CDN video distribution,” in Proc.
Aug. 2021. IEEE Int. Symp. Broadband Multimedia Syst. Broadcast. (BMSB),
[135] Y. Zhang et al., “Towards hit-interruption tradeoff in vehicular edge 2018, pp. 1–7.
caching: Algorithm and analysis,” IEEE Trans. Intell. Transp. Syst., [159] Y. Tan, C. Han, M. Luo, X. Zhou, and X. Zhang, “Radio network-aware
vol. 23, no. 6, pp. 5198–5210, Jun. 2022. edge caching for video delivery in MEC-enabled cellular networks,”
[136] J. Luo, F. R. Yu, Q. Chen, and L. Tang, “Adaptive video stream- in Proc. IEEE Wireless Commun. Netw. Conf. Workshops (WCNCW),
ing with edge caching and video transcoding over software-defined 2018, pp. 179–184.
mobile networks: A deep reinforcement learning approach,” IEEE [160] A. Martin, R. Viola, M. Zorrilla, J. Flórez, P. Angueira, and
Trans. Wireless Commun., vol. 19, no. 3, pp. 1577–1592, Mar. 2020. J. Montalbán, “MEC for fair, reliable and efficient media streaming in
[137] Y. Jiang, Y. Hu, M. Bennis, F. Zheng, and X. You, “A mean field mobile networks,” IEEE Trans. Broadcast., vol. 66, no. 2, pp. 264–278,
game-based distributed edge caching in fog radio access networks,” Jun. 2020.
IEEE Trans. Commun., vol. 68, no. 3, pp. 1567–1580, Mar. 2020. [161] C. Quadri, S. Gaito, R. Bruschi, F. Davoli, and G. P. Rossi, “A
[138] W. Sun, Y. Li, C. Hu, and M. Peng, “Joint optimization of cache MEC approach to improve QoE of video delivery service in urban
placement and bandwidth allocation in heterogeneous networks,” IEEE spaces,” in Proc. IEEE Int. Conf. Smart Comput. (SMARTCOMP),
Access, vol. 6, pp. 37250–37260, 2018. 2018, pp. 25–32.
[139] Y. Guan, Y. Xiao, H. Feng, C. Shen, and L. J. Cimini, “MobiCacher: [162] C. Wang, Z. Lin, S. Yang, and P. Lin, “Mobile edge computing-enabled
Mobility-aware content caching in small-cell networks,” in Proc. IEEE channel-aware video streaming for 4G LTE,” in Proc. 13th Int. Wireless
Global Commun. Conf., 2014, pp. 4537–4542. Commun. Mobile Comput. Conf. (IWCMC), 2017, pp. 564–569.
[140] F. Gabry, V. Bioglio, and I. Land, “On energy-efficient edge caching in [163] Y. Li, P. A. Frangoudis, Y. Hadjadj-Aoul, and P. Bertin, “A mobile
heterogeneous networks,” IEEE J. Sel. Areas Commun., vol. 34, no. 12, edge computing-assisted video delivery architecture for wireless het-
pp. 3288–3298, Dec. 2016. erogeneous networks,” in Proc. IEEE Symp. Comput. Commun. (ISCC),
[141] C. Liang, Y. He, F. R. Yu, and N. Zhao, “Enhancing QoEAware wireless 2017, pp. 534–539.
edge caching with software-defined wireless networks,” IEEE Trans. [164] S. Yang, Y. Tseng, C. Huang, and W. Lin, “Multi-access edge comput-
Wireless Commun., vol. 16, no. 10, pp. 6912–6925, Oct. 2017. ing enhanced video streaming: Proof-of-concept implementation and
[142] J. Dai, Z. Zhang, and D. Liu, “Proactive caching over cloud radio access prediction/QoE models,” IEEE Trans. Veh. Technol., vol. 68, no. 2,
network with user mobility and video segment popularity awared,” pp. 1888–1902, Feb. 2019.
IEEE Access, vol. 6, pp. 44396–44405, 2018. [165] X. Zhou, Z. Chang, C. Sun, and X. Zhang, “Demo abstract: Context-
[143] S. Vural, P. Navaratnam, N. Wang, C. Wang, L. Dong, and R. Tafazolli, aware video streaming with Q-learning for MEC-enabled cellu-
“In-network caching of Internet-of-Things data,” in Proc. IEEE Int. lar networks,” in Proc. IEEE INFOCOM Conf. Comput. Commun.
Conf. Commun. (ICC), 2014, pp. 3185–3190. Workshops (INFOCOM WKSHPS), 2018, pp. 1–2.

