1

A Survey on Mobile Edge Computing: The

Communication Perspective
Yuyi Mao, Changsheng You, Jun Zhang, Kaibin Huang, and Khaled B. Letaief

Abstract—Driven by the visions of Internet of Things and their processor speeds are growing exponentially, following
5G communications, recent years have seen a paradigm shift in Moore’s Law. Harvesting the vast amount of the idle compu-
mobile computing, from the centralized Mobile Cloud Computing tation power and storage space distributed at the network edges
towards Mobile Edge Computing (MEC). The main feature of
MEC is to push mobile computing, network control and storage can yield sufficient capacities for performing computation-
to the network edges (e.g., base stations and access points) so as intensive and latency-critical tasks at mobile devices. This
arXiv:1701.01090v4 [cs.IT] 13 Jun 2017

to enable computation-intensive and latency-critical applications paradigm is called Mobile Edge Computing (MEC) [5]. While
at the resource-limited mobile devices. MEC promises dramatic long propagation delays remain a key drawback for Cloud
reduction in latency and mobile energy consumption, tackling Computing, MEC, with the proximate access, is widely agreed
the key challenges for materializing 5G vision. The promised
gains of MEC have motivated extensive efforts in both academia to be a key technology for realizing various visions for next-
and industry on developing the technology. A main thrust of generation Internet, such as Tactile Internet (with millisecond-
MEC research is to seamlessly merge the two disciplines of scale reaction time) [6], Internet of Things (IoT) [7], and
wireless communications and mobile computing, resulting in a Internet of Me [8]. Presently, researchers from both academia
wide-range of new designs ranging from techniques for compu- and industry have been actively promoting MEC technology
tation offloading to network architectures. This paper provides a
comprehensive survey of the state-of-the-art MEC research with by pursuing the fusion of techniques and theories from both
a focus on joint radio-and-computational resource management. disciplines of mobile computing and wireless communications.
We also discusse a set of issues, challenges and future research This paper aims at providing a survey of key research progress
directions for MEC research, including MEC system deployment, in this young field from the communication perspective. We
cache-enabled MEC, mobility management for MEC, green shall also present a research outlook containing an ensemble
MEC, as well as privacy-aware MEC. Advancements in these
directions will facilitate the transformation of MEC from theory of promising research directions for MEC.
to practice. Finally, we introduce recent standardization efforts
on MEC as well as some typical MEC application scenarios.
A. Mobile Computing for 5G: From Clouds to Edges
Index Terms—Mobile edge computing, fog computing, mobile
cloud computing, computation offloading, resource management, In the past decade, the popularity of mobile devices and the
green computing. exponential growth of mobile Internet traffic have been driving
the tremendous advancements in wireless communications and
I. I NTRODUCTION networking. In particular, the breakthroughs in small-cell net-
The last decade has seen Cloud Computing emerging as a works, multi-antenna, and millimeter-wave communications
new paradigm of computing. Its vision is the centralization of promise to provide users gigabit wireless access in next-
computing, storage and network management in the Clouds, generation systems [9]. The high-rate and highly-reliable air
referring to data centers, backbone IP networks and cellular interface allows to run computing services of mobile devices
core networks [1], [2]. The vast resources available in the at the remote cloud data center, resulting in the research area
Clouds can then be leveraged to deliver elastic computing called Mobile Cloud Computing (MCC). However, there is
power and storage to support resource-constrained end-user an inherent limitation of MCC, namely, the long propagation
devices. Cloud Computing has been driving the rapid growth distance from the end user to the remote cloud center, which
of many Internet companies. For example, the Cloud business will result in excessively long latency for mobile applications.
has risen to be the most profitable sector for Amazon [3], and MCC is thus not adequate for a wide-range of emerging
Dropbox’s success depended highly on the Cloud service of mobile applications that are latency-critical. Presently, new
Amazon. network architectures are being designed to better integrate the
However, in recent years, a new trend in computing is hap- concept of Cloud Computing into mobile networks, as will be
pening with the function of Clouds being increasingly moving discussed in the latter part of this article.
towards the network edges [4]. It is estimated that tens of bil- In 5G wireless systems, ultra-dense edge devices, including
lions of Edge devices will be deployed in the near future, and small-cell base stations (BSs), wireless access points (APs),
laptops, tablets, and smartphones, will be deployed, each
Y. Mao, J. Zhang and K. B. Letaief are with the Dept. of Elec- having a computation capacity comparable with that of a
tronic and Computer Engineering, The Hong Kong University of Science
and Technology, Hong Kong (Email: [email protected], [email protected], computer server a decade ago. As such, a large population of
[email protected]). K. B. Letaief is also affiliated with Hamad bin Khalifa devices will be idle at every time instant. It will, in particular,
University, Doha, Qatar. be harvesting enormous computation and storage resources
C. You and K. Huang are with the Dept. of Electrical and Elec-
tronic Engineering, The University of Hong Kong, Hong Kong (Email: available at the network edges, which will be sufficient to
[email protected], [email protected]). enable ubiquitous mobile computing. In a nutshell, the main
 2

target of wireless systems, from 1G to 4G, is the pursuit of

increasingly higher wireless speeds to support the transition Image Face PreͲ Feature
from voice-centric to multimedia-centric traffic. As wireless Aquisition DigitalImage Detection processing Extraction
FaceImage Normalized
speeds approach the wireline counterparts, the mission of 5G FaceImage
Feature
is different and much more complex, namely to support the Classification Database Vector
explosive evolution of ICT and Internet. In terms of func- Results
Classification
(TrainingSamples)
tions, 5G systems will support communications, computing,
control and content delivery (4C). In terms of applications,
MECServer
a wide-range of new applications and services for 5G are
emerging, such as real-time online gaming, virtual reality
(VR) and ultra-high-definition (UHD) video streaming, which Fig. 1. Main computation components in a face recognition application [17].
require unprecedented high access speed and low latency.
The past decade also saw the take-off of different visions of Resolution ImageFormat

next-generation Internet including IoT, Tactile Internet (with

Video
millisecond latency), Internet-of-Me, and social networks. In Renderer
Source
particular, it was predicted by Cisco that about 50 billion
IoT devices (e.g., sensors and wearable devices) will be
added to the Internet by 2020, most of which have limited
resources for computing, communication and storage, and Object
have to rely on Clouds or edge devices for enhancing their Recognizer
Tracker Mapper
capabilities [10]. It is now widely agreed that relying only
on Cloud Computing is inadequate to realize the ambitious NumberofFeatures MECServer
millisecond-scale latency for computing and communication
in 5G. Furthermore, the data exchange between end users and
remote Clouds will allow the data tsunami to saturate and Fig. 2. Main computation components in an AR application [18].
bring down the backhaul networks. This makes it essential to
supplement Cloud Computing with MEC that pushes traffic,
computing and network functions towards the network edges. network functions [15]. On the other hand, ICN provides an
This is also aligned with a key characteristic of next-generation alternative end-to-end service recognition paradigm for MEC,
networks that information is increasingly generated locally shifting from a host-centric to an information-centric one
and consumed locally, which arises from the booming of for implementing context-aware computing. Last, SDN allows
applications in IoT, social networks and content delivery [4]. MEC network administrators to manage services via function
The concept of MEC was firstly proposed by the Euro- abstraction, achieving scalable and dynamic computing [16].
pean Telecommunications Standard Institute (ETSI) in 2014, A main focus of MEC research is to develop these general
and was defined as a new platform that “provides IT and network technologies so that they can be implemented at the
cloud-computing capabilities within the Radio Access Network network edges.
(RAN) in close proximity to mobile subscribers” [5]. The There is an increasing number of emerging mobile ap-
original definition of MEC refers to the use of BSs for plications that will benefit from MEC, by offloading their
offloading computation tasks from mobile devices. Recently, computation-intensive tasks to the MEC servers for cloud
the concept of Fog Computing has been proposed by Cisco as a execution. In the following, we will provide two examples to
generalized form of MEC where the definition of edge devices illustrate the basic principles of MEC. One is the face recog-
gets broader, ranging from smartphones to set-top boxes [11]. nition application as shown in Fig. 1, which typically con-
This led to the emergence of a new research area called sists of five main computation components, including image
Fog Computing and Networking [4], [12], [13]. However, acquisition, face detection, pre-processing, feature extraction,
the areas of Fog Computing and MEC are overlapping and and classification [17]. While the image acquisition component
the terminologies are frequently used interchangeably. In this needs to be executed at the mobile device for supporting the
paper, we focus on MEC but many technologies discussed are user interface, the other components could be offloaded for
also applicable to Fog Computing. cloud processing, which contain complex computation such
MEC is implemented based on a virtualized platform that as signal processing and machine learning (ML) algorithms.
leverages recent advancements in network functions virtualiza- Another popular stream of applications that can leverage the
tion (NFV), information-centric networks (ICN) and software- rich resources at the network edges are augmented reality
defined networks (SDN). Specifically, NFV enables a single (AR) applications, which are able to combine the computer-
edge device to provide computing services to multiple mobile generated data with physical reality. AR applications as shown
devices by creating multiple virtual machines (VMs) 1 for si- in Fig. 2 have five critical components [18]–[20], namely,
multaneously performing different tasks or operating different the video source (which obtains raw video frames from the
1 The VM is a virtual computer mapped to the physical machine’s hard-
mobile camera), a tracker (which tracks the position of the
wares, providing virtual CPU, memory, hard drive, network interface, and user), a mapper (which builds a model of the environment),
other devices [14]. an object recognizer (which identifies known objects in the
 3

environment), and a renderer (which prepares the processed cooperate and perform sophisticated tasks such as surveillance,
frame for display). Among these components, the video source crowd-sensing and health monitoring [38]. Powering the tens
and renderer should be executed locally, while the most of billions of IoT devices remains a key challenge for de-
computation-intensive components, i.e., the tracker, mapper signing IoT given that frequent battery recharging/replacement
and object recognizer, can be offloaded for cloud execution. is impractical if not impossible. By effectively supporting
In this way, mobile users can enjoy various benefits from computation offloading, MEC stands out as a promising solu-
MEC such as latency reduction and energy savings, as will tion for prolonging battery lives of IoT devices. Specifically,
be elaborated in the next subsection. computation-intensive tasks can be offloaded from IoT devices
to edge devices so as to reduce their energy consumption.
B. Mobile Edge Computing Versus Mobile Cloud Computing Significant energy savings by computation offloading have
As shown in Table I, there exist significant disparities been demonstrated in experiments, e.g., the completion of up
between MEC and MCC systems in terms of computing server, to 44-time more computation load for a multimedia application
distance to end users and typical latency, etc. Compared with eyeDentify [39] or the increase of battery life by 30-50% for
MCC, MEC has the advantages of achieving lower latency, different AR applications [40].
saving energy for mobile devices, supporting context-aware Context-Awareness: Another key feature that differentiates
computing, and enhancing privacy and security for mobile MEC from MCC is the ability of an MEC server for lever-
applications. These advantages are briefly described through aging the proximity of edge devices to end users to track
some examples and applications in the following. their real-time information such as behaviors, locations, and
Low Latency: The latency for a mobile service is the environments. Inference based on such information allows the
aggregation of three components: propagation, computation, delivery of context-aware services to end users [41]–[43]. For
and communication latency, depending on the propagation dis- instance, the museum video guide, an AR application, can
tance, computation capacity, and data rate, respectively. First, predict users’ interests based on their locations in the museum
the information-propagation distances for MEC are typically to automatically deliver contents related to e.g., artworks and
tens-of-meters for the cases of dense small-cell networks or antiques [44]. Another example is the CTrack system that uses
device-to-device (D2D) transmissions, and typically no longer the BS fingerprints to track and predict the trajectories of a
than 1km for general cases. In contrast, Cloud Computing large number of users for the purposes of traffic monitoring,
requires transmissions from end users to nodes in core net- navigation and routing, and personalized trip management
works or data centers with distances ranging from tens of [45].
kilometers to that across continents. This results in much Privacy/Security Enhancement: The capability of enhanc-
shorter propagation delay for MEC than that for MCC. Second, ing the privacy and security of mobile applications is also
MCC requires the information to pass through several net- an attractive benefit brought by MEC compared to MCC.
works including the radio-access network, backhaul network In MCC systems, the Cloud Computing platforms are the
and Internet, where traffic control, routing and other network- remote public large data centers, such as the Amazon EC2
management operations can contribute to excessive delay. With and Microsoft Azure, which are susceptible to attacks due to
the communication constrained at the network edges, MEC their high concentration of information resources of users. In
is free from these issues. Last, for the computation latency, addition, the ownership and management of users’ data are
a Cloud has a massive computation power that is several separated in MCC, which shall cause the issues of private
orders of magnitude higher than that of an edge device (e.g., data leakage and loss [46]. The use of proximate edge servers
a BS). However, the Cloud has to be shared by a much larger provides a promising solution to circumvent these problems.
number of users than an edge device, reducing their gap in the On one hand, due to the distributed deployment, small-scale
computation latency. Furthermore, a modern BS is powerful nature, and the less concentration of valuable information,
enough for running highly sophisticated computing programs. MEC servers are much less likely to become the target of a
For instance, the edge cloud at a BS has 102 -104 times higher security attack. Second, many MEC servers could be private-
computation capability than the minimum requirement (e.g., owned cloudlets, which shall ease the concern of informa-
a CPU over 3.3GHz, 8GB RAM, 70GB storage space) for tion leakage. Applications that require sensitive information
running the Call-of-Duty 13, a popular shooter game2 . In exchange between end users and servers would benefit from
general, experiments have shown that the total latency for MEC. For instance, the enterprise deployment of MEC could
MCC is in the range of 30-100ms [31]. This is unacceptable help avoid uploading restricted data and material to remote
for many latency-critical mobile applications such as real- data centers, as the enterprise administrator itself manages the
time online gaming, virtual sports and autonomous driving, authorization, access control, and classifies different levels of
which may require tactile speed with latency approaching service requests without the need of an external unit [47].
1ms [37]. In contrast, with short propagation distances and
simple protocols, MEC has the potential of realizing tactile- C. Paper Motivation and Outline
level latency for latency-critical 5G applications.
Mobile Energy Savings: Due to their compact forms, MEC has emerged as a key enabling technology for realiz-
IoT devices have limited energy storage but are expected to ing the IoT and 5G visions [15], [48], [49]. MEC research lies
at the intersection of mobile computing and wireless commu-
2 https://www.callofduty.com/ nications, where the existence of many research opportunities
 4

TABLE I
C OMPARISON OF MEC AND MCC S YSTEMS .

MEC MCC
Small-scale data centers Large-scale data centers (each contains
Server hardware
with moderate resources [5], [21] a large number of highly-capable servers) [22], [23]
Co-locate with wireless gateways, Installed at dedicated buildings,
Server location
WiFi routers, and LTE BSs [5] with size of several football fields [24], [25]
Densely deployed by telecom operators, Deployed by IT companies, e.g., Google
MEC vendors, enterprises, and and Amazon, at a few locations
Deployment
home users. Require lightweight over the world. Require sophisticated
configuration and planning [5] configuration and planning [22]
Small Large
Distance to end users
(tens to hundreds of meters) [15] (may across the country border) [26]
Infrequent use Frequent use
Backhaul usage
Alleviate congestion [27] Likely to cause congestion [27]
Hierarchical control Centralized control [28]
System management
(centralized/distributed) [28]
Supportable latency Less than tens of milliseconds [15], [29] Larger than 100 milliseconds [30], [31]
Latency-critical and computation-intensive Latency-tolerant and computation-intensive
Applications applications, e.g., AR, automatic driving, applications, e.g., online social networking,
and interactive online gaming [5], [32]. and mobile commerce/health/learning [33]–[36].

has resulted in a highly active area. In recent years, researchers still lacks a systematic survey article providing comprehensive
from both academia and industry have investigated a wide- and concrete discussions on specific MEC research results
range of issues related to MEC, including system and network with a deep integration of mobile computing and wireless
modeling, optimal control, multiuser resource allocation, im- communications, which motivates the current work. This paper
plementation and standardization. Subsequently, several survey differs from existing surveys on MEC in the following aspects.
articles have been published to provide overviews of the First, the current survey summarizes existing models of com-
MEC area with different focuses, including system models, puting and communications in MEC to facilitate theoretical
architectures, enabling techniques, applications, edge caching, analysis and provide a quick reference for both researchers
edge computation offloading, and connections with IoT and 5G and practitioners. Next, we present a comprehensive literature
[27], [28], [50]–[56]. Their themes are summarized as follows. review on joint radio-and-computational resource allocation
An overview of MEC platforms is presented in [50] where for MEC, which is the central theme of the current paper. The
different existing MEC frameworks, architectures, and their literature review in our paper shall be a valuable addition to the
application scenarios, including FemtoClouds, REPLISM, and existing survey literature on MEC, which can benefit readers
ME-VOLTE, are discussed. The survey of [51] focuses on from the research community in building up a systematic
the enabling techniques in MEC such as cloud computing, understanding of the state-of-the-art resource management
VM, NFV, SDN that allow the flexible control and multi- techniques for MEC systems. Furthermore, we identify and
tenancy support. In [52], the authors categorize diverse MEC discuss several research challenges and opportunities in MEC
applications, service models, deployment scenarios, as well from the communication perspective, for which potential so-
as network architectures. The survey in [53] presents a tax- lutions are elaborated. In addition, to bridge the gap between
onomy for MEC applications and identifies potential direc- theoretical research and real implementation of MEC, recent
tions for research and development, such as content scaling, standardization efforts and use scenarios of MEC will then be
local connectivity, augmentation, and data aggregation and introduced.
analytics. In [28], emerging techniques of edge computing, This paper is organized as follows. In Section II, we summa-
caching, and communications (3C) in MEC are surveyed, rize the basic MEC models, comprising models of computation
showing the convergence of 3C. Besides, key enablers of MEC tasks, communications, mobile devices and MEC servers,
such as cloud technology, SDN/NFV, and smart devices are based on which the models of MEC latency and energy con-
also discussed. The survey in [54] focuses on three critical sumption are developed. Next, a comprehensive review is pre-
design problems in computation offloading for MEC, namely, sented in Section III, focusing on the research of joint radio-
the offloading decision, computation resource allocation, and and-computational resource management for different types
mobility management. In addition, the role of MEC in IoT, of MEC systems, including single-user, multiuser systems as
i.e., creating new IoT services, is highlighted in [55] through well as multi-server MEC. Subsequently, a set of key research
MEC deployment examples with reference to IoT use cases. issues and future directions are discussed in Section III-D3
Several attractive use scenarios of MEC in 5G networks including 1) deployment of MEC systems, 2) cache-enabled
are also introduced in [27], ranging from mobile-edge or- MEC, 3) mobility management for MEC, 4) green MEC,
chestration, collaborative caching and processing, and multi- and 5) security-and-privacy issues in MEC. Specifically, we
layer interference cancellation. Furthermore, potential business analyze the design challenges for each research problem
opportunities related to MEC are discussed in [56] from the and provide several potential research approaches. Last, the
perspectives of application developers, service providers, and MEC standardization efforts and applications are reviewed and
network equipment vendors. In view of prior work, there discussed in Section V, followed by concluding remarks in
 5

TABLE II
S UMMARY OF I MPORTANT ACRONYMS .

Acronym Definition Acronym Definition

AF application function MEC mobile edge computing
AR augmented reality ML machine learning
AP access point mMTC massive machine type communication
BS base station NEF network exposure function
CAPEX capital expenditure NFC near-filed communications
C-RAN cloud radio access network NFV network functions virtualization
CSI channel-state information OFDMA orthogonal frequency division multiple access
DAG directed acyclic graph PCF policy control function
DCN data-center network PMR peak-to-mean ratio
DNS domain name system PoC proof of concept
DP dynamic programming QoS quality of service
DPP determinantal point process RAM random access memory
DVFS dynamic frequency and voltage scaling RAN radio access network
D2D device-to-device RFID radio frequency identification
EH energy harvesting RNIS radio network information services
eMBB enhanced mobile broadband SDN software-defined networks
ESI energy side information SINR signal-to-interference-plus-noise ratio
ETSI European Telecommunications Standard Institute TOF traffic offloading function
GLB geographical load balancing UE user equipment
Het-MEC heterogeneous MEC UHD ultra-high-definition
HetNets heterogeneous networks UPF user plane function
HPPP homogeneous Poisson point process UPS uninterrupted power supply
IaaS Infrastructure as a Service URLLC ultra-reliable and low latency communication
ICN information-centric networks VM virtual machine
ISG industry specification group VR virtual reality
ISI inter-symbol interference V2X vehicular-to-everything
IoT Internet of Things WPT wireless power transfer
KKT Karush-Kuhn-Tucker 3C computing, caching, and communications
LP linear programming 3GPP 3rd Generation Partnership Project
LTE long-term evolution 4C communications, computing, control and content delivery
MCC mobile cloud computing 5GPPP European 5G infrastructure Public Private Partnership
MDP Markov decision process 5QI 5G QoS Indicator

Section VI. We summarze the definitions of the acronyms that A. Computation Task Models
will be frequently use in this paper in TABLE II for ease of
There are various parameters that play critical roles in
reference.
modeling the computation tasks, including latency, bandwidth
utilization, context awareness, generality, and scalability [57].
II. MEC C OMPUTATION AND C OMMUNICATION M ODELS Though it is highly sophisticated to develop accurate models
for tasks, there exist simple ones that are reasonable and allow
In this section, system models are introduced for the key mathematical tractability. In this subsection, we introduce two
computation/communication components of the typical MEC computation-task models popularly used in existing literature
system. The models provide mechanisms for abstracting var- on MCC and MEC, corresponding to binary and partial
ious functions and operations into optimization problems and computation offloading, respectively.
facilitating theoretical analysis as discussed in the following 1) Task Model for Binary Offloading: A highly integrated
sections. or relatively simple task cannot be partitioned and has to be
For the MEC system shown in Fig. 3, the key components executed as a whole either locally at the mobile device or
include mobile devices (a.k.a. end users, clients, service sub- offloaded to the MEC server, called binary offloading. Such a
scribers) and MEC servers. The MEC servers are typically task can be represented by a three-field notation A (L, τd , X).
small-scale data centers deployed by the cloud and telecom This commonly-used notation contains the information of
operators in close proximity with end users and can be co- the task input-data size L (in bits), the completion deadline
located with wireless APs. Through a gateway, the servers τd (in second), and the computation workload/intensity X
are connected to the data centers via Internet. Mobile devices (in CPU cycles per bit). These parameters are related to
and servers are separated by the air interface where reliable the nature of the applications and can be estimated through
wireless links can be established using advanced wireless task profilers [58], [59]. The use of these three parameters
communication and networking technologies. In the following not only captures essential properties of mobile applications
subsections, we will introduce the models for different compo- such as the computation and communication demands, but
nents of MEC systems, including models for the computation also facilitates simple evaluation of the execution latency and
tasks, wireless communication channels and networks, as well energy consumption performance (which will be analyzed in
as the computation latency and energy consumption models of Section II-C).
mobile devices and MEC servers. The task A (L, τd , X) is required to be completed before
 6

Gaming

ARApps.

3DModeling

Socialnetworking

Smartdevicesapplications
Connectedvehicles
MECServer

'DWDFHQWHU
$63V
*DWHZD\ ,QWHUQHW
EDFNERQH
0RELOHFRUHQHWZRUN &'1V

DP DP

DP DP

Healthmonitoring Surveillancenetworks

Fig. 3. Architecture of the MEC systems.

a hard deadline τd . This model can also be generalized to • Second, due to either software or hardware constraints,
handle the soft deadline requirement which allows a small some functions or routines can be offloaded to the server
portion of tasks to be completed after τd [60]. In this case, for remote execution, while the ones can only be executed
the number of CPU cycles needed to execute 1-bit of task locally such as the image display function.
input data is modeled as a random variable X. Specifically,
define x0 as a positive integer such that Pr(X > x0 ) ≤ ρ
where ρ is a small real number: 0 < ρ  1. It follows that
Pr(LX > Wρ ) ≤ ρ where Wρ = Lx0 . Then given the L-bit This calls for task models that are more sophisticated
task-input data, Wρ upper bounds the number of required CPU than the mentioned data-partition model that can capture
cycles almost surely. the inter-dependency among different computation functions
2) Task Models for Partial Offloading: In practice, and routines in an application. One such model is called the
many mobile applications are composed of multiple proce- task-call graph. The graph is typically a directed acyclic
dures/components (e.g., the computation components in an graph (DAG), which is a finite directed graph with no
AR application as shown in Fig. 2), making it possible directed cycles. We shall denote it as G (V, E), where the set
to implement fine-grained (partial) computation offloading. of vertices V represents different procedures in the application
Specifically, the program can be partitioned into two parts with and the set of edges E specifies their call dependencies. There
one executed at the mobile device and the other offloaded for are three typical dependency models of sub-tasks (i.e., task
edge execution. components such as functions or routines), namely sequential,
parallel, and general dependency [61], [62], as illustrated
The simplest task model for partial offloading is the data-
in Fig. 4. For the mobile initiated applications, the first and
partition model, where the task-input bits are bit-wise indepen-
the last steps, e.g., collecting the I/O data and displaying
dent and can be arbitrarily divided into different groups and
the computation results on the screen, are normally required
executed by different entities in MEC systems, e.g., parallel
to be executed locally. Thus, node 1 and node N in Fig.
execution at the mobiles and MEC server.
4(a)-4(c) are components that must be executed locally.
Nevertheless, the dependency among different proce- Besides, the required computation workloads and resources
dures/components in many applications cannot be ignored of each procedure, e.g., the number of required CPU cycles
as it significantly affects the procedure of execution and and the amount of needed memory, can also be specified
computation offloading due to the following reasons: in the vertices of the task-call graph, while the amount of
• First, the execution order of functions or routines cannot input/output data of each procedure can be characterized by
be arbitrarily chosen because the outputs of some com- imposing weights on the edges.
ponents are the inputs of others.
 7

F
F F
 w w
F   
w F FN
 F
w « ͼ
w wNN F  N wNN
« ͼ w wN ͼ
   N N   N 
w

w
« wNN-1 N N
F F F FN FN F FN F FN FN
F

«
w
FN 
ͼ
N F

(a) Sequential dependency (b) Parallel dependency (c) General dependency

Fig. 4. Typical topologies of the task-call graphs.

B. Communication Models quency and space make it important for designing efficient
In the literature of MCC, communication channels between MEC systems to seamlessly integrate control of computation
the mobile devices and cloud servers are typically abstracted as offloading and radio resource management. For instance, when
bit pipes with either constant rates or random rates with given the wireless channel is in deep fade, the reduction on execution
distributions. Such coarse models are adopted for tractability latency by remote execution may not be sufficient to compen-
and may be reasonable for the design of MCC systems sate for the increase of transmission latency due to the steep
where the focuses are to tackle the latency in the core drop in transmission-data rates. For such cases, it is desirable
networks and management of large-scale cloud but not the to defer offloading till the channel gain is favorable or switch
wireless-communication latency. The scenario is different for to an alternative frequency/spatial channel with a better quality
MEC systems. Given small-scale edge clouds and targeting for offloading. Furthermore, increasing transmission power can
latency-critical applications, reducing communication latency increase the data rate, but also lead to a larger transmission
by designing a highly efficient air interface is the main energy consumption. The above considerations necessitate the
design focus. Consequently, the mentioned bit-pipe models joint design of offloading and wireless transmissions, which
are insufficient as they overlook some fundamental properties should be adaptive to the time-varying channels based on the
of wireless propagation and are too simplified to allow the accurate channel-state information (CSI).
implementation of advanced communication techniques. To be In MEC systems, communications are typically between
specific, wireless channels differ from the wired counterparts APs and mobile devices with the possibility of direct D2D
in the following key aspects [63]: communications. The MEC servers are small-scale data cen-
1) Due to atmospheric ducting, reflection and refraction ters deployed by the Cloud Computing/telecom operators,
from scattering objects in the environment (e.g., build- which can be co-located with the wireless APs, e.g., the
ings, walls and trees), there exists the well-known multi- public WiFi routers and BSs, as so to reduce the capital
path fading in wireless channels, making the channels expenditure (CAPEX) (e.g., site rental). As shown in Fig. 3,
highly time-varying and can cause severe inter-symbol the wireless APs not only provide the wireless interface for
inference (ISI). Thus, effective ISI suppression tech- the MEC servers, but also enable the access to the remote
niques, such as equalization and spread spectrum, are data center through backhaul links, which could help the MEC
needed for reliable transmissions. server to further offload some computation tasks to other MEC
2) The broadcast nature of wireless transmissions results servers or to large-scale cloud data centers. For the mobile
in a signal being interfered by other signals occupy- devices that cannot communicate with MEC servers directly
ing the same spectrum, which reduces their respective due to insufficient wireless interfaces, D2D communications
receive signal-to-interference-plus-noise ratios (SINRs) with neighboring devices provide the opportunity to forward
and thereby results in the probabilities of error in detec- the computation tasks to MEC servers. Furthermore, D2D
tion. To cope with the performance degradation, inter- communications also enable the peer-to-peer cooperation on
ference management becomes one of the most important resource sharing and computation-load balancing within a
design issues for wireless communication systems and cluster of mobile devices.
has attracted extensive research efforts [64]–[66]. Presently, there exist different types of commercialized
3) Spectrum shortage has been the main foe for very technologies for mobile communications, including the near-
high-rate radio access, motivating extensive research on filed communications (NFC), radio frequency identification
exploiting new spectrum resources [67], [68], designing (RFID), Bluetooth, WiFi, and cellular technologies such as the
novel transceiver architectures [69]–[71] and network long-term evolution (LTE). Besides, the 5G network, which
paradigms [72], [73] to improve the spectrum efficiency, will be realized by the development of LTE in combination
as well as developing spectrum sharing and aggregation with new radio-access technologies, is currently being stan-
techniques to facilitate efficient use of fragmented and dardized and will be put into commercial use as early as
underutilized spectrum resources [74]–[76]. 2020 [77]. These technologies can support wireless offloading
The random variations of wireless channels in time, fre- from mobiles to APs or peer-to-peer mobile cooperation for
 8

TABLE III
C HARACTERISTICS OF T YPICAL W IRELESS C OMMUNICATION T ECHNOLOGIES .

NFC RFID Bluetooth WiFi LTE 5G

Max. Coverage 10cm 3m 100m 100m up to 5km Excellent coverage
LF: 120-134kHz
Operation Freq. 13.56MHz HF: 13.56MHz 2.4GHz 2.4GHz, 5GHz TDD: 1.85-3.8GHz 6-100GHz
UHF: 850-960MHz FDD: 0.7-2.6GHz
Indoor/dense outdoor:
Data Rate
106, 212, Low (LF) to 135Mbps DL: 300Mbps up to 10Gbps
414kbps high (UHF) 22Mbps (IEEE 802.11n) UL: 75Mbps Urban/suburban:
> hundreds of Mbps

varying data rates and transmission ranges. We list the key the circuit theory [78]–[81], the CPU power consumption can
characteristics of typical wireless communication technologies be divided into several factors including the dynamic, short-
in Table III, which differ significantly in terms of the operation circuit, and leakage power consumption3 , where the dynamic
frequency, maximum coverage range, and data rate. For NFC, power consumption dominates the others. In particular, it
the coverage range and data rate are very low and thus is shown in [80] that the dynamic power consumption is
2
the technology is suitable for applications that require little proportional to the product of Vcir fm where Vcir is the circuit
information exchange, e.g., e-payment and physical access supplied voltage. It is further noticed in [78], [81] that, the
authentication. RFID is similar to NFC, but only allows one- clock frequency of the CPU chip is approximately linear
way communications. Bluetooth is a more powerful tech- proportional to the voltage supply when operating at the low
nique to enable short-range D2D communications in MEC voltage limits. Thus, the energy consumption of a CPU cycle
2
systems. For long-range communications between mobiles and is given by κfm , where κ is a constant related to the hardware
MEC servers, WiFi and LTE (or 5G in the future) are two architecture. For the computation task A (L, τ, X) with CPU
primary technologies enabling the access to MEC systems, clock speed fm , the energy consumption can be derived:
which can be adaptively switched depending on their link 2
Em = κLXfm . (2)
reliability. For the deployment of wireless technologies in
MEC systems, the communication and networking protocols One can observe from (1) and (2) that the mobile device may
need to be redesigned to integrate both the computing and not be able to complete a computation-intensive task within
communication infrastructures, and effectively improve the the required deadline, or else the energy consumption incurred
computation efficiency that is more sophisticated than the data by mobile execution is so high that the onboard battery will be
transmission. depleted quickly. In such cases, offloading the task execution
process to an MEC server is desirable.
C. Computation Models of Mobile Devices Besides CPUs, other hardware components in the mobile
In this subsection, we introduce the computation models of devices, e.g., the random access memory (RAM) and flash
mobile devices and discuss methodologies of evaluating the memory, also contribute to the computation latency and
computation performance. energy consumption [82], while detailed discussions are
The CPU of a mobile device is the primary engine for beyond the scope of this survey.
local computation. The CPU performance is controlled by
the CPU-cycle frequency fm (also known as the CPU clock
speed). The state-of-the-art mobile CPU architecture adopts
D. Computation Models of MEC Servers
the advanced dynamic frequency and voltage scaling (DVFS)
technique, which allows stepping-up or -down of the CPU- In this subsection, we introduce the computation models
cycle frequency (or voltage), resulting in growing and reducing of the MEC servers. Similar as the mobile devices, the
energy consumption, respectively. In practice, the value of computation latency and energy consumption are of particular
max interests.
fm is bounded by a maximum value, fCPU , which reflects
the limitation of the mobile’s computation capability. Based The server-computation latency is negligible compared with
on the computation task model introduced in Section II-A, communication or local-computation latency in MEC systems
the execution latency for task A (L, τ, X) can be calculated where the computation loads for servers are much lower than
accordingly to their computation capacities [81], [83]. This model can be also
LX relevant for multiuser MEC systems with resource-constrained
tm = , (1)
fm
3 The dynamic power consumption comes from the toggling activities of the
which indicates that a high CPU clock speed is desirable in
logic gates inside a CPU, which shall charge/discharge the capacitors inside
order to reduce the execution latency, at the cost of higher the logic gates. When a logic gate toggles, some of its transistors may change
CPU energy consumption. states, and thus, there might be a short period of time when some transistors
As the mobile devices are energy-constrained, the energy are conducting simultaneously. In this case, the direct path between the source
and ground will result in some short-circuit power loss. The leakage power
consumption for local computation is another critical mea- dissipation is due to the flowing current between doped parts of the transistors
surement for the mobile computing efficiency. According to [80], available on https://en.wikipedia.org/wiki/CPU power dissipation.
 9

servers if the servers’ computation loads are regulated by mul- of MEC servers. One model is based on the DVFS technique
tiuser resource management under latency and computation- described as follows. Consider an MEC server that handles K
capacity constraints [84]. computation tasks and the k-th task is allocated with wk CPU
On the other hand, as edge servers have relatively limited cycles with CPU-cycle frequency fs,k . Hence, the total energy
computation resources, it is necessary to consider the non- consumed by the CPU at the MEC server, denoted by Es , can
negligible server execution time in the general design of be expressed as
MEC systems, yielding the computation model for the severs K
X
discussed in the remainder of this subsection. Two possible 2
Es = κwk fs,k , (4)
models are considered in the literature, corresponding to the k=1
deterministic and stochastic server-computation latency. The which is similar to that for the mobile devices. The other model
deterministic model is proposed to consider the exact server- is based on an observation in recent works [91]–[93] that the
computation latency for latency-sensitive applications, which server-energy consumption is linear to the CPU utilization
is implemented using techniques such as VMs and DVFS. ratio which depends on the computation load. Moreover, even
Specifically, assume the MEC server allocates different VMs for an idle server, it still, on average, consumes up to 70%
for different mobile devices, allowing independent computa- of the energy consumption for the case with the full CPU
tion [85]. Let fs,k denote the allocated servers’ CPU-cycle speed. Thus, the energy consumption at the MEC server can
number for mobile device k. Similar to Section II-C, it be calculated according to
follows that the server execution time denoted by ts,k can
wk Es = αEmax + (1 − α)Emax u, (5)
be calculated as ts,k = , where wk is the number of
fs,k
required CPU cycles for processing the offloaded computation where Emax is the energy consumption for a fully-utilized
workload. This model has been widely used for designing server, α is the fraction of the idle energy consumption (e.g.,
computation-resource allocation policies [86]–[88]. A similar 70%) and u denotes the CPU utilization ratio. This model
model was proposed in [84], where the MEC server is assumed suggests that energy-efficient MEC should allow servers to be
to perform load balancing for the total offloaded computation switched into the sleep mode in the case of light load and
workloads. In other words, the CPU cycles at the MEC server consolidation of computation loads into fewer active servers.
are proportionally allocated to each mobile device such that
they experience the same execution latency. Furthermore, in E. Summary and Insights
addition to the CPU processing time, the server scheduling The MEC computation and communication models are
queuing delay should be accounted for MEC servers with summarized in Fig. 5, laying the foundation for the analysis of
relatively small computation capacities, where parallel com- MEC resource management in the next section. These models
puting via virtualization techniques is not feasible and thus shed several useful insights on the offloading design, listed as
it needs to process the computation workloads sequentially. follows.
Without loss of generality, denote k as the processing order for • The effective design of MEC should leverage and inte-
a mobile device and name it as mobile k. Thus, the total server- grate advanced techniques from both areas of wireless
computation latency including the queuing delay for device k communications and mobile computing.
denoted by Ts,k can be given as • It is vital to choose suitable computation task models