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
14330 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 16, 15 AUGUST 2023

[166] Z. Chang, X. Zhou, Z. Wang, H. Li, and X. Zhang, “Edge-assisted [185] Y. Cao, S. Chen, P. Hou, and D. Brown, “FAST: A fog computing
adaptive video streaming with deep learning in mobile edge networks,” assisted distributed analytics system to monitor fall for stroke miti-
in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), 2019, gation,” in Proc. IEEE Int. Conf. Netw. Archit. Storage (NAS), 2015,
pp. 1–6. pp. 2–11.
[167] F. Qian, B. Han, J. Pair, and V. Gopalakrishnan, “Toward practical volu- [186] R. W. L. Coutinho and A. Boukerche, “Modeling and analysis of
metric video streaming on commodity smartphones,” in Proc. 20th Int. a shared edge caching system for connected cars and industrial
Workshop Mobile Comput. Syst. Appl. (HotMobile), 2019, pp. 135–140. IoT-based applications,” IEEE Trans. Ind. Informat., vol. 16, no. 3,
[Online]. Available: https://doi.org/10.1145/3301293.3302358 pp. 2003–2012, Mar. 2020.
[168] M. Ghazal, Y. AlKhalil, A. Mhanna, and F. Dehbozorgi, “Mobile [187] L. Qian, Y. Wu, F. Jiang, N. Yu, W. Lu, and B. Lin, “NOMA assisted
panoramic video maps over MEC networks,” in Proc. IEEE Wireless multi-task multi-access mobile edge computing via deep reinforcement
Commun. Netw. Conf., 2016, pp. 1–6. learning for Industrial Internet of Things,” IEEE Trans. Ind. Informat.,
[169] Y. Liu, J. Liu, A. Argyriou, and S. Ci, “MEC-assisted panoramic VR vol. 17, no. 8, pp. 5688–5698, Aug. 2021.
video streaming over millimeter wave mobile networks,” IEEE Trans. [188] S. Xu, Y. Li, S. Guo, C. Lei, D. Liu, and X. Qiu, “Cloud–edge col-
Multimedia, vol. 21, no. 5, pp. 1302–1316, May 2019. laborative SFC mapping for Industrial IoT using deep reinforcement
[170] S. K. Datta, “Virtual reality mobile application testing in a 5G Testbed,” learning,” IEEE Trans. Ind. Informat., vol. 18, no. 6, pp. 4158–4168,
in Proc. 11th Int. Conf. Ubiquitous Future Netw. (ICUFN), 2019, Jun. 2022.
pp. 455–459. [189] Q. Chen, X. Ma, S. Tang, J. Guo, Q. Yang, and S. Fu,
[171] M. Schneider, J. Rambach, and D. Stricker, “Augmented reality based “F-cooper: Feature based cooperative perception for autonomous
on edge computing using the example of remote live support,” in Proc. vehicle edge computing system using 3D point clouds,” in Proc.
IEEE Int. Conf. Ind. Technol. (ICIT), 2017, pp. 1277–1282. SEC, 2019, pp. 88–100. [Online]. Available: https://doi.org/10.1145/
[172] J. Cho, K. Sundaresan, R. Mahindra, J. Van der Merwe, and 3318216.3363300
S. Rangarajan, “ACACIA: Context-aware edge computing for con- [190] Q. Chen et al., “Low-latency high-level data sharing for connected and
tinuous interactive applications over mobile networks,” in Proc. 12th autonomous vehicular networks,” in Proc. ICII, 2019, pp. 287–296.