Ts,k =
X
ts,i . (3) for different MEC applications. For example, the soft-
i≤k
deadline task model can be applied for social network-
ing applications but is not suitable for AR applications
For latency-tolerant applications, the average server- due to the stringent computation latency requirements.
computation time can be derived based on stochastic models. Moreover, for a specific application, the task model also
For example, in [89], the task arrivals and service time depends on the offloading scenario, e.g., the data-partition
are modeled by the Poisson and exponential processes, model can be used when the input-data is offloaded, and
respectively. Thus, the average server-computation time can the task-call graph should be considered when each task
be derived using techniques from queuing theory. Last, component can be offloaded as a whole.
for all above models, as investigated in [1], multiple VMs • The wireless channel condition significantly affects the
sharing the same physical machine will introduce the I/O amount of energy consumption for computation offload-
interference among different VMs. It results in the longer ing. MEC has the potential to reduce the transmission
0
computation latency for each VM denoted by Ts,k , which energy consumption due to short distances between users
0
can be modeled by Ts,k = Ts,k (1 + )n where  is the and MEC servers. Advanced wireless communication
performance degradation factor as the percentage increasing techniques, such as interference cancelation and adaptive
of the latency [90]. power control, can further reduce the offloading energy
The energy consumption of an MEC server is jointly consumption.
determined by the usage of the CPU, storage, memory, and • Dynamic CPU-cycle frequency control is the key tech-
network interfaces. Since the CPU contribution is dominant nique for controlling the computation latency and en-
among these factors, it is the main focus in the literature. Two ergy consumption for both mobile devices and MEC
tractable models are widely used for the energy consumption servers. Specifically, increasing the CPU-cycle frequency
 10

Fig. 5. Summary of MEC models.

can reduce the computing time but contributes to higher considerations and basic design methodologies. Subsequently,
energy consumption. The effective CPU-cycle frequency more complex multiuser MEC systems are considered where
control should approach the optimal tradeoff between multiple offloading users compete for the use of both the radio
computation latency and energy consumption. and server-computation resources and have been coordinated.
• Apart from the task-execution latency, the computation Last, we extend the discussion to MEC systems with het-
scheduling delay is non-negligible if the MEC server has erogeneous servers which not only provide the freedom of
a relatively small computation capacity or heavy compu- server selection but also allow the cooperation among servers.
tation loads are offloaded to the server. Load-balancing Such network-level operations can significantly enhance the
and intelligent scheduling policies can be designed to performance of MEC systems.
reduce the total computation latency.
A. Single-User MEC Systems
III. R ESOURCE M ANAGEMENT IN MEC S YSTEMS This subsection focuses on the simple single-user MEC sys-
The joint radio-and-computational resource management tems and reviews a set of recent research efforts for this case.
plays a pivotal role in realizing energy-efficient and low- The discussion is divided according to three popularly-used
latency MEC. The implementation of relevant techniques is task models, namely, deterministic task model with binary
facilitated by the network architecture where MEC servers offloading, deterministic task model with partial offloading,
and wireless APs (e.g., BSs and WiFi routers) are co-located. and stochastic task model.
In this section, we provide a comprehensive overview of 1) Deterministic Task Model with Binary Offloading:
the literature on resource management for MEC systems Consider the mentioned single-user MEC system where the
summarized in Fig. 6. Our discussion starts from the simple binary offloading decision is on whether a particular task
single-user systems comprising a single mobile device and a should be offloaded for edge execution or local computation.
single MEC server, allowing the exposition of the key design The investigations for the optimal offloading policies can be
 11

1. Deterministic Task Model with Binary Offloading

Single-User MEC
Systems 2. Deterministic Task Model with Partial Offloading

3. Stochastic Task Model

1. Joint Radio-and-Computational Resource Allocation

Resource Management
Multiuser MEC Systems 2. MEC Server Scheduling
in MEC Systems
3. Multiuser Cooperative Edge Computing

1. Server Selection
MEC Systems with
Heterogeneous Servers 2. Server Cooperation

3. Computation Migration

Fig. 6. Classification of resource management techniques for MEC.

dated back to those for conventional Cloud Computing sys- power-rate function. The optimization problem is convex and
tems, where the communication links were typically assumed can be solved in closed form. In particular, task offloading
to have a fixed rate B. In [94] and [95], general guidelines is desirable when the channel power gain is greater than a
are developed for determining the offloading decision for the threshold and the server CPU is fast enough, which reveals
purposes of minimizing the mobile-energy consumption and the effects of wireless channels on the offloading decision.
computation latency. Denote w as the amount of computation A further study was conducted by Zhang et al. in [81] to
(in CPU cycles) for a task, fm as the CPU speed of the mobile minimize the energy consumption for executing a task with
device, d as the input data size, and fs as the CPU speed at the a soft real-time requirement, targeting e.g., multimedia appli-
cloud server. Offloading the computation to the cloud server cations, which requires the task to be completed within the
can improve the latency performance only when deadline with a given probability ρ. The offloading decision
w d w was determined by the computation mode (either offloading
> + , (6) or local computing) that incurs less energy consumption. On
fm B fs
one hand, the energy consumption for local execution was op-
which holds for applications that require heavy computation timized using the DVFS technique, which was formulated as a
and have small amount of data input, or when the cloud server convex optimization problem with the objective function being
is fast, and the transmission rate is sufficiently high. Moreover, the expected energy consumption of the Wρ CPU cycles and
let pm represent the CPU power consumption at the mobile a time duration constraint for these CPU cycles. The optimal
device, and pt as the transmission power, pi as the power CPU-cycle frequencies over the computation duration were
consumption at the device when the task is running at the derived in closed form by solving the Karush-Kuhn-Tucker
server. Offloading the task could help save mobile energy when (KKT) conditions, suggesting that the processor should speed
w d w up as the number of completed CPU cycles increases. On the
pm × > pt × + pi × (7) other hand, the expected energy consumption for task offload-
fm B fs
ing was minimized via data transmission scheduling. Under the
holds, i.e., applications with heavy computation and light Gilbert-Elliott channel model, the optimal data transmission
communication should be offloaded. scheduling was obtained through dynamic programming (DP)
Nevertheless, the data rates for wireless communications are techniques, and the scaling law of the minimum expected
not constant and change with the time-varying channel gains energy consumption with respect to the execution deadline was
as well as depend on the transmission power. This calls for also derived. This framework was further developed in [83]
the design of control policies for power adaptation and data where both the local computing and offloading are powered
scheduling to streamline the offloading process. In addition, as by wireless energy transfer. Specifically, the optimal CPU-
the CPU power consumption increases super-linearly with the cycle frequencies for local computing and time division for
CPU-cycle frequency, the computation energy consumption for offloading should be adaptive to the transferred power.
mobile execution can be minimized using DVFS techniques. 2) Deterministic Task Model with Partial Offloading: The
These issues led to the active field of adaptive MEC as running of a relatively sophisticated mobile application can be
summarized below. decomposed into a set of smaller sub-tasks. Inspired by recent
In [96], the problem of transmission-energy minimization advancements of parallel computing, partial offloading (also
under a computation-deadline constraint was formulated with known as program partitioning) schemes were proposed to
the optimization variable being the input-data transmission further optimize MEC performance in [61], [62], [97]–[102].
time, where the famous Shannon-Hartley formula gives the In [97], full granularity in program partitioning was con-
 12

sidered where the task-input data can be arbitrarily divided controls the states of the local processing and transmission
for local and remote executions. Joint optimization of the units and the task buffer queue length based on the channel
offloading ratio, transmission power and CPU-cycle frequency state. It was shown that the optimal task-scheduling policy
was performed to minimize the mobile-energy consumption significantly outperforms the greedy scheduling policy (i.e.,
(or latency) subject to a latency (or energy consumption) tasks are scheduled to the local CPU/transmission unit when-
constraint. Both the energy and latency minimization problems ever they are idle). To jointly optimize the computation latency
are non-convex in contrast to the ones for binary-offloading. and energy consumption, the problem of minimizing the long-
The former problem can be solved optimally with a variable- term average execution cost was considered in [102] and [106],
substitution technique while a sub-optimal algorithm was where the former only optimized the offloading data size
proposed for the latter one in [97]. based on the MDP theory while the latter jointly controlled
In [61], [62], [98]–[102], applications were modeled by the local CPU frequency, modulation scheme as well as data
task-call graphs discussed earlier that specify the dependency rates under a semi-MDP framework. In [107], the energy-
among different sub-tasks, and the code partitioning schemes latency tradeoff in MEC systems with heterogeneous types of
designed to dynamically generate the optimal set of tasks for applications was investigated, including the non-offloadable
offloading. In [61], by leveraging the concept of load balancing workload, cloud-offloadable workload and network traffic.
between the mobile device and the server, a heuristic program- A Lyapunov optimization-based algorithm was proposed to
partitioning algorithm was developed to minimize the execu- jointly decide the offloading policy, task allocation, CPU clock
tion latency. Kao et al. investigated the latency minimization speed, and selected network interface. It was also shown
problem with a prescribed resource utilization constraint in that the energy consumption decreases inversely proportional
[98], and proposed a polynomial-time approximate solution to V while the latency increases linearly with V , where V
with guaranteed performance. To maximize the energy savings is a control parameter in the proposed algorithm. Similar
achieved by computation offloading, the scheduling and cloud investigation was conducted for MEC systems with a multi-
offloading decisions were jointly optimized using an integer core mobile device in [108].
programming approach in [62]. In [99], considering the wire- 4) Summary and Insight: The comparison of resource
less channel models including the block fading channel, in- management schemes for single-user MEC systems is shown
dependent and identical distributed (i.i.d.) stochastic channel, in Table IV. This series of work yields a number of useful
and the Markovian stochastic channel, the expected energy insights on controlling computation offloading as summarized
consumption minimization problem with a completion time below.
constraint was found to be a stochastic shortest-path problem, • Consider binary offloading. For energy savings, com-
and the one-climb policies (i.e., the execution only migrates putation offloading is preferred to local computation
once from the mobile device to the server) were shown to be when the user has desirable channel condition or small
optimal. In addition, the program-partitioning schemes were local computation capability. Moreover, beamforming and
also optimized together with the physical layer parameters, MIMO techniques can be exploited to reduce the energy
such as the transmission and reception power, constellation consumption for offloading. For latency reduction, com-
size, as well as the data allocation for different radio interfaces putation offloading is advantageous over local computa-
[100]–[102]. tion when the user has a large bandwidth and the MEC
3) Stochastic Task Model: Resource management policies server is provisioned with huge computation capacity.
have been also developed for MEC systems with stochastic • Partial offloading allows flexible components/data par-
task models characterized by random task arrivals, where the titioning. By offloading time-consuming or energy-
arrived but not yet executed tasks join the queues in buffers consuming sub-tasks to MEC servers, partial offloading
[103]–[108]. For such systems, the long-term performance, can achieve larger energy savings and smaller compu-
e.g., the long-term average energy consumption and execution tation latency compared with binary offloading. Graph
latency, are more relevant compared with those of determin- theory is a powerful tool for designing the offloading
istic task arrivals, and the temporal correlation of the optimal scheduling according to the task dependency graph.
system operations makes the design more challenging. As a • For stochastic task models, the temporal correlation of
result, the design of MEC systems with random task arrivals task arrivals and channels can be exploited to design
is an area less explored compared with the simpler cases adaptive dynamic computation offloading policies. More-
with deterministic task models. In [103], in order to minimize over, it is critical to maintain the task buffer stability at
the mobile-energy consumption while keeping the proportion the user and MEC server via offloading rate control.
of executions violating the deadline requirement below a
threshold, a dynamic offloading algorithm was proposed to
determine the offloaded software components from an applica- B. Multiuser MEC Systems
tion running at a mobile user based on Lyapunov optimization While the preceding subsection aims at resource manage-
techniques, where 3G and WiFi networks are accessible to the ment policies for single-user MEC systems with a dedicated
device but their rates vary at different locations. Assuming that MEC server, this subsection considers the multiuser MEC
concurrent local and edge executions are feasible, the latency- systems comprising multiple mobile devices that share one
optimal task scheduling policies were designed in [104] based edge server. Several new challenges are investigated in the
on the theory of Markov decision process (MDP), which sequel, including the multiuser joint radio-and-computational
 13

TABLE IV
T HE COMPARISON OF PAPERS FOCUSING ON SINGLE - USER MEC SYSTEMS .

Task model Design Objective Reference Proposed Solution

[81] Optimize local computing and offloading by controlling the CPU
frequency and transmission rate
[83] Propose a novel framework of wirelessly powered MEC and optimize
Energy
both local computing and offloading
Binary Offloading
[94] Propose general guidelines to make offloading decision for energy
consumption minimization
[96] Propose the optimal binary computation offloading decision using
convex optimization
[95] Propose general guidelines to make offloading decision for energy-
Energy and latency
consumption and computation-latency minimization
[62] Propose a joint scheduling and computation offloading algorithm by
parallel processing appropriate components in the mobile and cloud
[99] Formulate a stochastic shortest-path problem and derive the one-climb
optimal policy
[101] Jointly optimize the program partitioning with the selection of transmit
Energy power and constellation size
[102] Propose an iterative algorithm for the optimal offloading scheduling as
Partial Offloading
well as the percentage of the data to be carried on each radio interface
[61] Propose a heuristic load-balancing program-partitioning algorithm
Latency
[98] Propose a polynomial-time approximate solution with guaranteed
performance
[97] Jointly optimize the offloading ratio, transmission power and CPU-
Energy and latency cycle frequency using variable-substitution technique
[100] Propose an algorithmic to leverage the structure of the call graphs by
means of message passing under both serial and parallel implementa-
tions of processing and communication
[103] Propose a Lyapunov optimization-based dynamic computation offload-
Energy
ing policy
[104] Dynamically control the local processing and transmission using MDP
[105] Optimize local computing and transmission using semi-MDP and
Latency
propose a one-dimensional heuristic search algorithm
[106] Jointly control the local CPU frequency, modulation scheme as well
Stochastic Model
as the data rates under a semi-MDP framework
[107] Propose a Lyapunov optimization-based algorithm to decide the of-
floading policy, task allocation, CPU clock speed, and selected network
Energy and Latency
interface
[108] Propose a Lyapunov optimization-based scheme for cloud offloading
scheduling, as well as download scheduling for cloud execution output

resource allocation, MEC server scheduling, and multiuser with priorities above and below a given threshold will perform
cooperative edge computing. full and minimum offloading (so as to meet a given compu-
1) Joint Radio-and-Computational Resource Allocation: tation deadline), respectively. This result was also extended
Compared with the central cloud, the MEC servers have much to the OFDMA-based MEC systems for designing a close-
less computational resources. Therefore, one key issue in to-optimal computation offloading policy. In [86], instead
designing a multiuser MEC system is how to allocate the of controlling the offloading data size and time, the MEC
finite radio-and-computational resources to multiple mobiles server determined the mobile-transmission power and assigned
for achieving a system-level objective, e.g., the minimum server CPU cycles to different users in order to reduce the
sum mobile-energy consumption. Both the centralized and sum mobile-energy consumption. The optimal solution shows
distributed resource allocation schemes have been studied for that, there exists an optimal one-to-one mapping between the
different MEC systems as reviewed in the following.. transmission power and the number of allocated CPU cycles
For centralized resource allocation [84], [86], [101], [109]– for each mobile device. This work was further extended in
[114], the MEC server obtains all the mobile information, in- [101] to account for the optimal binary offloading based on
cluding the CSI and computation requests, makes the resource- the model of task-call graphs. In [112], the authors con-
allocation decisions, and informs the mobile devices about sidered the multiuser video compression offloading in MEC
the decisions. In [84], mobile users time-share a single edge and minimized the latency in local compression, edge cloud
server and have different computation workloads and local- compression and partial compression offloading scenarios.
computation capacities. A convex optimization problem was Besides, in order to minimize the energy and delay cost for
formulated to minimize the sum mobile-energy consumption. multi-user MEC systems where each user has multiple tasks,
The key finding is that the optimal policy for controlling Chen et al. jointly optimized the offloading decisions and the
offloading data size and time allocation has a simple threshold- allocation of communication resource via a separable semidef-
based structure. Specifically, an offloading priority function inite relaxation approach in [113], which was later extended
was firstly derived according to mobile users’ channel condi- in [114] by taking the computational resource allocation and
tions and local computing energy consumption. Then, the users processing cost into account. Different from [84], [86], [101],
 14

[112]–[114], the revenue of service providers was maximized latency using queuing theory. Second, even for synchronized
in [109] under constraints of quality of service (QoS) require- task arrivals, the latency requirements can differ significantly
ments for all mobile devices. The assumed fixed resource over users running different types of applications ranging from
usage of each user results in a semi-MDP problem, which latency-sensitive to latency-tolerant applications. This fact
was transformed into a linear programming (LP) model and calls for the server scheduling assigning users different levels
efficiently solved. In [110], assuming a stochastic task arrival of priorities based on their latency requirements. In [121], after
model, the energy-latency tradeoff in multiuser MEC systems the pre-resource allocation, the MEC server will check the
was investigated via a Lyapunov optimization-based online deadline of different tasks during the server computing process
algorithm, which jointly manages the available radio-and- and adaptively adjust the task execution order to satisfy the
computational resources. Centralized resource management for heterogeneous latency requirements. Last, some computation
multiuser MEC system based on cloud radio access network tasks each consists of several dependent sub-tasks such that the
(C-RAN) has also been investigated in [111]. scheduling of these modules must satisfy the task-dependency
Another thrust of research targets distributed resource allo- requirements. The task model with a sequential sub-task
cation for multiuser MEC systems which were designed using arrangement was considered in [122] that jointly optimizes
game theory and decomposition techniques [87], [88], [115]– the program partitioning for multiple users and the server-
[119]. In [115] and [87], the computation tasks were assumed computation scheduling to minimize the average completion
to be either locally executed or fully offloaded via single time. As a result, a heuristic algorithm was proposed to
and multiple interference channels, respectively. With fixed solve the formulated mixed-integer problem. Specifically, it
mobile-transmission power, an integer optimization problem first optimizes the computation partition for each user. Under
was formulated to minimize the total energy consumption and these partitions, it will search the time intervals violating the
offloading latency, which was proved to be NP-hard. Instead of resource constraint and adjust them accordingly. Furthermore,
designing a centralized solution, the game-theoretic techniques the general dependency-task model as shown in Fig. 4(c) was
were applied to develop a distributed algorithm that is able considered for multiple users in [118]. This model drastically
to achieve a Nash equilibrium. Moreover, it was shown that complicates the computing time characterization. To address
for each user, offloading is beneficial only when the received this challenge, a measure of ready time was defined for each
interference power is lower than a threshold. Furthermore, this sub-task as the earliest time when all the predecessors have
work was extended in [116] and [117], where each mobile has been computed. Then, the offloading decision, mobile CPU-
multiple tasks and can offload computation to multiple APs cycle frequency and mobile-transmission power were jointly
connected by a common edge-server, respectively. For the of- optimized to reduce the sum mobile-energy consumption and
floading process, in addition to transmission energy, this work computation latencies with a proposed distributed algorithm.
has also accounted for the scanning energy of the APs and the 3) Multiuser Cooperative Edge Computing: Multiuser
fixed circuit power. The proposed distributed offloading policy cooperative computing is envisioned as a promising technique
showed that a mobile device should handover the computation to improve the MEC performance by providing two advantages
to a different AP only when a new user choosing the same [123]–[129]. First, MEC servers with limited computational
AP achieves a larger benefit. Building on the system model resources may be overloaded when they have to serve a
in [87], the joint optimization for the mobile-transmission large number of offloading mobile users. In such cases, the
power and the CPU-cycle allocation of the edge server was burdens on the servers can be lightened via peer-to-peer mobile
investigated in [88]. To solve the formulated mixed-integer cooperative computing. Second, sharing the computational
problem, the decomposition technique was utilized to optimize resources among the users can balance the uneven distribution
the resource allocation and offloading decision sequentially. of the computation workloads and computation capabilities
Specifically, the offloading decision problem was reduced to a over users. In [123], D2D communication was proposed to
sub-modular maximization problem and solved by designing enable multiuser cooperative computing. In particular, this
a heuristic greedy algorithm. Similar decomposition technique work studied how to detect and utilize computational resources
and successive convex approximation technique were utilized on other users. This idea was adopted in [124] to propose
in [118] and [119] respectively to design distributed resource a D2D-based heterogeneous MCC networks. Such a novel
allocation algorithm for MEC systems. framework was shown to enhance the network capacity and
2) MEC Server Scheduling: The works discussed earlier offloading probability. Moreover, for wireless sensor networks,
[84], [86]–[88], [109], [117] are based on the assumptions of cooperative computing was proposed in [125] to enhance
user synchronization and the feasibility of parallel local-and- its computation capability. First, the optimal computation
edge computation. However, studying practical MEC server partition for minimizing the total energy consumption of
scheduling requires relaxation of these assumptions as dis- two cooperative nodes was investigated. This result was then
cussed below together with the resultant designs. First, the utilized to design the fairness-aware energy-efficient cooper-
arrival times of different users are in general asynchronous so ative node selection. Furthermore, Song et al. showed that
that it is desirable for the edge server with finite computational sharing computation results among the peer users can sig-
resource to buffer and compute the tasks sequentially, which nificantly reduce the communication traffic for a multiuser
incurs the queuing delay. In [120], to cope with the bursty MEC system [126]. Assuming the task can either be offloaded
task arrivals, the server scheduling was integrated with uplink- or computed locally, a mixed-integer optimization problem
downlink transmission scheduling to minimize the average was formulated to minimize the total energy consumption
 15

TABLE V
T HE COMPARISON OF PAPERS FOCUSING ON MULTIUSER MEC SYSTEMS .

Theme Design Design Objective Reference Proposed Solution

Type/Motivation
[84] Design the optimal threshold-based resource allocation policy
based on defined offloading priority function for TDMA and
OFDMA systems
[86] Jointly optimize the allocation of communication and com-
Energy putation resources
[101] Design the optimal resource allocation and code partitioning
by call-graph selection approach
Centralized
[111] Solve the non-convex resource allocation problem for C-RAN
using iterative algorithms
[112] Minimize the latency in multiuser video compression via
Latency
resource allocation
[110] Propose a Lyapunov optimization-based dynamic computa-
tion offloading policy
Energy and latency
[113], [114] Jointly optimize the offloading decisions and the allocation
of resource via semidefinite relaxation
Joint radio-and- Revenue [109] Design the optimal resource allocation based on semi-MDP
computational
resource [119] Propose a distributed iterative algorithm using successive
Energy
allocation convex approximation technique
[87], [115] Develop a distributed algorithm that is able to achieve a Nash
equilibrium
[116] Propose a distributed algorithm for multi-user MEC systems
Distributed where each user has multiple tasks
Energy and latency [117] Consider multiple servers and develop a distributed algorithm
admitting the Nash equilibrium
[118] Propose a decomposition algorithm to control the computation
offloading selection, clock frequency control and transmission
power allocation iteratively
[88] Propose a decomposition algorithm to optimize the resource
Utility
allocation and offloading decisions
Bursty data [120] Optimize the uplink and downlink scheduling using queuing
Latency
arrivals theory
Heterogeneous [121] Propose a pre-resource allocation and joint scheduling scheme
Energy
MEC server deadlines
scheduling [122] Propose heuristic algorithm with searching and adjusting
Latency
phases based on constraint relaxation
Task dependency
[118] Propose a decomposition algorithm to control the computation
Energy and latency
offloading selection, clock frequency control and transmission
power allocation iteratively
D2D Task success rate [123] Propose the optimal and periodic mobile cloud access scheme
communication Network capacity and [124] Propose D2D communication techniques in heterogeneous
offloading probability MEC systems
[125] Propose a fairness-aware energy-efficient cooperative node
selection scheme
Cooperation Energy
[127] Propose a four-slot protocol to enable joint computation and
Cooperative communication cooperation
computing Share [126] Propose a Lyapunov optimization-based cooperative comput-
Energy
computation ing policy
results
Share [128] Propose a “string-pulling” offloading policy based on con-
Energy
computational structed offloading feasibility tunnel
resource
Small BSs cooper- [129] Propose a peer offloading framework that allows both cen-
Delay cost
ation tralized and autonomous decision making

under the constraint of the system communication traffic. To Specifically, an offloading feasibility tunnel was constructed
tackle this challenging problem, two online task scheduling based on the helper’s CPU profile and buffer size. Given the
algorithms were proposed based on pricing and Lyapunov tunnel, the optimal offloading was shown to be achieved by
optimization theories. In addition, by employing a helper, the well-known “string-pulling” strategy, graphically referring
a four-slot joint computation-and-communication cooperation to pulling a string across the tunnel. Last, Chen et al. proposed
protocol was proposed in [127], where the helper not only an online peer offloading framework based on Lyapunov
computes part of the tasks offloaded from the user, but also optimization and game theoretic approaches in [129], which
acts as a relay node to forward the tasks to the MEC server. enables small BSs cooperation to handle the spatially uneven
Another recent work [128] investigated the optimal offloading computation workloads in the network.
policies in a peer-to-peer cooperative computing system where
the computing helper has time-varying computation resources. 4) Summary and Insight: The comparison of resource man-
agement schemes for multiuser MEC systems is provided in
 16

Table V. We draw several conclusions on resource allocation, device to offload tasks to multiple MEC servers was proposed
MEC server scheduling and mobile cooperative computing as in [133], and semidefinite relaxation-based algorithms were
follows. proposed to determine the task allocation decisions and CPU
• Consider multiuser MEC systems with finite radio-and- frequency scaling.
computational resources. For system-lever objectives, 2) Server Cooperation: Resource sharing via server co-
e.g., to minimize the sum mobile energy consumption, the operation can not only improve the resource utilization and
users with large channel gains and low local-computation increase the revenue of computing service providers, but also
energy consumption have higher priorities for offloading provide more resources for mobile users to enhance their
computation since they can contribute to larger energy user experience. This framework was originally proposed in
savings. Too many offloading users, however, will cause [134], which includes components such as resource allocation,
severe inter-user interference of communication and com- revenue management and service provider cooperation. First,
putation, which will, in turn, reduce the system revenue. resource allocation was optimized for cases with deterministic
• To effectively reduce the sum computation latency of and random user information to maximize the total revenues.
multiple users, the scheduling design for a MEC server Second, considering self-interested cloud service providers, a
should assign higher priorities to the users with more distributed algorithm based on game theory was proposed to
stringent latency requirements and heavy computation maximize service providers’ own profits, which was shown
loads. Moreover, parallel computing can further boost the to achieve the Nash equilibrium. This study was further
computation speed at the server. extended in [135], which considered both the local and remote
• Scavenging the enormous amount of distributed computa- resource sharing. The former refers to resource sharing among
tion resources can not only alleviate the network conges- different service providers within the same data center, while
tion, but also improves resource utilization and enables the latter one means the cooperation across different data
ubiquitous computing. This vision can be materialized centers. To realize the resource sharing and cooperation among
by peer-to-peer mobile cooperative edge computing. The different servers, a coalition game was formulated and solved
key advantages include short-range transmission via D2D by a game-theoretic algorithm with stability and convergence
techniques and computation resource and result sharing. guarantees. Moreover, the recent work [136] proposed a new
server cooperation scheme where edge servers exploit both the
computational and storage resources by proactively caching
C. MEC Systems with Heterogeneous Servers computation results to minimize the computation latency. The
To enable ubiquitous edge computing, heterogeneous MEC corresponding task distributing problem was formulated as a
(Het-MEC) systems were proposed in [130] comprising one matching game and solved by an efficient algorithm based on
central cloud and multiple edge servers. The coordination a proposed deferred-acceptance algorithm.
and interaction of multi-level central/edge clouds introduce 3) Computation Migration: In [137]–[139], apart from
many interesting new research challenges and recently have optimizing the offloading decisions, the authors also inves-
attracted extensive relevant investigations on server selection, tigated the computation migration among different remote
cooperation and computation migration, as discussed in the servers. Specifically, the computation migration over MEC
sequel. servers was motivated by the mobility of offloading users.
1) Server Selection: For users served by a Het-MEC When a user moves closer to a new MEC server, the network
system, a key design issue is to determine the destination of controller can choose to migrate the computation to this
computation offloading, i.e., either the edge or central cloud server, or compute the task in the original server and then
server. In [131], the server selection problem was studied forward the results back to the user via the new server. The
for a multiuser system comprising a single edge server and computation migration problem was formulated as an MDP
a single central cloud. To maximize the total successful problem based on a random-walk mobility model in [137].
offloading probability, a heuristic scheduling algorithm was It was shown that the optimal policy has a threshold-based
proposed to leverage both the low communication latency due structure, i.e., the migration should be selected only when the
to the proximity of the MEC server and the low computation distance of two servers is bounded by two given thresholds.
latency arising from abundant computational resources at the This work was further extended in [138] where the workload
central-cloud server. Specifically, when the computation load scheduling in edge servers was integrated with the service
of the MEC server exceeds a given threshold, latency-tolerant migration to minimize the average overall transmission and
tasks are offloaded to the central cloud to spare enough reconfiguration costs using Lyapunov optimization techniques.
computational resources at the edge server for processing Another computation migration framework was proposed in
latency-sensitive tasks. In addition, [132] explored the problem [139], where the MEC server can either process offloaded
of server selection over multiple MEC servers. The major computation tasks locally or migrate them to the central cloud
challenge arises from the correlation between the amounts server. An optimization problem was formulated to minimize
of the offloaded computation and selected edge servers for the sum mobile-energy consumption and computation latency.
multiple users. To cope with this issue, a congestion game This problem was solved by a heuristic two-stage algorithm,
was formulated and solved to minimize the sum energy which first determines the offloading decision for each user
consumption of mobile users and edge servers. Most recently, by the semi-definite relaxation and randomization techniques,
a computation offloading framework that allows a mobile and then performs the resource allocation optimization for all
 17

TABLE VI
T HE COMPARISON OF PAPERS FOCUSING ON MEC SYSTEMS WITH HETEROGENEOUS SERVERS .

Theme Design Type Design Objective Reference Proposed Solution

Edge/central Successful offload- [131] Propose a heuristic server selection algorithm according to
server selection ing probability the deadline requirements
Edge server selec- Energy [132] Formulate a congestion game and propose a distributed algo-
Server selection tion rithm admitting the Nash equilibrium
Multiple edge [133] Propose semidefinite relaxation-based algorithms for task
Energy and latency
servers allocation decisions and frequency scaling
Edge server coop- [134] Propose a distributed resource allocation algorithm admitting
Revenue
eration the Nash equilibrium
Server cooperation Edge/remote [135] Formulate a coalition game and propose a game-theoretic
Utility
server cooperation algorithm
Edge server [136] Study the distribution and proactive caching of computing
Latency
proactive caching tasks in MEC
[137] Propose a threshold-based computation migration scheme
Computation Edge server according to the distance
Cost
migration migration [138] Propose online workload scheduling and migration algorithms
using Lyapunov optimization techniques
Remote server mi- [139] Propose a heuristic two-stage algorithm including migration
Energy and latency
gration decision and resource allocation

the users. 1) Two-Timescale Resource Management: In most exist-

4) Summary and Insight: Table VI provides the summary ing works, e.g., [87], [88], [96], [119], [121], [140], wireless
of resource management schemes for MEC systems with channels were assumed to remain static during the whole task
heterogeneous servers. The literature provides a set of insights execution process for simplicity. Nevertheless, this assumption
on server selection, cooperation, and computation migration, may be unreasonable when the channel coherence time is
described as follows. much shorter than the latency requirement. For instance, at a
carrier frequency of 2GHz, the channel coherence time can be
• Consider MEC systems with multiple computation tasks
as small as 2.5ms when the speed is 100km/h. For some mobile
and heterogeneous servers. To reduce the sum computa-
applications such as the MMORPG game PlaneShift4 , the
tion latency, it is desirable to offload latency-insensitive
acceptable response time is 440ms and the excellent latency
but computation-intensive tasks to remote central cloud
is 120ms [141]. In such scenarios, the task offloading process
server and latency-sensitive ones to the edge servers.
may be across multiple channel blocks, necessitating the two-
• Server cooperation can significantly improve the compu-
timescale resource management for MEC. This problem is
tation efficiency and resource utilization at MEC servers.
very challenging even for a single-user MEC system with
More importantly, it can balance the computation load
deterministic task arrivals [81].
distribution over the networks so as to reduce sum com- 2) Online Task Partitioning: For ease of optimization,
putation latency while the resources are better utilized. existing literature tackling the task partitioning problems ig-
Moreover, the server cooperation design should con- nores the fluctuation of the wireless channels, and obtain the
sider temporal-and-spatial computation task arrivals and task partitioning decision before the start of the execution
server’s computation capacities, time-varying channels, process. With such an offline task partitioning decision, the
and servers’ individual revenue. change of the channel condition may lead to inefficient or
• Computation migration is an effective approach for mo- even infeasible offloading, which shall severely degrade the
bility management in MEC. The decision of migrate-or- computation performance. To develop online task partitioning
not depends on the migration overhead, distances between policies, one should incorporate the channel statistics into the
users and servers, channel conditions, and servers’ com- formulated task partitioning problem, which may easily belong
putation capacities. Specifically, when a user moves far to an NP-hard problem even under a static channel. In [99] and
away from its original MEC server, it is preferred to [142], approximate online task partitioning algorithms were
migrate the computation to nearby servers. derived for applications with serial and tree-topology task-
call graphs, respective, while solutions for general task models
remain unexploited.
D. Challenges 3) Large-Scale Optimization: The collaboration of multi-
ple MEC servers allows their resources to be jointly managed
In the preceding subsections, we have conducted a compre-
for serving a large number of mobile devices simultaneously.
hensive survey on the state-of-the-art resource management
However, the increase of the network size renders the re-
techniques for MEC systems. However, the progress is still
source management a large-scale optimization problem with
in the infancy stage and many critical factors have been
respect to a large number of offloading decisions as well as
overlooked for simplicity, which need to be addressed in future
radio-and-computational resource allocation variables. Con-
research efforts. In the following, we identify three critical
ventional centralized joint radio-and-computational resource
research challenges for resource management in MEC that
remain to be solved. 4 http://www.planeshift.it/
 18

1. Site Selection for MEC Servers

Deployment of MEC
Systems 2. MEC Network Architecture

3. Server Density Planning

1. Service Caching for MEC Resource Allocation

Cache-Enabled MEC
2. Data Caching for MEC Data Analytics

1. Mobility-Aware Online Prefetching

Mobility Management 2. Mobility-Aware Offloading Using D2D Communications

Future Research
Directions for MEC for MEC
3. Mobility-Aware Fault-Tolerant MEC

4. Mobility-Aware Server Scheduling

1. Dynamic Right-Sizing for Energy-Proportional MEC

Green MEC 2. Geographical Load Balancing for MEC

3. Renewable Energy-Powered MEC Systems

1. Trust and Authentication Mechanisms

Security and Privacy
2. Networking Security
Issues in MEC
3. Secure and Private Computation

Fig. 7. Future research directions for MEC.