Int. Conf. Emerg. Netw. Exp. Technol. (CoNEXT), 2016, pp. 375–389. [191] R. Bi, J. Xiong, Y. Tian, Q. Li, and X. Liu, “Edge-cooperative
[Online]. Available: https://doi.org/10.1145/2999572.2999604 privacy-preserving object detection over random point cloud shares
[173] S. Kumar, N. Wang, C. Ge, and B. Evans, “Optimising layered for connected autonomous vehicles,” IEEE Trans. Intell. Transp. Syst.,
video content delivery based on satellite and terrestrial integrated vol. 23, no. 12, pp. 24979–24990, Dec. 2022.
5G networks,” in Proc. Eur. Conf. Netw. Commun. (EuCNC), 2019, [192] B. Chen, J. Wan, L. Shu, P. Li, M. Mukherjee, and B. Yin, “Smart
pp. 161–166. factory of industry 4.0: Key technologies, application case, and chal-
[174] R. Pietrini, D. Manco, M. Paolanti, V. Placidi, E. Frontoni, and lenges,” IEEE Access, vol. 6, pp. 6505–6519, 2018.
P. Zingaretti, “An IoT edge-fog-cloud architecture for vision based [193] B. Chen, J. Wan, A. Celesti, D. Li, H. Abbas, and Q. Zhang, “Edge
Planogram integrity,” in Proc. 13th Int. Conf. Distrib. Smart Cameras computing in IoT-based manufacturing,” IEEE Commun. Mag., vol. 56,
(ICDSC), 2019, pp. 1–8. [Online]. Available: https://doi.org/10.1145/ no. 9, pp. 103–109, Sep. 2019.
3349801.3349807 [194] W. Zhuang, Q. Ye, F. Lyu, N. Cheng, and J. Ren, “SDN/NFV-
[175] M. I. Al-Saleh, M. A. Alsmirat, Y. Jararweh, and I. Obaidat, “A unified empowered future IoV with enhanced communication, computing, and
key distribution and session management protocol for mobile video caching,” Proc. IEEE, vol. 108, no. 2, pp. 274–291, Feb. 2020.
surveillance systems,” in Proc. 5th Int. Conf. Internet Things Syst. [195] M. Garrich, F.-J. Moreno-Muro, M.-V. B. Delgado, and P. P. Marino,
Manag. Security, 2018, pp. 234–238. “Open-source network optimization software in the open SDN/NFV
[176] H. B. Pasandi and T. Nadeem, “CONVINCE: Collaborative cross- transport ecosystem,” J. Lightw. Technol., vol. 37, no. 1, pp. 75–88,
camera video Analytics at the edge,” in Proc. IEEE Int. Conf. Jan. 1, 2019.
Pervasive Comput. Commun. Workshops (PerCom Workshops), 2020, [196] B. Sonkoly et al., “5G applications from vision to reality: Multi-
pp. 1–5. operator orchestration,” IEEE J. Sel. Areas Commun., vol. 38, no. 7,
[177] P. Bellavista, P. Chatzimisios, L. Foschini, M. Paradisioti, and pp. 1401–1416, Jul. 2020.
D. Scotece, “A support infrastructure for machine learning at the edge [197] S. Guo, Y. Dai, S. Xu, X. Qiu, and F. Qi, “Trusted cloud-edge network
in smart city surveillance,” in Proc. IEEE Symp. Comput. Commun. resource management: DRL-driven service function chain orchestra-
(ISCC), 2019, pp. 1189–1194. tion for IoT,” IEEE Internet Things J., vol. 7, no. 7, pp. 6010–6022,
[178] G. Avino et al., “A MEC-based extended virtual sensing for automotive Jul. 2020.
services,” in Proc. AEIT Int. Conf. Elect. Electron. Technol. Autom. [198] A. M. Zarca, J. B. Bernabe, A. Skarmeta, and J. M. A. Calero, “Virtual