management algorithms require a huge amount of information the latency caused by congestion and propagation delays in
and computation when applied to large-scale MEC systems, the core network. However, there is no formal definition of
which will inevitably incur a significant execution delay and what an MEC server should be, and the server locations
may whittle away the potential performance improvement, in the system are not specified. These invoke the site se-
e.g., latency reduction, brought by the MEC paradigm. To lection problems for MEC servers, which are significantly
achieve efficient resource management, it is required to design different from the conventional BS site selection problems,
distributed low-complexity large-scale optimization algorithms as the optimal placement of edge servers is coupled with the
with light signaling and computation overhead. Although computational resource provisioning, and both of them are
the recent advancements in large-scale convex optimization constrained by the deployment budget. Besides, the efficiency
[143] provide powerful tools for radio resource management, of an MEC system relies heavily on its architecture, which
they cannot be directly applied to optimize the computation should account for various aspects such as workload intensity
offloading decision due to its combinatorial and non-convex and communication rate statistics. In addition, it is critical
nature, which calls for new algorithmic techniques. for MEC vendors to determine the required server density
for catering the service demand, which is closely related to
IV. I SSUES , C HALLENGES , AND F UTURE R ESEARCH
the infrastructure deployment cost and marketing strategies.
D IRECTIONS
Nonetheless, the large-scale nature of MEC systems makes
Recent years have witnessed substantial research efforts on traditional simulation-based methods inapplicable, and thus
resource management for MEC as surveyed in the preceding solutions based on network-scale analysis are preferred. In this
section. However, there are lots of emerging research direc- subsection, we will discuss three research problems related
tions of MEC that are still largely uncharted. In this section, to MEC deployment, including the site selection for MEC
technical issues, challenges and research opportunities will be servers, the MEC network architecture, and server density
identified and discussed as summarized in Fig. 7, including planning.
the large-scale MEC system deployment, cache-enabled MEC, 1) Site Selection for MEC Servers: Selecting the sites
mobility management, green MEC and security-and-privacy for MEC infrastructures, especially MEC servers, is the first
issues in MEC. step towards building up the MEC system. To make the
A. Deployment of MEC Systems cost-effective server-site selection, the system planners and
The primary motivation of MEC is to shift the Cloud administrators should account for two important factors: site
Computing capability to the network edges in order to reduce rentals and computation demands. In general, given the system
 19

deployment budget, more MEC servers should be installed at

regions with higher computation demands, such as business 'DWD&HQWHU

districts, commercial areas and densely populated areas. This,

7LHU
however, contradicts the cost requirement as such areas are
'RZQORDG /7(%6
likely to have high site rentals. Fortunately, thanks to the well- 8SORDG
deployed telecom networks, it is a promising idea to install the
MEC servers co-located with the existing infrastructures such 7LHU
as macro BSs, which is even more attractive for the telecom
operators who would like to participate in the MEC market. 6PDOOFHOO%6
However, this would not solve all the problems. On one
7LHU
hand, due to the ever-increasing computation-quality require-
:L)L5RXWHU
ment and ubiquitous smart devices, satisfactory user experi-
ence cannot be guaranteed due to the poor signal quality and
(QG8VHUV
congestion in the macro cells. For some applications, e.g.,
smart home [144], it is desirable to move the computation
capability even closer to the end users. This can be achieved $GKRF&ORXG
by injecting some computational resources at small-cell BSs
[72], [73], which are low-cost and small-size BSs. Despite the
potential benefits, there are still obstacles on the way: Fig. 8. A 3-tier heterogeneous MEC system. Tier-1 servers are located in close
proximity to the end users, such as at WiFi routers and small-cell BSs, which
• First, due to physical limitations, the computation ca- are of relatively small computation capabilities. Tier-2 servers are deployed
pabilities of such kind of MEC servers will be much at LTE BSs with moderate computation capabilities. Tier-3 servers are the
smaller than those at macro BSs, making it challenging to existing Cloud Computing infrastructures, such as data centers.
handle computation-intensive tasks. One feasible solution
is to build a hierarchical network architecture for MEC
2) MEC Network Architecture: The promotion of MEC
systems comprising MEC servers with heterogeneous
does not mean the extinction of the data-center networks
communication-and-computation capabilities as detailed
(DCNs). Instead, future mobile computing networks are en-
in the sequel.
visioned to be consisted of three layers as shown in Fig. 8,
• Second, some of the small-cell BSs may be self-deployed
i.e., cloud, edge (a.k.a. fog layer), and the service subscriber
by the home users, and many femto BS owners may not
layer [130], [145]. While the cloud layer is mature and well-
have the motivation to collaborate with MEC vendors.
deployed, there is still some flexibility and uncertainty in
To overcome this issue, MEC vendors need to design
designing the edge layer.
a proper incentive mechanism in order to stimulate the
By analogy to the heterogeneous networks (HetNets) in
owners of small-cell BSs for renting the sites.
cellular systems, it is intuitive to design the Het-MEC systems,
• Moreover, deploying MEC servers at small-cell BSs may
which consist of multiple tiers. Specifically, the MEC servers
incur security problems as they are easy-to-reach and
in different tiers have distinct computation and communication
vulnerable to external attacks, which shall degrade the
capabilities. Such kinds of hierarchical MEC system structures
levels of reliability.
can not only preserve the advantage of efficient transmission
On the other hand, the computation hot spots do not always offered by HetNets, but also possess strong ability to handle
coincide with the communication hot spots. In other words, for the peak computation workloads by distributing them across
some of the computation hot spots, there exists no available different tiers [146]. However, the computation capacity provi-
communication infrastructure (either macro or small-cell BS). sioning problem is highly challenging and remains unsolved,
For these circumstances, we need to deploy edge servers with as it should account for many different factors, such as
wireless transceivers by properly choosing new locations. the workload intensity, communication cost between different
Besides, the site selection for MEC servers is dependent tiers, workload distribution strategies, etc.
on the computational resource-allocation strategy, which poses Another thrust of research efforts focuses on exploiting
extra challenges compared to the conventional BS site selec- the potential of the service subscriber layer, and utilizing
tion. Intuitively, concentrating the computational resources at the undedicated computational resources, e.g., laptops, smart
a few MEC servers can help save the site rentals. However, phones, and vehicles, overlaid with dedicated edge nodes.
this comes at the prices of potential degradation of the service This paradigm is termed as the Ad-hoc mobile cloud in
coverage and communication quality. In addition, the optimal literature [147]–[150]. The ad-hoc mobile cloud enjoys the
computational resource allocation should take into account benefits of amortizing the stress of MEC systems, increasing
both site rentals and computation demands. For example, for the utilization of the computational resources, and reducing
an MEC server at a site with a high site rental, it is preferred the deployment cost. However, it also brings difficulties in
to allocate huge computational resource and thus serve a large resource management and security issues due to its ad-hoc
number of users, for achieving the high revenue. Hence, a and self-organized nature.
joint site selection and computational resource provisioning 3) Server Density Planning: As mentioned in Sec-
problem needs to be solved before deploying MEC systems. tion IV-A2, the MEC infrastructure may be a combination
 20

500

400

300

200

100 Building B
Building A
Meter

−100

−200

−300
Building C
−400 MEC server
Mobile device Building D
−500
−500 −400 −300 −200 −100 0 100 200 300 400 500
Meter Fig. 10. Cache-enabled MEC systems.

Fig. 9. Illustration of the clustering behavior of the computation demands. by this fact, wireless content caching or FemtoCaching was
The mobile devices requesting for MEC services will be more concentrated
around the MEC servers.
proposed in [157]–[160] to avoid frequent replication for the
same contents by caching them at BSs. This technology has
attracted extensive attention from both academia and industry
of different types of edge servers, which provides various due to its striking advantages on reducing content acquisition
levels of computation experience and contributes different latency, as well as relieving heavy overhead burden of the
deployment costs. Hence, it is critical to determine the number network backhaul. While caching is to move popular contents
of edge nodes as well as the optimal combination of different close to end users, MEC is to deploy edge servers to handle
types of MEC servers with a given deployment budget and computation-intensive tasks for edge users to enhance user
computation demand statistics. Conventionally, this problem experience. Note that these two techniques seem to target
can only be addressed by numerical simulations, which is for diverse research directions, i.e., one for popular content
time-consuming and has poor scalability. Fortunately, owing delivery and the other for individual computation offloading.
to the recent development of stochastic geometry theory and However, they will be integrated seamlessly in this subsection
its successful applications in performance analysis for wireless and envisioned to create a new research area, namely, the
networks [151]–[154], as well as the similarity between Het- cache-enabled MEC.
MEC systems and HetNets, it is feasible to conduct per- Consider the novel cache-enabled MEC system shown in
formance analysis for MEC systems using techniques from Fig. 10. In such systems, the MEC server can cache several
stochastic geometry theory. Such analysis of MEC systems application services and their related database, called service
should address the following challenges: 1) The timescales caching (or service placement [161]) and data caching, respec-
of computation and wireless channel coherence time may tively, and handle the offloaded computation from multiple
be different [81], [104], which makes existing results for users. To efficiently reduce the computation latency, several
wireless networks not readily applicable for MEC systems. key and interesting problems need to be solved, which are
One possible solution is to combine the Markov chain and described in the following with potential solutions.
stochastic geometry theories to capture the steady behavior 1) Service Caching for MEC Resource Allocation: Unlike
of computations. 2) The computation offloading policy will the central cloud server that is always assumed with huge and
affect the radio resource management policy, which should diverse resources (e.g., computing, memory and storage), the
be taken into consideration. 3) The computation demands are current edge server has much less resources, making it unable
normally non-uniformly distributed and clustered (see Fig. 9), to accommodate all users’ computation requests. On the other
prohibiting the use of the homogeneous Poisson point process hand, different mobile services require different resources,
(HPPP) model for edge servers and service subscribers. It thus based on which, they can be classified into CPU-hungry (e.g.,
calls for the investigation of more advanced point processes, cloud chess and VR), memory-hungry (e.g., online Matlab),
e.g., the Ginibre α-determinantal point process (DPP), to and storage-hungry (e.g., VR) applications. Such a mismatch
capture the clustering behaviors of edge nodes [155]. between resource and demand introduces a key challenge on
how to allocate heterogeneous resources for service caching.
B. Cache-Enabled MEC Note that similar problems have been investigated in con-
It has been predicted by Cisco that mobile video streaming ventional Cloud Computing systems [162]–[165], termed as
will occupy up to 72% of the entire mobile data traffic by 2019 VM placement, as well as MCC systems [161]. Specifically, the
[156]. One unique property of such services is that the content authors in [162] proposed a novel architecture for VM manage-
requests are highly concentrated and some popular contents ment and optimized the VM placement over multiple clouds
will be asynchronously and repeatedly requested. Motivated to reduce the deployment costs and improve user experience,
 21

given constraints on hardware configuration, the number of data caching that only reserves frequently-used database. From
VMs as well as load balancing. Similar VM-placement prob- another perspective, caching parts of computation-result data
lems were also investigated in [163], [164] for maximizing the that is likely to be reused by others can further boost the
energy savings of cloud servers and in [165] for different cloud computation performance of the entire MEC system. One
scheduling strategies. Recently, the authors in [161] extended typical example is mobile cloud gaming, which enables fast
the VM placement idea to MCC systems and studied the and energy-efficient gaming by shifting game computing
joint optimization of service caching/placement over multiple engines from mobiles to edge servers and supporting real-
clouds and load dispatching for end users’ requests. As a time gaming by game video streaming. Thus, it emerges as
result, one efficient algorithm was proposed to minimize both a leading technique for next generation mobile computing
the computation latency and service placement transition cost. infrastructures [167]. Since certain game rendered videos, e.g.,
These works, however, cannot be directly applied to design gaming scenes, can be reused by other players, caching these
efficient service caching policies for MEC systems, since it computation results would not only significantly reduce the
should take into account more refined information including computation latency of the players with the same computation
users’ location, preference, experience as well as edge servers’ request, but also ease the computation burden for edge servers.
capacities in terms of the memory, storage and VM instance. Similar idea has been proposed in [168], which investigated
To this end, two possible approaches are described as follows. collaborative multi-bitrate video caching and processing in
The first one is spatial popularity-driven service caching, MEC.
referring to caching different combinations and amounts of For MEC data caching at a single edge server, one key
services in different MEC servers according to their specific problem is how to balance the tradeoff between massive
locations and surrounding users’ common interests. This idea database and finite storage capacity. Unlike FemtoCaching
is motivated by the fact that users in one small region are likely networks where content (data) caching mainly introduces
to request similar computing services. For example, visitors in a new multiple-access mechanism termed as cache-enabled
a museum tend to use AR for better sensational experience. access [169], data caching in MEC systems brings about
Thus, it is desirable to cache multiple AR services at the MEC manifold effects on the computation accuracy, latency and
server of this region for providing the real-time service. To edge server-energy consumption, which, however, have not
achieve the optimal spatial service caching, it is essential to been characterized in existing literature. This calls for model
construct a spatial-application popularity distribution model building research efforts for accurately quantifying the men-
for characterizing the popularity of each application over tioned effects for various MEC applications. Furthermore, it
different locations. Based on this, we can design resource- is also essential to establish a practical database popularity
allocation policies using various optimization algorithms, e.g., distribution model that is able to statistically characterize the
the game theory and convex optimization techniques. usage of each database set for different MEC applications.
An alternative approach is temporal popularity-driven ser- Based on the above models, the said tradeoff can be achieved
vice caching. The main idea is similar to that of the spatial by solving an optimization problem that maximizes the achiev-
counterpart, but it exploits the popularity information in the able QoS and minimizes the storage cost in MEC systems
temporal domain, since the computation requests also depend simultaneously.
on the time period. One example is that users are apt to play The above framework can be further extended to MEC
mobile cloud gaming after dinner. This kind of information systems with multiple servers where each server can serve
will suggest MEC operators to cache several gaming services multiple users and each user can offload computation to mul-
during this typical period for handling the huge computation tiple edge servers. The fundamental problem is similar to that
loads. One disadvantage of this temporal-based approach is of the cache-enabled HetNets [170], that is, how to spatially
the additional server cost resulted from frequent cache-and- distribute the database over heterogeneous edge servers under
tear operations since popularity information is time-varying both storage and computation-load constraints on each of
and MEC servers possess finite resources. them, for increasing network-wide revenue. Intuitively, for
2) Data Caching for MEC Data Analytics: Many modern each MEC server, it is desirable to spare more storage to cache
mobile applications involve intensive computation based on the database of the most popular applications in its cell, and
data analytics, e.g., ranking and classification. Take VR as an it also needs to utilize partial storage to accommodate less
instance. It creates an imaginary environment similar to the popular ones, whose computation performance will be further
real world by generating realistic images, sounds and other improved by cooperative caching in different MEC servers.
sensations for enhancing users’ experience. Achieving this end Moreover, the performance of large-scale cache-enabled MEC
is nontrivial as it requires the MEC server to finish multiple networks can be analyzed using stochastic geometry by mod-
complicated processes within the ultra-short duration (e.g., eling nearby users as clusters [171].
1ms), such as recognizing users’ actions via pattern recog-
nition, “understanding” users’ requests via data mining, as
well as rendering virtual settings via video streaming or other C. Mobility Management for MEC
sensation techniques [166]. All the above data-analytics based Mobility is an intrinsic trait of many MEC applications,
techniques should be supported by comprehensive database, such as VR assisted museum tour to enhance experience of
which, however, imposes extremely heavy burden on the edge visitors. In these applications, the movement and trajectory of
server storage. This challenge can be relieved by intelligent users provide location and personal preference information for
 22

Mobile Device’ Trajectory

Fig. 11. Mobility management for MEC.

the edge servers to improve the efficiency of handling users’ users with latency-tolerant tasks via designing intelligent cell
computation requests. On the other hand, mobility also poses association mechanisms. In [160], edge caching was integrated
significant challenges for realizing ubiquitous and reliable with mobility prediction in Follow-Me Cloud for enhancing
computing (i.e., without interruptions and errors) due to the the content-caches migration located at the edges. Recent
following reasons. First, MEC will be typically implemented proposals on mobility-aware wireless caching in [180] also
in the HetNet architecture comprising of multiple macro, provided valuable guidelines on mobility management in MEC
small-cell BSs and WiFi APs. Thus, users’ movement will systems.
call for frequent handovers among the small-coverage edge Note that most of the existing works focused on optimizing
servers as shown in Fig. 11, which is highly complicated mobility-aware server selection. However, to achieve better
due to the diverse system configurations and user-server as- user experience and higher network-wide profit, the offloading
sociation policies. Next, users moving among different cells techniques at mobile devices and scheduling policies at MEC
will incur severe interference and pilot contamination, which servers should be jointly considered. This introduces a set
shall greatly degrade the communication performance. Last, of interesting research opportunities with some described as
frequent handovers will increase the computation latency and follows.
thus deteriorate users’ experience. 1) Mobility-Aware Online Prefetching: In practice, the
Mobility management has been extensively studied for tra- full information of the user trajectory may be unavailable.
ditional heterogeneous cellular networks [172]–[174]. In these Conventional design for mobile computation offloading will
prior works, users’ mobility is modeled by the connectivity fetch a computation task to another server only when it is
probability or the link reliability according to such information handoverred. This mechanism requires excessive fetching of
as the users’ moving speeds. Based on such models, dynamic a large volume of data for handover and thus brings long
mobility management has been proposed to achieve high data fetching latency. Moreover, it also causes heavy loads on the
rate and low bit-error rate. However, these policies cannot MEC network. One promising solution to handle this issue is
be directly applied for MEC systems with moving users, to leverage the statistical information of the user trajectory and
since they neglect the effects of the computation resources prefetch parts of future computation data to potential servers
at edge servers on the handover policies. Recent works in during the server-computation time, referred to as online
[175]–[178] have made initial efforts to design mobility-aware prefetching [181]. This technique can not only significantly
MEC systems. Specifically, the inter-contact time and contact reduce the handover latency via mobility prediction, but also
rate were defined in [175] to model users’ mobility. An enable energy-efficient computation offloading by enlarging
opportunistic offloading policy was then designed by solving the transmission time. However, it also encounters several
a convex optimization problem for maximizing the successful challenges with two most critical ones described as follows.
task offloading probability. Alternatively, to account for the The first challenge arises from the trajectory prediction. Ac-
mobility, the number of edge servers that users can access was curate prediction can allow seamless handovers among edge
modeled by an HPPP in [176]. Then, the offloading decision servers and reduce the prefetching redundancy. Achieving it,
was optimized by addressing the formulated MDP problem however, requires precise modeling and high-complexity ML
to minimize the offloading cost including mobile-energy con- techniques, e.g., Bayesian, reinforcement and deep learning.
sumption, latency and failure penalty. Other mobility models For example, the trajectory of a typical visitor in a museum
were also proposed in [177], [178], which characterize the can be predicted according to his own interest-information
mobility by a sequence of networks that users can connect to and statistical route-information of some previous visitors
and a two-dimensional location-time workflow, respectively. with similar interests that can be obtained by ML algorithms.
In addition, mobility management for MEC was integrated Therefore, it is important to balance the tradeoff between the
with traffic control in [179] to provide better experience for modeling accuracy and computation complexity. The second
 23

challenge lies in the selection of the prefetched computation MEC applications. Next, fault detection is to collect fault
data. To maximize the successful offloading probability of information, which can be realized by setting intelligent timing
edge users, the computation-intensive components should be checks or receiving feedbacks for MEC services. In addition,
prefetched earlier with adaptive transmission power control in channel and mobility estimation techniques can also be applied
dynamic fading channels. to estimate the fault so as to reduce the detection time.
2) Mobility-Aware Offloading Using D2D Communica- Last, for detected MEC faults, recovery approaches should be
tions: D2D communications was first proposed in [182] to performed to continue and accelerate the MEC service. The
improve the network capacity and alleviate the data traffic suspended service can be switched to more reliable backup
burden in cellular systems. This paradigm can also be used wireless links with adaptive power control for higher-speed
to handle the user mobility problems in MEC systems [123], offloading. Alternative recovery approaches include migrating
which creates numerous D2D communication links. These the workloads to neighboring MEC systems directly or through
links allow the computation of a user to be offloaded to its ad-hoc relay nodes as proposed in [185].
nearby users which have more powerful computation capabil- 4) Mobility-Aware Server Scheduling: For multiuser MEC
ities. The short-range communication offered by D2D links systems, traditional MEC server scheduling servers users
reduces energy consumption of data transmission as well. according to the offloading priority order that depends on
However, user mobility brings new design issues as follows. users’ distinct local computing information, channel gains
The first one is how to exploit the advantages of both D2D and latency requirements [84]. However, this static scheduling
and cellular communications. One possible approach is to design cannot be directly applied for the multiuser MEC
offload the computation-intensive data to the edge servers systems with mobility due to dynamic environments, e.g.,
at BSs that have huge computation capabilities in order to time-varying channels and intermittent connectivities. Such
reduce the server-computing time; while the components of dynamics motivate the design of adaptive server scheduling
large data sizes and strict computation requirements should that regenerates the scheduling order from time to time,
be fetched to nearby users via D2D communications for incorporating the real-time user information. In such adaptive
higher energy efficiency. Next, the selection of surrounding scheduling mechanisms, users with worse conditions will
users for offloading should be optimized to account for users’ be allocated with higher offloading priorities to meet their
mobility information, dynamic channels and heterogeneous computing deadlines. Another potential approach is to design
users’ computation capabilities. Last, massive D2D links will mobility-aware offloading priority function by the following
introduce severe interference for reliable communications. two steps. The first step is to accurately predict users’ mobility
This issue is more complicated in the mobility-based MEC profiles and channels, where the major challenge is how
systems due to the fast-changing wireless fading environments. to reflect the mobility effects and re-define the offloading
Hence, advanced interference cancellation and cognitive radio priority function. The second step is resource reservation that
techniques can be applied for MEC systems, together with can enhance the server scheduling performance [186], [187].
mobility prediction to increase the offloading rate and reduce Specifically, to guarantee the QoS of latency-sensitive and
the service latency. high-mobility users, MEC servers can reserve some dedicated
3) Mobility-Aware Fault-Tolerant MEC: User mobility computational resources and provide reliable computing ser-
poses significant challenges for providing reliable MEC ser- vice for such users. While for other latency-tolerant users, the
vices due to dynamic environments. Computation offloading MEC server can perform on-demand provisioning. For such a
may fail due to intermittent connections and rapid-changing hybrid MEC server provisioning scheme, the server scheduling
wireless channels. The induced failure is catastrophic for can be optimized for serving the maximum number of users
the latency-sensitive and resource-demanding applications. For with QoS guarantees, as well as maximizing MEC servers’
instance, AR-based museum video guide aims to provide fluent revenue.
and fancy virtual sensations for visitors, and the disruption
or failure of video streaming due to intermittent connections
would upset visitors. Another example is the military operation D. Green MEC
which always requires fast and ultra-reliable computation, MEC servers are small-scale data centers, each of which
even in high-mobility environments. Any computation failure consumes substantially less energy than the conventional cloud
would bring serious consequences. These facts necessitate the data center. However, their dense deployment pattern raises a
design for mobility-aware fault-tolerant MEC systems [183]– big concern on the system-wide energy consumption. There-
[185], with three major and interesting problems illustrated as fore, it is unquestionably important to develop innovative tech-
follows, including fault prevention, fault detection and fault niques for achieving green MEC [188], [189]. Unfortunately,
recovery. Fault prevention is to avoid or prevent MEC fault designing green MEC is much more challenging compared
by backing up extra stable offloading links. Macro BSs or to green communication systems or green DCNs. Compared
central clouds can be chosen as protection-clouds, since they to green communication systems, the computational resource
have large network coverage that allows continuous MEC needs to be managed to guarantee satisfactory computation
service. The key design challenges lie in how to balance the performance, making the traditional green radio techniques not
tradeoff between QoS (i.e., the failure probability) and energy readily applicable. On the other hand, the previous research
consumption due to extra offloading links for the single- efforts on green DCNs have not considered the radio resource
user case, and how to allocate protection-clouds for multiuser management, which makes them not suitable for green MEC.
 24

Besides, the highly unpredictable computation workload pat- edge servers as well as user experience. On the other hand, it
tern in MEC servers poses another big challenge for resource can prolong the battery lives of mobile devices, as offloading
management in MEC systems, calling for advanced estimation the tasks through the nearby server could save transmission
and optimization techniques. In this subsection, we will intro- energy. It is worthwhile to note that the implementation of
duce different approaches on designing green MEC systems, GLB requires efficient resource management techniques at
including dynamic right-sizing for energy-proportional MEC, edge servers, such as dynamic right-sizing and VM manage-
geographical load balancing (GLB) for MEC, and MEC ment [194]–[197].
systems powered by renewable energy. Meanwhile, there are many factors to be incorporated when
1) Dynamic Right-Sizing for Energy-Proportional MEC: applying GLB in MEC environments. Firstly, since the mi-
The energy consumption of an MEC server highly depends grated tasks should go through the cellular core network, the
the utilization radio [see Eq. (5)]. Even when the server is network congestion state should be monitored and considered
idling, it still consumes around 70% of the energy as it when making the GLB decisions. Secondly, to enable seamless
operates at the full speed. This fact motivates the design task migration, a VM should be migrated/set up in another
of energy-proportional (or power-proportional) servers, i.e., edge server beforehand, which may cause additional energy
the energy consumption of a server should be proportional consumption. Thirdly, the mutual interests of MEC operators
to its computation load [190]. One way to realize energy- and edge computing service subscribers should be carefully
proportional servers is to switch off/slow down the processing considered when performing GLB, due to the tradeoff between
speeds of some edge servers with light computation loads. the energy savings and latency reduction. Last but not least,
Such an operation is termed as dynamic right-sizing in the the existence of conventional Cloud Computing infrastructures
literature on green DCNs [191]. However, along with the endows the edge servers with an extra option of offloading
potential energy savings, toggling servers between the active the latency-critical and computation-intensive tasks to remote
and sleep modes could bring detrimental effects. First of all, cloud data centers, creating a new design dimension and
it will incur the switching energy cost and application data- further complicating the optimization.
migration latency. Also, user experience may be degraded 3) Renewable Energy-Powered MEC Systems: Traditional
due to the less amount of allocated computational resources, grid energy is normally generated by coal-fired power plants.
which may, in turn, reduce the operator’s revenue. Besides, Hence, powering mobile systems with grid energy inevitably
the risk associated with server toggling as well as the wear- causes a huge amount of carbon emission, which opposes the
and-tear cost of the servers might be increased, which can in target of green computing. Off-grid renewable energy, such
turn increase the maintenance costs of MEC vendors. As a as solar radiation and wind energy, recently, has emerged as
result, switching off the edge servers in a myopic manner is a viable and promising power source for various IT systems
not always beneficial. thanks to the recent advancements of energy harvesting (EH)
In order to make an effective decision on dynamic right- techniques [198], [199]. This fact motivates the design of
sizing, the profile of computation workload at each edge server innovative MEC systems, called renewable energy-powered
should be accurately forecasted. In conventional DCNs, this MEC systems, which are shown in Fig. 12 comprising both
can be achieved rather easily as the workload at each data EH-powered MEC servers and mobile devices. On one hand,
center is an aggregation of the computation requests across as the MEC servers are expected to be densely-deployed
a large physical region, e.g., several states in the United and have low power consumption similar to that of small-
States, which is relatively stable so that it can be estimated cell BSs [200], it is reasonable and feasible to power the
by referring to the readily available historical data at the MEC infrastructures with the state-of-the-art EH techniques.
data centers. However, for MEC systems, the serving area On the other hand, the mobile devices can also get benefits
of each edge server is much smaller, and hence its workload from using renewable energy as EH is able to prolong their
pattern is affected by many factors, such as the location of the battery lives, which is one of the most favorable features for
server, time, weather, the number of nearby edge servers, and mobile phones [201]. Besides, the use of renewable energy
user mobility. This leads to a fast-changing workload pattern, sources eliminates the need of human intervention such as
and requires more advanced prediction techniques. Moreover, replacing/recharging the batteries, which is difficult if not
online dynamic right-sizing algorithms that require less future impossible for certain types of application scenarios where
information need to be developed. the devices are hard and dangerous to reach. Meanwhile, these
2) Geographical Load Balancing for MEC: GLB is an- advantages of using renewable energy are accompanied with
other key technique for green DCNs [192], [193], which new design challenges.
leverages the spatial diversities of the workload patterns, A fundamental problem to be addressed for renewable
temperatures, and electricity prices, to make workload routing energy-powered MEC systems is the green energy-aware
decision among different data centers. This technique can also resource allocation and computation offloading. Instead of
be applied to MEC systems. For instance, a cluster of MEC minimizing the energy consumption subject to satisfactory
servers can coordinate together to serve a mobile user, i.e., user experience, the design principle for the renewable energy-
the tasks can be routed from the edge server located in a powered MEC systems should be changed to optimizing the
hot spot (such as a restaurant) to a nearby edge server with achievable performance given the renewable energy constraint,
light workload (such as the one in a park). On one hand, this as the renewable energy almost comes for free. Also, with
helps to improve the energy efficiency of the lightly-loaded renewable energy supplies, the energy side information (ESI),
 25

Fig. 12. Renewable energy-powered MEC systems.

which indicates the amount of available renewable energy, will energy-powered edge servers can be densely deployed
play a key role in the decision making. Initial investigations over the system to provide more offloading opportunities
on renewable energy-powered MEC systems were conducted for the users. The resultant overlapping serving areas
in [202] and [203], which focused on EH-powered MEC offer the offloading diversity in the available energy to
servers and EH-powered mobile devices, respectively. For EH- avoid performance degradation. A similar idea has been
powered MEC servers, the system operator should decide proposed for EH cooperative communication systems in
the amount of workload required to be offloaded from the [208].
edge server to the central cloud, as well as the processing • Second, the chance of energy shortage can be reduced
speed of the edge server, according to the information of by properly selecting the renewable energy sources. It
the core network congestion state, computation workload, and was found in [192] that solar energy is more suitable
ESI. This problem was solved by a learning-based online for workloads with a high peak-to-mean ratio (PMR),
algorithm in [202]. While for EH-powered mobile devices, a while wind energy fits better for workloads with a small
dynamic computation offloading policy has been proposed in PMR. This provides guidelines for renewable energy
[203] using Lyapunov optimization techniques based on both provisioning for edge servers.
the CSI and ESI. However, these two works only considered • Third, MEC servers can be powered by hybrid energy
small-scale MEC systems that consist of either one edge sources to improve reliability [209]–[211], i.e., powered
server (in [202]) or one mobile device (in [203]). Thus, they by both the electric grid and the harvested energy. Also,
cannot provide a comprehensive solution for large-scale MEC equipping uninterrupted power supply (UPS) units at the
systems. edge servers can provide a short period of stable energy
For large-scale MEC systems where multiple MEC servers supply when green energy is in deficit, and it can be
are deployed across a large geographic region, the concept recharged when the surrounding energy condition returns
of GLB could be modified as the green energy-aware GLB to a good state.
to optimize the MEC systems by further utilizing the spatial • Moreover, wireless power transfer (WPT), which charges
diversity of the available renewable energy. This idea was mobile devices using RF wave [212], [213], is a newly-
originally proposed for green DCNs, where the “follow the emerged solution that enables wireless charging and
renewables” routing scheme offers a huge opportunity in extends the battery life. This technique has been provided
reducing the grid energy consumption [192], [204]–[207]. in modern mobile phones such as Samsung Galaxy S6.
Moreover, as mentioned before, there exist significant differ- In renewable energy-powered MEC systems, the edge
ences between MEC systems and conventional DCNs in terms servers can be powered by WPT when the renewable
of the wireless channel fluctuation and resource-management energy is insufficient for reliability [214]. This technology
design freedom of system operators. These factors make the also applies to the computation offloading for mobile
offloading decision making for the green energy-aware GLB in devices in MEC systems [83] and data offloading for
MEC systems much more complicated, as it needs to consider collaborate mobile clouds [215]. However, novel en-
the CSI and ESI in the whole system. ergy beamforming techniques are needed to increase the
The randomness of renewable energy may introduce the charging efficiency. Moreover, due to the double near-
offloading unreliability and risks of failure, bringing about far problem in wireless powered systems, it requires a
a major concern for using renewable energy to power MEC delicate scheduling to guarantee fairness among multiple
systems. Fortunately, there are several potential solutions to mobile devices.
circumvent this issue as described below.
• First, thanks to the low deployment cost, renewable
 26

E. Security and Privacy Issues in MEC Moreover, the large number of devices and entities in MEC
systems increase the chance of successfully attacking a single
There are increasing demands for secure and privacy- device, which provides means to launch an attack to the
preserving mobile services. While MEC enables new types whole system [225]. Therefore, novel and robust security
of services, its unique features also bring new security and mechanisms, such as hypervisor introspection, run-time mem-
privacy issues. First of all, the innate heterogeneity of MEC ory analysis, and centralized security management [226], are
systems makes the conventional trust and authentication mech- needed to guarantee a secured networking environment for
anisms inapplicable. Second, the diversity of communication MEC systems.
technologies that support MEC and the software nature of
3) Secure and Private Computation: Migrating
the networking management mechanisms bring new security
computation-intensive applications to the edge servers is
threats. Besides, secure and private computation mechanisms
the most important function and motivation of building MEC
become highly desirable as the edge servers may be an
systems. In practice, the task input data commonly contains
eavesdropper or an attacker. These motivate us to develop
sensitive and private information such as personal clinical
effective mechanisms as described in the following.
data and business financial records. Therefore, such data
1) Trust and Authentication Mechanisms: Trust is an should be properly pre-processed before being offloaded to
important security mechanism in almost every mobile system, edge servers, especially the untrusted ones, in order to avoid
behind which, the basic idea is to know the identity of the entity information leakage. In addition to information leakage,
that the system is interacting with. Authentication management the edge servers may return inaccurate and even incorrect
provides a possible solution to ensure “trust” [216]. However, computation results due to either software bugs or financial
the inherent heterogeneity of MEC systems, i.e., different incentives, especially for tasks with huge computation
types of edge servers may be deployed by multiple vendors demands [227]. To achieve secure and private computation,
and different kinds of mobile devices coexist, makes the it is highly preferred that the edge platforms can execute the
conventional trust and authentication mechanisms designed computation tasks without the need of knowing the original
for Cloud Computing systems inapplicable. For example, the user data and the correctness of the computation results can
reputation-based trust model will lead to severe trust threats be verified, which can be realized by encryption algorithms
in MEC systems, as demonstrated in [217]. This fact calls and verifiable computing techniques [228]. An interesting
for a unified trust and authentication mechanism that is able example of secure computation mechanisms for LP problems
to assess the reliability of edge servers and identify the was developed in [227], where the LP problem is decomposed
camouflaged edge servers. Besides, within the mobile network, into the public-owned solvers and the private-owned data.
there will be a large number of edge servers serving mas- By using a privacy-preserving transformation, the customer
sive mobile devices. This makes the trust and authentication offloads the encrypted private data for cloud execution, and
mechanism design much more complicated compared with the server returns the results for the transformed LP problem.
that in conventional Cloud Computing systems, since edge A set of necessary and sufficient conditions for verifying the
servers are of small computation capabilities and designed to correctness of the results were developed based on duality
enable latency-sensitive applications. Therefore, it is critical theory. Upon receiving the correct result, the clients can
to minimize the overhead of authentication mechanisms and map back the desired solution for the original problem using
design distributed policies [218], [219]. the secret transformation. This method of result validation
2) Networking Security: The communication technologies achieves a big improvement in computation efficiency via
to support MEC systems, e.g., WiFi, LTE and 5G, have their high-level LP computation compared to the generic circuit
own security protocols to protect the system from attacks and representation, and it incurs close-to-zero additional overhead
intrusions. However, these protocols inevitably create different on both the client and cloud server, which provides hints to
trust domains. The first challenge of networking security in develop secure and private computation mechanisms for other
MEC systems comes from the difficulties in the distribution cloud applications.
of credentials, which can be used to negotiate session keys
among different trust domains [216]. In existing solutions,
the certification authority can only distribute the credentials V. S TANDARDIZATION E FFORTS AND U SE S CENARIOS OF
to all the elements located within its own trust domain [216], MEC
making it hard to guarantee the privacy and data integrity for Standardization is an indispensable step for successful pro-
communications among different trust domains. To address motion of a new technology, which documents the consensus
this problem, we can use the cryptographic attributes as among multiple players and defines voluntary characteristics
credentials in order to exchange session keys [220], [221]. and rules in a specific industry. Due to the availability of
Also, the concept of federated content networks, which defines structured methods and reliable data, standardization helps to
how multiple trust domains can negotiate and maintain inter- promote innovation and disseminate groupbreaking ideas and
domain credentials [222], can be utilized. knowledge about cutting-edge techniques. More importantly,
Besides, techniques such as SDN and NFV are introduced standardization can build customer trust in products, services
to MEC systems to simplify the networking management as and systems, which helps to develop favorable market condi-
well as to provide isolation [5]. However, these techniques tion. The technical standards for MEC are being developed
are softwares by nature and thus vulnerable [223], [224]. by ETSI, and a new industry specification group (ISG) was
 27

0(&$SS 0(&$SS 0(&$SS 0(&$SS 0(&$SS

$SSOLFDWLRQ0DQDJHPHQW
90 90 90 90 90 6\VWHPV

$3, $3, $3, $3,

7UDIILFRIIORDGLQJ 5DGLR1HWZRUN &RPPXQLFDWLRQ

,QIRUPDWLRQ 6HUYLFH5HJLVWU\
IXQFWLRQ 72) 6HUYLFHV
6HUYLFH 51,6
0(&3ODWIRUPV
0(&$SSOLFDWLRQ3ODWIRUP6HUYLFHV 0DQDJHPHQW6\VWHP

0(&9LUWXDOL]DWLRQ0DQDJHU,DD6
0(&$SSOLFDWLRQ3ODWIRUP

0(&9LUWXDOL]DWLRQ/D\HU
0(&+RVWLQJ
,QIUDVWUXFWXUH
0DQDJHPHQW6\VWHP

0(&+DUGZDUH5HVRXUFHV
0(&+RVWLQJ,QIUDVWUXFWXUH

*335DGLR1HWZRUN(OHPHQW

Fig. 13. MEC platform overview [5].