(AEIT AUTOMOTIVE), 2019, pp. 1–6. IoT HoneyNets to mitigate cyberattacks in SDN/NFV-enabled IoT
[179] O. Mämmelä, T. Ojanperä, J. Mäkelä, O. Martikainen, and J. Väisänen, networks,” IEEE J. Sel. Areas Commun., vol. 38, no. 6, pp. 1262–1277,
“Evaluation of LiDAR data processing at the mobile network edge for Jun. 2020.
connected vehicles,” in Proc. Eur. Conf. Netw. Commun. (EuCNC), [199] I. Farris, T. Taleb, Y. Khettab, and J. Song, “A survey on emerging
2019, pp. 83–88. SDN and NFV security mechanisms for IoT systems,” IEEE Commun.
[180] T. Ojanperä, J. Mäkelä, O. Mämmelä, M. Majanen, and O. Martikainen, Surveys Tuts., vol. 21, no. 1, pp. 812–837, 1st Quart., 2018.
“Use cases and communications architecture for 5G-enabled road [200] S. Liu, Z. Cai, H. Xu, and M. Xu, “Security-aware virtual network
safety services,” in Proc. Eur. Conf. Netw. Commun. (EuCNC), 2018, embedding,” in Proc. IEEE Int. Conf. Commun. (ICC), 2014,
pp. 335–340. pp. 834–840.
[181] S. Lu, X. Yuan, and W. Shi, “Edge compression: An integrated [201] J. Xiong, R. Bi, M. Zhao, J. Guo, and Q. Yang, “Edge-assisted privacy-
framework for compressive imaging processing on CAVs,” in Proc. preserving raw data sharing framework for connected autonomous vehi-
IEEE/ACM Symp. Edge Comput. (SEC), 2020, pp. 125–138. cles,” IEEE Wireless Commun., vol. 27, no. 3, pp. 24–30, Jun. 2020.
[182] A. Akter, A. Islam, and S. Y. Shin, “Mobile edge computing based [202] J. Xiong, M. Zhao, M. Z. A. Bhuiyan, L. Chen, and Y. Tian, “An
mixed reality application for the assistance of blind and visually AI-enabled three-party game framework for guaranteed data privacy in
impaired people,” in Proc. 7th Int. Conf. Inf. Commun. Technol. mobile edge Crowdsensing of IoT,” IEEE Trans. Ind. Informat., vol. 17,
(ICoICT), 2019, pp. 1–5. no. 2, pp. 922–933, Feb. 2021.
[183] A. Islam and S. Y. Shin, “BHMUS: Blockchain based secure outdoor [203] R. Bi, M. Zhao, Z. Ying, Y. Tian, and J. Xiong, “Achieving dynamic
health monitoring scheme using UAV in smart city,” in Proc. 7th Int. privacy measurement and protection based on reinforcement learning
Conf. Inf. Commun. Technol. (ICoICT), 2019, pp. 1–6. for mobile edge crowdsensing of IoT,” Digit. Commun. Netw., to be
[184] X. Li et al., “COMEC: Computation offloading for video-based heart published. [Online]. Available: https://www.sciencedirect.com/science/
rate detection APP in mobile edge computing,” in Proc. IEEE Int. article/pii/S2352864822001614
Conf. Parallel Distrib. Process. Appl. Ubiquitous Comput. Commun. [204] D. C. Nguyen, M. Ding, P. N. Pathirana, A. Seneviratne, J. Li,
Big Data Cloud Comput. Soc. Comput. Netw. Sustain. Comput. and H. V. Poor, “Federated learning for Internet of Things: A com-
Commun. (ISPA/IUCC/BDCloud/SocialCom/SustainCom), 2018, prehensive survey,” IEEE Commun. Surveys Tuts., vol. 23, no. 3,
pp. 1038–1039. pp. 1622–1658, 3rd Quart., 2021.