established within ETSI by Huawei, IBM, Nokia Networks, well as the technical challenges and requirements of MEC
NTT docomo and Vodafone. The aim of the ISG is to build systems. Typical use scenarios of MEC will be then elaborated.
up a standardized and open environment, which will allow the In addition, we will discuss MEC-related issues in 5G stan-
efficient and seamless integration of applications from vendors, dardizations, including the functionality supports for MEC,
service providers, and third-parties across multi-vendor MEC and the innovative features in 5G systems with the potential
platforms [229]. In September 2014, an introductory technical to help realize MEC.
white paper on MEC was published by ETSI, which defined
the concept of MEC, proposed the referenced MEC platform, A. Referenced MEC Server Framework
as well as pointed out a set of technical requirements and
challenges for MEC [5]. Also, typical use scenarios and their In the MEC introductory technical white paper [5], the
relationships with MEC have been discussed. These aspects ETSI MEC ISG has defined a referenced framework for MEC
have also been documented in the ETSI specifications in 2015 servers (a.k.a. MEC platforms), where each server consists of a
[47], [230]–[232]. Most recently, ETSI has announced six hosting infrastructure and an application platform as shown in
Proofs of Concepts (PoCs) that were accepted by the MEC ISG Fig. 13. The hosting infrastructure includes the hardware com-
in MEC World Congress 2016, which will assist the strategic ponents (such as the computation, memory, and networking
planning and decision-making of organizations, as well as resources) and an MEC virtualization layer (which abstracts
help to identify which MEC solutions may be viable in the the detailed hardware implementation to the MEC application
network [233]. This provides the community with confidence platform). Also, the MEC host infrastructure provides the
in MEC and will accelerate the pace of the standardization. It interface to the host infrastructure management system as well
is interesting to note that, in this congress, the ETSI MEC ISG as the radio network elements, which, however, are beyond the
has renamed Mobile Edge Computing as Multi-access Edge scope of the MEC initiative due to the availability of multiple
Computing in order to reflect the growing interest in MEC implementation options.
from non-cellular operators, which will take effects starting The MEC application platform includes an MEC virtual-
from 2017 [234]. Most recently, the 3rd Generation Partner- ization manager together with an Infrastructure as a Service
ship Project (3GPP) shows a growing interest in including (IaaS) controller, and provides multiple MEC application
MEC into its 5G standard, and functionality supports for platform services. The MEC virtualization manager supports
edge computing has been identified and reported in a recent a hosting environment by providing IaaS facilities, while the
technical specification document [235]. In this section, we IaaS controller provides a security and resource sandbox (i.e.,
will first introduce the recent standardization efforts from the a virtual environment) for both the applications and MEC
industry, including the referenced MEC server framework as platform. The MEC application platform offers four main
categories of services, i.e., traffic offloading function (TOF),
 28

radio network information services (RNIS), communication them from adversely affecting other normal operations of the
services, and service registry. An MEC application platform network.
management interface is used by the operators for MEC 6) Operation: The virtualization and Cloud technologies
application platform management, supporting the application make it possible for various parties to participate in the
configuration and life cycle control, as well as VM operation management of MEC systems. Thus, the implementation of
management. the management framework should also consider the diversity
On top of the MEC application platform, the MEC appli- of potential deployments.
cations are deployed and executed within the VMs, which 7) Regulatory and Legal Considerations: The develop-
are managed by their related application management systems ment of MEC systems should meet the regulatory and legal
and agnostic to the MEC server/platform and other MEC requirements, e.g., the privacy and charging.
applications. Besides the aforementioned challenges and requirements,
there still exist more aspects that should be considered in the
B. Technical Challenges and Requirements final MEC standards, such as the support for user mobility,
applications/traffic migration, and requirements on the con-
In this subsection, we will briefly summarize the technical nectivity and storage. However, currently, the standardization
challenges and requirements specified in [5], [232]. efforts and even efforts from the research communities are still
1) Network Integration: As MEC is a new type of service on their infant stages.
deployed on top of the communication networks, the MEC
platform is supposed to be transparent to the 3GPP network
architectures, i.e., the existing 3GPP specifications should not C. Use Scenarios
be largely affected by the introduction of MEC. MEC will enable numerous mobile applications. In this
2) Application Portability: Application portability requires subsection, we will introduce four typical use scenarios that
MEC applications to be seamlessly loaded and executed by the have been documented by ETSI MEC ISG in [47].
MEC servers deployed by multiple vendors. This eliminates 1) Video Stream Analysis Service: Video stream analysis
the need for dedicated development or integration efforts for has a broad range of applications such as the vehicular
each MEC platform, and provides more freedom on optimizing license plate recognition, face recognition, and home security
the location and execution of MEC applications. It requires surveillance, for which, the basic operations include object
the consistency of the MEC application platform management detection and classification. The video analysis algorithms
systems, as well as mechanisms used to package, deploy and normally have a high computation complexity, and thus it is
manage applications from different platforms and vendors. preferable to move the analysis jobs away from the video-
3) Security: The MEC systems face more security chal- capturing devices (e.g., the camera) to simplify the device
lenges than communication networks due to the integration design and reduce the cost. If these processing tasks are
of computing and IT services. Hence, the security require- handled in the central cloud, the video stream should be routed
ments for the 3GPP networks and the IT applications (e.g., to the core network [236], which will consume a great amount
isolating different applications as much as possible) should of network bandwidth due to the nature of video stream.
be simultaneously satisfied. Besides, because of the nature By performing the video analysis in the place close to edge
of proximity, the physical security of the MEC servers is devices, the system can not only enjoy the benefits of low
more vulnerable compared to conventional data centers. Thus, latency, but also avoid the problem of network congestion
the MEC platforms need to be designed in a way that both caused by the video stream uploading. The MEC-based video
logical intrusions and physical intrusions are well protected. analysis system is shown in Fig. 14, where the edge server
Moreover, authorization is an important aspect to prevent the should have the ability to conduct video management and
unauthorized/untrusted third-party applications from destroy- analysis, and only the valuable video clips (screenshots) will
ing MEC hosts as well as the valued radio access network. be backed up to the cloud data centers.
4) Performance: As mentioned previously, the telecom 2) Augmented Reality Service: AR is a live direct or
operators expect that introducing MEC will have minimal indirect view of a physical, real-world environment whose
impacts on the network performance, e.g., the throughput, elements are augmented (or supplemented) by computer-
latency, and packet loss. Thus, sufficient capacity should be generated sensory inputs such as sound, video, graphics,
provisioned to process the user traffic in the system deploy- or GPS data5 . Upon analyzing such information, the AR
ment stage. Also, because of the highly-virtualized nature, the applications can provide additional information in real-time.
provided performance may be impaired especially for those The AR applications are highly localized and require low
applications that require intensive use of hardware resources latency as well as intensive data processing. One of the
or have low latency requirements. As a result, how to improve most popular applications is the museum video guides, i.e., a
the efficiency of virtualized environments becomes a big handheld mobile device that provides the detailed information
challenge. of some exhibits that cannot be easily shown on the scene.
5) Resilience: The MEC systems should offer certain Online games, such as the Pokémon Go6 , is another important
level of resilience and meet the high-availability requirements
demanded by their network operators. The MEC platforms 5 https://en.wikipedia.org/wiki/Augmented reality
and applications should have fault-tolerant abilities to prevent 6 http://www.pokemongo.com/
 29

9LGHR 9LGHR
0DQDJHPHQW $QDO\WLFV
9LGHR&OLSV6QDSVKRWV

0(&6HUYHU &RUH1HWZRUN 9LGHR6WRUDJH

9LGHR6WUHDP

Fig. 14. MEC for video stream analysis [5].

$52EMHFW
'DWD&DFKH
&HQWUDO
9LGHR $5&DFKH

2EMHFW,'
0RELOH'HYLFH 0(&6HUYHU &RUH1HWZRUN ,QWHUQHW&RQWHQW6HUYHU
Object
Tracker Mapper
Recognizer

Fig. 15. MEC for AR services [5].

application that AR techniques play a critical role. An MEC- is yet to come as the latency requirement cannot be met with
based AR application system is shown in Fig. 15, where the existing connected car clouds, which contributes to an end-
the MEC server should be able to distinguish the requested to-end latency between 100ms to 1s. MEC is a key enabling
contents by accurately analyzing the input data, and then technique for connected vehicles by adding computation and
transmit the AR data back to the end user. Much attention geo-distributed services to roadside BSs. By receiving and
has been paid on the MEC-enabled AR systems recently, and analyzing the messages from proximate vehicles and roadside
one demo has been implemented by Intel and roadshowed in sensors, the connected vehicle cloudlets are able to propagate
the Mobile World Congress 2016 [237]. the hazard warnings and latency-sensitive messages within
3) IoT Applications: To simplify the hardware complexity a 20ms end-to-end delay, allowing the drivers to react im-
of IoT devices and prolong their battery lives, it is promising to mediately (as shown in Fig. 16) and make it possible for
offload the computation-intensive tasks for remote processing autonomous driving. The connected vehicle technology has
and retrieve the results (required action) once the processing already attracted extensive attention from the automobile man-
is completed. Also, some IoT applications need to obtain ufacturers (e.g., Volvo, Peugeot), automotive supplier (e.g.,
distributed information for computation, which might be dif- BOSCH), telecom operators (e.g., Orange, Vodafone, NTT
ficult for an IoT device without the aid of an external entity. docomo), telecom vendors (e.g., QualComm, Nokia, Huawei),
Since the MEC servers host high-performance computation as well as many research institutes. In November 9 2015,
capabilities and are able to collect distributed information, Nokia7 presented two use cases for connected vehicles on
their deployment will significantly simplify the design of IoT an automotive driving testbed, including the emergency brake
devices, without the need to have strong processing power light and cooperative passing assistance.
and capability to receive information from multiple sources In addition to connected vehicle systems with automobiles,
for performing meaningful computation. Another important MEC will also be applicable for enabling connected unmanned
feature of IoT is the heterogeneity of the devices running aerial vehicles (UAVs), which play an increasingly important
different forms of protocols, and their management should be role in various scenarios such as photography, disaster re-
accomplished by a low-latency aggregation point (gateway), sponse, inspection and monitoring, precision agriculture, etc.
which could be the MEC server. In 2016, Nokia proposed the UAV traffic management (UTM)
based MEC architecture for connected UAVs in [241], where
4) Connected Vehicles: The connected vehicle technology the UTM unit provides functions of fleet management, auto-
can enhance safety, reduce traffic congestion, sense vehicles’ mated UAV missions, 3D navigation, and collision avoidance.
behaviors, as well as provide opportunities for numerous However, as existing mobile networks are mainly designed for
value-added services such as the car finder and parking lo-
cation [238]–[240]. However, the maturity of such technology 7 https://networks.nokia.com/solutions/mobile-edge-computing
 30

Fig. 16. MEC for connected vehicles.

users on the ground, UAVs will have very limited connectivity 1) Functionality Supports Offered by 5G Networks: From
and bandwidth. Therefore, reconfiguring the mobile networks the 5G network operators’ point of view, reducing the end-
to guarantee the connectivity and low latency between the to-end latency and load on the transport networks are two
UAVs and the infrastructure becomes a critical task for de- dominant design targets, which could possibly be achieved
signing MEC systems for connected UAVs. with MEC as operators and third part applications could
Due to limited space, we omit the description of some other be hosted close to the user equipment’s (UE’s) associated
interesting application scenarios, such as active device track- wireless AP. To integrate MEC in 5G systems, the recent 5G
ing, RAN-aware content optimization, distributed content and technical specifications have explicitly pointed out necessary
Domain Name System (DNS) caching, enterprise networks, as functionality supports that should be offered by 5G networks
well as safe-and-smart cities. Interested readers may refer to for edge computing, as listed below:
the white papers on MEC [5], [21], [242] for details.
• The 5G core network should select the traffic to be routed
to the applications in the local data networks.
D. MEC in 5G Standardizations • The 5G core network selects a user plane function (UPF)
in proximity to the UE to route and execute the traffic
The 5G standard is currently under development, which is to
steering from the local data networks via the interface,
enable the connectivity of a broad range of applications with
which should be based on the UE’s subscription data,
new functionality, characteristics, and requirements [77]. To
UE location, and the data from the application function
achieve these visions, the network features and functionality
(AF).
in 5G networks are foreseen to be migrated from hardware to
• The 5G network should guarantee the session and service
software, thanks to the recent development of SDN and NFV
continuity to enable UE and application mobility.
techniques. Since 2015, MEC (together with SDN and VFN) is
• The 5G core network and AF should provide information
recognized by the European 5G infrastructure Public Private
to each other via the network exposure function (NEF)8 .
Partnership (5GPPP) research body as one of the key emerging
• The policy control function (PCF)9 provides rules for QoS
technologies for 5G networks as it is a natural development
control and charging for the traffic routed to the local data
in the evolution of mobile BSs and the convergence of IT and
network.
telecommunication networking [15]. In April 2017, 3GPP has
included supporting edge computing as one of the high level
features in 5G systems in the technical specification document 8 The NEF supports external exposure of capabilities of network functions,

[235], which will be introduced in this subsection. We have which can be categorized into monitoring capability, provisioning capability,
also identified some innovative features of 5G systems, which and policy/charging capability [235].
9 The PCF was defined as a stand-alone functional part of the 5G core
would pave the way for the realization, standardization and network that allows to shape the network behaviour based on the operator
commercialization of MEC. policies [243].
 31

2) Innovative Features in 5G to Facilitate MEC: Compared is regarded as an indispensable feature in 5G systems to

to previous generations of wireless networks, 5G networks support different services running across a single radio
possess various innovative features that are beneficial to the access network. Existing studies found that network slic-
realization, standardization, and commercialization of MEC. ing is of supreme need for three use scenarios, including
Three of them will be detailed in this subsection, including the ultra-reliable and low latency communication (URLLC),
support service requirement, mobility management strategy, massive machine type communication (mMTC), and en-
and capability of network slicing. hanced mobile broadband (eMBB) [245]. With the capa-
• Support Service Requirement: In 5G systems, the bility of network slicing in 5G systems, MEC applications
QoS characteristics (in terms of resource type, priority could be provisioned with optimized and dedicated net-
level, packet delay budget, and packet error rate), which work resources, which could help to reduce the latency
describe the packet forwarding treatment that a QoS flow incurred by the access networks substantially and support
receives edge-to-edge between the UE and the UPF, are intense access of MEC service subscribers.
associated with the 5G QoS Indicator (5QI). In [235],
a standardized 5QI to QoS mapping table is provided, VI. C ONCLUSION
showing a broad range of services that can be supported MEC is an innovative network paradigm to cater for the
in 5G systems. In particular, 5G systems are able to cater unprecedented growth of computation demands and the ever-
the requirements of latency-sensitive applications (e.g., increasing computation quality of user experience require-
real-time gaming and vehicular-to-everything (V2X) mes- ments. It aims at enabling Cloud Computing capabilities and
sages, which have a stringent packet budget delay require- IT services in close proximity to end users, by pushing
ment, i.e., <50ms, and a relatively small packet error abundant computational and storage resources towards the
rate < 10−3 ), and mission-critical services (e.g., push- network edges. The direct interaction between mobile devices
to-talk signaling that has both low delay (<60ms) and and edge servers through wireless communications brings the
small packet error rate (< 10−6 ) requirements). These possibility of supporting applications with ultra-low latency
applications coincide with typical MEC applications as requirement, prolonging device battery lives and facilitating
mentioned in Section V-C, i.e., 5G network is a viable highly-efficient network operations. However, they come along
choice for wireless communications in MEC systems. with various new design considerations and unique challenges
• Advanced Mobility Management Strategy: The con-
due to reasons such as the complex wireless environments and
cept of mobility pattern was introduced for designing mo- the inherent limited computation capacities of MEC servers.
bility management strategy for 5G systems. Such strate- In this survey, we presented a comprehensive overview and
gies may be used by the 5G core network to characterize research outlook of MEC from the communication perspective.
and optimize UE mobility. Specifically, the mobility To this end, we first summarized the modeling methodologies
pattern could be determined, monitored, and updated by on key components of MEC systems such as the computation
the 5G core network based on the subscription of the UE, tasks, communications, as well as mobile devices and MEC
statistics of UE mobility, network local policy, and UE servers computation. This help characterize the latency and
assisted information [235]. The mobility pattern not only energy performance of MEC systems. Based upon the system
plays a central role on designing advanced transmission modeling, we conducted a comprehensive literature review
schemes in wireless communication systems, but also on recent research efforts on resource management for MEC
becomes a non-negligible design consideration for many under various system architectures, which exploit the concepts
MEC applications discussed in Section V-C, e.g., the AR of computation offloading, joint radio-and-computational re-
services and connected vehicular applications. Thus, inte- source allocation, MEC server scheduling, as well as multi-
gration of advanced mobility management strategies that server selection and cooperation. A number of potential re-
make full use of the mobility pattern in 5G network can search directions were then identified, including MEC de-
help to develop an efficient wireless interface for MEC ployment issues, cache-enabled MEC, mobility management
systems. Besides, the mobility pattern obtained from the for MEC, green MEC, as well as security-and-privacy issues
5G core network can be further leveraged to design joint in MEC. Key research problems and preliminary solutions
radio-and-computational resource management strategies for each of these directions were elaborated. Finally, we
for MEC systems. introduced the recent standardization efforts from industry,
• Capability of Network Slicing: Network slicing is a
along with several typical use scenarios. The comprehensive
form of agile and virtual network architecture that allows overview and research outlook on MEC provided in this
multiple network instances to be created on top of a survey hopefully can serve as useful references and valuable
common shared physical infrastructure10 . Each of the guidelines for further in-depth investigations of MEC.
network instances is optimized for a specific service,
enabling resource isolation and customized network oper-
R EFERENCES
ations [244]. Due to the heterogeneous types of services
that 5G systems need to support (different requirements in [1] M. Armbrust, R. G. A. Fox, A. D. Joseph, R. H. Katz, A. Konwinski,
G. Lee, D. A. Patterson, A. Rabkin, I. Stoica, and M. Zaharia,
terms of functionality and performance), network slicing “Above the clouds: A berkeley view of cloud computing,” Feb. 2012.
[Online]. Available: https://www2.eecs.berkeley.edu/Pubs/TechRpts/
10 https://5g.co.uk/guides/what-is-network-slicing/ 2009/EECS-2009-28.pdf
 32