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.
CHIANG et al.: MANAGEMENT AND ORCHESTRATION OF EDGE COMPUTING FOR IoT: A COMPREHENSIVE SURVEY 14331

Yao Chiang (Student Member, IEEE) received the Guan-Hao Chen received the B.S. degree in elec-
B.S. degree and the M.S. degree in management trical engineering from National Taiwan University,
information systems from the National Pingtung Taipei, Taiwan, in 2020, where he is currently pur-
University of Science and Technology, Pingtung, suing the M.S. degree in electrical engineering.
Taiwan, in 2014 and 2016, respectively, and the His research interests include edge computing and
Ph.D. degree in electrical engineering from National machine learning.
Taiwan University (NTU), Taipei, Taiwan, in 2021.
He is currently a Postdoctoral Research Fellow
with NTU. His research interests include data min-
ing, machine learning, and mobile communications
design for multiaccess edge computing systems.

Huan-Ting Chen received the B.S. degree in elec-

trical engineering from National Taiwan University,
Taipei, Taiwan, in 2021, where he is currently pur-
suing the M.S. degree in electrical engineering.
His research interests include edge computing,
video analytics, and machine learning.
Yi Zhang (Member, IEEE) received the B.S.
degree in software engineering from Software
College, Xiamen University, Xiamen, China, in
2014, and the M.S. and Ph.D. degrees from the
Graduate Institute of Communication Engineering,
National Taiwan University, Taipei, Taiwan, in
2016.
He was an Assistant Engineer with Quanzhou
Institute of Equipment Manufacturing, Haixi
Institutes, Chinese Academy of Sciences,
Quanzhou, Fujian, China, from 2016 to 2017. He Yan-Jhu Wang is currently pursuing the M.S.
is currently an Assistant Professor with the Department of Information degree in communication engineering with National
and Communication Engineering, Xiamen University. His research interests Taiwan University, Taipei, Taiwan.
include mobile and wireless networking, fog/edge computing, and game Her research interests focused on related topics in
theoretical models for communications networks. virtual network embedding.

Hao Luo received the B.S. and M.S. degrees in elec- Hung-Yu Wei (Senior Member, IEEE) received the
trical engineering from National Taiwan University, B.S. degree in electrical engineering from National
Taipei, Taiwan, in 2020 and 2022, respectively. Taiwan University, New Taipei, Taiwan, and the
From 2018 to 2020, he was a Research Intern M.S. and the Ph.D. degrees in electrical engineering
with the Institute of Information Science, Academia from Columbia University, New York, NY, USA.
Sinica, Taipei. His research interests include wire- He is a Professor with the Department of
less communication, machine learning, and edge Electrical Engineering and Graduate Institute of
computing. Communications Engineering, National Taiwan
University. His research interests include 6G wire-
less, IoT, and fog/edge computing.
Dr. Wei serves as a Chair of IEEE P1935 for
edge/fog management standard and IEEE P1934 for edge/fog architecture
standard.

Chun-Ting Chou (Member, IEEE) received the

B.S. and M.S. degrees in electrical engineering from
National Taiwan University, Taipei, Taiwan, in 1995
and 1997, and the Ph.D. degree in electrical engi-
Tse-Yu Chen received the B.S. and M.S. degrees neering from the University of Michigan, Ann Arbor,
in electrical engineering from National Taiwan MI, USA, in 2004.
University, Taipei, Taiwan, in 2019 and 2021, He was a Senior Member Research Staff with
respectively. Philips Research North America, New York, NY,
His research interests include system design, USA, from 2004 to 2007. He is currently an
video streaming, and resource management for mul- Associate Professor with the Graduate Institute
tiaccess edge computing systems. of Communication Engineering, National Taiwan
University. His research interests include dynamic spectrum access, medium
access control design, wireless and mobile communications, and Internet of
Things.

Authorized licensed use limited to: Universidad de Castilla La Mancha. Downloaded on April 02,2024 at 09:43:41 UTC from IEEE Xplore. Restrictions apply.