[2] Q. Zhang, L. Cheng, and R. Boutaba, “Cloud computing: State-of-the- [26] S. Clinch, J. Harkes, A. Friday, N. Davies, and M. Satyanarayanan,
art and research challenges,” Journal Internet Services appl., vol. 1, “How close is close enough? Understanding the role of cloudlets in
no. 1, pp. 7–18, 2010. supporting display appropriation by mobile users,” in Proc. IEEE Int.
[3] N. Wingfield, “Amazon’s profits grow more than 800 percent, Conf. Pervasive Comput. Commun. (PerCom), Lugano, Switzerland,
lifted by cloud services,” The New York Times, Jul. 2016. Mar. 2012, pp. 122–127.
[Online]. Available: http://www.nytimes.com/2016/07/29/technology/ [27] T. X. Tran, A. Hajisami, P. Pandey, and D. Pompili, “Collaborative
amazon-earnings-profit.html? r=0 mobile edge computing in 5G networks: New paradigms, scenarios,
[4] M. Chiang and T. Zhang, “Fog and IoT: An overview of research and challenges,” IEEE Commun. Mag., vol. 55, no. 4, Apr. 2017.
opportunities,” IEEE Internet Things J., vol. PP, no. 99, pp. 1–1, 2016. [28] S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, and W. Wang, “A
[5] ETSI, “Mobile-edge computing introductory technical white survey on mobile edge networks: Convergence of computing, caching
paper,” White Paper, Mobile-edge Computing Industry Initiative. and communications,” IEEE Access, to appear.
[Online]. Available: https://portal.etsi.org/portals/0/tbpages/mec/docs/ [29] J. Zhang, W. Xie, F. Yang, and Q. Bi, “Mobile edge computing and
mobile-edge computing - introductory technical white paper v1 field trial results for 5G low latency scenario,” China Commun., vol. 13,
[6] G. P. Fettweis, “The tactile Internet: Applications and challenges,” no. 2 (Supplement), pp. 174–182, 2016.
IEEE Veh. Techn. Mag., vol. 9, no. 1, pp. 64–70, Mar. 2014. [30] E. Cuervo, A. Balasubramanian, D.-k. Cho, A. Wolman, S. Saroiu,
[7] A. A. Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and R. Chandra, and P. Bahl, “Maui: Making smartphones last longer
M. Ayyash, “Internet of Things: A survey on enabling technologies, with code offload,” in Proc. ACM Int. Conf. Mobile Syst. Appl. Serv.
protocols, and applications,” IEEE Commun. Surveys Tuts., vol. 17, (MobiSys), San Francisco, California, USA, Jun. 2010, pp. 49–62.
no. 4, pp. 2347–2376, 4th Quarter 2015. [31] M. Satyanarayanan, P. Bahl, R. Caceres, and N. Davies, “The case for
[8] Juniper, “Smart wireless devices and the Inter- VM-based cloudlets in mobile computing,” IEEE Pervasive Comput.,
net of me,” White paper, Mar. 2015. [Online]. vol. 8, no. 4, pp. 14–23, 2009.
Available: http://itersnews.com/wp-content/uploads/experts/2015/03/ [32] 5GPPP, “5g automotive vision,” White Paper. [On-
96079Smart-Wireless-Devices-and-the-Internet-of-Me.pdf line]. Available: https://5g-ppp.eu/wp-content/uploads/2014/02/
[9] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. Soong, 5G-PPP-White-Paper-on-Automotive-Vertical-Sectors.pdf
and J. C. Zhang, “What will 5G be?” IEEE J. Sel. Areas Commun., [33] O. Khalid, M. Khan, S. Khan, and A. Zomaya, “OmniSuggest: A
vol. 32, no. 6, pp. 1065–1082, 2014. ubiquitous cloud based context aware recommendation system for
[10] CISCO, “The Internet of Things how the next evolution of the mobile social networks,” IEEE Trans. Serv. Comput., vol. 7, no. 3,
Internet is changing everything,” White paper, Apr. 2011. [Online]. pp. 401–414, Dec. 2014.
Available: http://www.cisco.com/c/dam/en us/about/ac79/docs/innov/ [34] K. Goel and M. Goel, “Cloud computing based e-commerce model,”
IoT IBSG 0411FINAL.pdf in Proc. IEEE Int. Conf. Recent Trends in Electron., Info. & Commun.
[11] F. Bonomi, R. Milito, J. Zhu, and S. Addepalli, “Fog computing and its Techn. (RTEICT), Banglore, India, May 2016, pp. 27–30.
role in the Internet of Things,” in Proc. ACM 1st edition of the MCC [35] G. Riah, “E-learning systems based on cloud computing: A review,”
workshop on Mobile cloud computing, 2012, pp. 13–16. ELSEVIER Proc. Comput. Sci., vol. 62, pp. 352–359, Sep. 2015.
[12] S. Yi, C. Li, and Q. Li, “A survey of fog computing: Concepts, [36] A. Abbas and S. U. Khan, “A review on the state-of-the-art privacy-
applications and issues,” in Proc. ACM Workshop on Mobile Big Data, preserving approaches in the e-health clouds,” IEEE J. Biomed. Health
2015, pp. 37–42. Inform., vol. 18, no. 4, pp. 1431–1441, Apr. 2014.
[13] G. I. Klas, “Fog computing and mobile edge cloud gain momentum
[37] G. Intelligence, “Understanding 5G: Perspectives on future technolog-
open fog consortium, ETSI MEC and Cloudlets,” 2015. [Online].
ical advancements in mobile,” London, UK, 2014.
Available: http://www.engpaper.com/mobile-computing-2015.htm
[38] A. Somov and R. Giaffreda, “Powering IoT devices: Technologies
[14] R. P. Goldberg, “Survey of virtual machine research,” Computer, vol. 7,
and opportunities,” Available: http://iot.ieee.org/newsletter/november-
no. 6, pp. 34–45, 1974.
2015/powering-iot-devices-technologies-and-opportunities.html.
[15] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young, “Mobile
edge computing—A key technology towards 5G,” ETSI White Paper, [39] R. Kemp, N. Palmer, T. Kielmann, F. Seinstra, N. Drost, J. Maassen,
vol. 11, 2015. and H. Bal, “EyeDentify: Multimedia cyber foraging from a smart-
[16] C.-Y. Chang, K. Alexandris, N. Nikaein, K. Katsalis, and T. Spyropou- phone,” in Proc. IEEE Int. Symp. Multimedia, San Diego, CA, USA,
los, “MEC architectural implications for LTE/LTE-A networks,” in Dec. 2009, pp. 392–399.
Proc. ACM Workshop on Mobility in the Evolving Internet Architecture [40] B. Shi, J. Yang, Z. Huang, and P. Hui, “Offloading guidelines for
(MobiArch), New York, NY, Oct. 2016, pp. 13–18. augmented reality applications on wearable devices,” in Proc. ACM
[17] Z. Q. Jaber and M. I. Younis, “Design and implementation of real time Int. Symp. Multimedia, Brisbane, Australia, Oct. 2015, pp. 1271–1274.
face recognition system (RTFRS),” Int. J. Compt. Appl., vol. 94, no. 12, [41] W. N. Schilit, “A system architecture for context-aware mobile com-
pp. 15–22, May 2014. puting,” Ph.D. dissertation, Columbia University, 1995.
[18] T. Verbelen, P. Simoens, F. D. Turck, and B. Dhoedt, “Leveraging [42] C. Perera, A. Zaslavsky, P. Christen, and D. Georgakopoulos, “Context
cloudlets for immersive collaborative applications,” IEEE Pervasive aware computing for the Internet of Things: A survey,” IEEE Commun.
Comput., vol. 12, no. 4, pp. 30–38, Oct.-Dec. 2013. Surveys Tuts., vol. 16, no. 1, pp. 414–454, 1st Quater 2014.
[19] A. A. Shuwaili and O. Simeone, “Energy-efficient resource allocation [43] S. Nunna, A. Kousaridas, M. Ibrahim, M. Dillinger, C. Thuemmler,
for mobile edge computing-based augmented reality applications,” H. Feussner, and A. Schneider, “Enabling real-time context-aware
IEEE Wireless Commun. Lett., vol. PP, no. 99, Apr. 2017. collaboration through 5G and mobile edge computing,” in Proc. IEEE
[20] A.-S. Ali and O. Simeone, “Energy-efficient resource allocation for Int. Conf. Inf. Techn. New Generations (ITNG), Las Vegas, NV, Apr.
mobile edge computing-based augmented reality applications,” IEEE 2015, pp. 601–605.
Wireless Commun. Lett., to appear. [44] X. Luo, “From augmented reality to augmented computing: A look
[21] Juniper, “White paper: Mobile edge computing use cases & at cloud-mobile convergence,” in Proc. IEEE Int. Symp. Ubiquitous
deployment options.” [Online]. Available: https://www.juniper.net/ Virtual Reality, Gwangju, South Korea, Jul. 2009, pp. 29–32.
assets/us/en/local/pdf/whitepapers/2000642-en.pdf [45] A. Thiagarajan, L. Ravindranath, H. Balakrishnan, S. Madden, and
[22] M. Armbrust, A. Fox, R. Griffith, A. D. Joseph, R. Katz, A. Konwinski, L. Girod, “Accurate, low-energy trajectory mapping for mobile de-
G. Lee, D. Patterson, A. Rabkin, I. Stoica, and M. Zaharia, “A overview vices.” in Proc. USENIX Symp. Networked Systems Design and Imple-
of cloud computing,” Commun. ACM, vol. 53, no. 4, pp. 52–58, Apr. mentation (NSDI), Boston, MA, Mar. 2011, pp. 1–14.
2010. [46] H. Suo, Z. Liu, J. Wan, and K. Zhou, “Security and privacy in
[23] M. Othman, S. A. Madani, S. U. Khan et al., “A survey of mobile mobile cloud computing,” in Proc. IEEE Int. Wireless Commun. Mobile
cloud computing application models,” IEEE Commun. Surveys Tuts., Comput. Conf. (IWCMC), Cagliari, Italy, Jul. 2013, pp. 655–659.
vol. 16, no. 1, pp. 393–413, 1st Quater 2014. [47] ETSI, “Mobile-edge computing (MEC): Service scenarios.” [Online].
[24] M. F. Bari, R. Boutaba, R. Esteves, L. Z. Granville, M. Podlesny, Available: http://www.etsi.org/deliver/etsi gs/MEC-IEG/001 099/004/
M. G. Rabbani, Q. Zhang, and M. F. Zhani, “Data center network 01.01.01 60/gs mec-ieg004v010101p.pdf
virtualization: A survey,” IEEE Commun. Surveys Tuts., vol. 15, no. 2, [48] W. Shi and S. Dustdar, “The promise of edge computing,” Comput,
pp. 909–928, 2nd Quarter 2013. vol. 49, no. 5, pp. 78–81, May 2016.
[25] A. Ghiasi and R. Baca, “Overview of largest data centers,” IEEE [49] O. Salman, I. Elhajj, A. Kayssi, and A. Chehab, “Edge computing
802.3bs Task Force Interim Meeting, May 2014. [Online]. Available: enabling the Internet of Things,” in Proc. IEEE World Forum Internet
http://www.ieee802.org/3/bs/public/14 05/ghiasi 3bs 01b 0514.pdf of Things (WFIOT), Dec. 2015, pp. 603–608.
 33

[50] A. Ahmed and E. Ahmed, “A survey on mobile edge computing,” in [74] S. Han, Y.-C. Liang, and B.-H. Soong, “Spectrum refarming: A new
Proc. IEEE Int. Conf. Intell. Syst. Control (ISCO), Coimbatore, India, paradigm of spectrum sharing for cellular networks,” IEEE Trans.
Jan. 2016, pp. 1–8. Commun., vol. 63, no. 5, pp. 1895–1906, May 2016.
[51] T. Taleb, K. Samdanis, B. Mada, H. Flinck, S. Dutta, and D. Sabella, [75] Q. Chen, G. Yu, and Z. Ding, “Optimizing unlicensed spectrum
“On multi-access edge computing: A survey of the emerging 5g net- sharing for LTE-U and WiFi network coexistence,” IEEE J. Sel. Areas
work edge architecture &amp; orchestration,” IEEE Commun. Surveys Commun., vol. 34, no. 10, pp. 2562–2574, Oct. 2016.
Tuts., to appear. [76] P. Kryszkiewicz, A. Kliks, and H. Bogucka, “Small-scale spectrum
[52] H. Liu, F. Eldarrat, H. Alqahtani, A. Reznik, X. de Foy, and aggregation and sharing,” IEEE J. Sel. Areas Commun., vol. 34, no. 10,
Y. Zhang, “Mobile edge cloud system: Architectures, challenges, and pp. 2630–2641, Oct. 2016.
approaches,” IEEE Syst. J., to appear. [77] ERICSSON, “5G radio access - Capabilities and technologies,”
[53] M. T. Beck, M. Werner, S. Feld, and S. Schimper, “Mobile edge White paper, Apr. 2016. [Online]. Available: https://www.ericsson.
computing: A taxonomy,” in Proc. Int. Conf. Advances Future Internet com/assets/local/publications/white-papers/wp-5g.pdf
(AFIN), Lisbon, Portugal, Nov 2014, pp. 48–54. [78] T. Burd and R. Broderson, “Processor design for portable systems,”
[54] P. Mach and Z. Becvar, “Mobile edge computing: A survey on archi- Kluwer J. VLSI Signal Process. Syst., vol. 13, no. 2/3, pp. 203–221,
tecture and computation offloading,” IEEE Commun. Surveys Tuts., to Aug. 1996.
appear. [79] W. Yuan and K. Nahrstedt, “Energy-efficient CPU scheduling for
[55] D. Sabella, A. Vaillant, P. Kuure, U. Rauschenbach, and F. Giust, multimedia applications,” ACM Trans. Compt. Syst., vol. 24, no. 3,
“Mobile-edge computing architecture: The role of MEC in the Internet pp. 292–331, Aug 2006.
of Things.” IEEE Consum. Electron. Mag., vol. 5, no. 4, pp. 84–91, [80] K. D. Vogeleer, G. Memmi, P. Jouvelot, and F. Coelho, “The en-
Oct. 2016. ergy/frequency convexity rule: Modeling and experimental validation
[56] E. Ahmed and M. H. Rehmani, “Mobile edge computing: Opportu- on mobile devices,” in Proc. Springer Int. Conf. Parallel Process. Appl.
nities, solutions, and challenges,” Future Generation Comput. Syst., Mathematics (PPAM), Warsaw, Poland, Sep. 2013, pp. 793–803.
vol. 70, pp. 59–63, May 2017. [81] W. Zhang, Y. Wen, K. Guan, D. Kilper, H. Luo, and D. O. Wu, “Energy-
[57] A. u. R. Khan, M. Othman, S. A. Madani, and S. U. Khan, “A survey of optimal mobile cloud computing under stochastic wireless channel,”
mobile cloud computing application models,” IEEE Commun. Surveys IEEE Trans. Wireless Commun., vol. 12, no. 9, pp. 4569–4581, Sep.
Tuts., vol. 16, no. 1, pp. 393–413, 1st Quater 2014. 2013.
[58] A. P. Miettinen and J. K. Nurminen, “Energy efficiency of mobile [82] A. Carroll and G. Heiser, “An analysis of power consumption in a
clients in cloud computing,” in Proc. USENIX Conf. Hot Topics Cloud smartphone,” in Proc. USENIX Annual Technical Conf., Boston, MA,
Comput. (HotCloud), Boston, MA, Jun. 2010, pp. 1–7. Jun. 2010, pp. 1–14.
[59] S. Melendez and M. P. McGarry, “Computation offloading decisions [83] C. You, K. Huang, and H. Chae, “Energy efficient mobile cloud
for reducing completion time,” 2016. [Online]. Available: http: computing powered by wireless energy transfer,” IEEE J. Sel. Areas
//arxiv.org/pdf/1608.05839.pdf Commun., vol. 34, no. 5, pp. 1757–1771, May 2016.
[60] W. Yuan and K. Nahrstedt, “Energy-efficient soft real-time CPU [84] C. You, K. Huang, H. Chae, and B.-H. Kim, “Energy-efficient resource
scheduling for mobile multimedia systems,” in Proc. ACM Symp. allocation for mobile-edge computation offloading,” IEEE Trans. Wire-
Operat. Syst. Principles (SOSP), Bolton Landing, NY, USA, Oct. 2003, less Commun., vol. PP, no. 99, 2016.
pp. 149–163. [85] P. Barham, B. Dragovic, K. Fraser, S. Hand, T. Harris, A. Ho,
R. Neugebauer, I. Pratt, and A. Warfield, “Xen and the art of virtual-
[61] M. Jia, J. Cao, and L. Yang, “Heuristic offloading of concurrent tasks
ization,” in Proc. ACM Symp. Operat. Syst. Principles (SOSP), Bolton
for computation-intensive applications in mobile cloud computing,”
Landing, NY, Oct. 2003, pp. 164–177.
in Proc. IEEE Int. Conf. Comput. Commun. (INFOCOM WKSHPS),
[86] S. Barbarossa, S. Sardellitti, and P. Di Lorenzo, “Joint allocation of
Toronto, Canada, Apr. 2014, pp. 352–357.
computation and communication resources in multiuser mobile cloud
[62] S. E. Mahmoodi, R. N. Uma, and K. P. Subbalakshmi, “Optimal joint
computing,” in Proc. IEEE Int. Workshop Signal Process. Advances
scheduling and cloud offloading for mobile applications,” IEEE Trans.
Wireless Commun. (SPAWC), Darmstadt, Germany, Jun. 2013, pp. 26–
Cloud Comput., vol. PP, no. 99, pp. 1–13, 2016.
30.
[63] A. Goldsmith, Wireless Communications. New York, NY, USA: [87] X. Chen, L. Jiao, W. Li, and X. Fu, “Efficient multi-user computation
Cambridge University Press, 2005. offloading for mobile-edge cloud computing,” IEEE Trans. Netw.,
[64] D. Gesbert, S. Hanly, H. Huang, S. S. Shitz, O. Simeone, and W. Yu, vol. 24, pp. 2795–2808, Oct. 2016.
“Multi-cell MIMO cooperative networks: A new look at interference,” [88] X. Lyu, H. Tian, P. Zhang, and C. Sengul, “Multi-user joint task
IEEE J. Sel. Areas Commun., vol. 28, no. 9, pp. 1380–1408, Dec. 2010. offloading and resources optimization in proximate clouds,” IEEE
[65] S. A. Jafar, “Topological interference management through index Trans. Veh. Techn., vol. PP, no. 99, pp. 1–13, Jul. 2016.
coding,” IEEE Trans. Inf. Theory, vol. 60, no. 1, pp. 529–568, Jan [89] S. Vakilinia, M. M. Ali, and D. Qiu, “Modeling of the resource
2014. allocation in cloud computing centers,” Elsevier Comput. Netw., vol. 91,
[66] C. Li, J. Zhang, M. Haenggi, and K. B. Letaief, “User-centric intercell pp. 453–470, Nov. 2015.
interference nulling for downlink small cell networks,” IEEE Trans. [90] D. Bruneo, “A stochastic model to investigate data center performance
Commun., vol. 63, no. 4, pp. 1419–1431, Apr. 2015. and QoS in IaaS cloud computing systems,” IEEE Trans. Parallel
[67] E. Torkildson, U. Madhow, and M. Rodwell, “Indoor millimeter wave Distrib. Syst., vol. 25, no. 3, pp. 560–569, 2014.
MIMO: Feasibility and performance,” IEEE Trans. Wireless Commun., [91] X. Fan, W.-D. Weber, and L. A. Barroso, “Power provisioning for a
vol. 10, no. 12, pp. 4150–4160, Dec. 2011. warehouse-sized computer,” in ACM SIGARCH Comput. Archit. News,
[68] X. Yu, J. C. Shen, J. Zhang, and K. B. Letaief, “Alternating mini- vol. 35, no. 2, 2007, pp. 13–23.
mization algorithms for hybrid precoding in millimeter wave MIMO [92] C.-C. Lin, P. Liu, and J.-J. Wu, “Energy-efficient virtual machine
systems,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 485– provision algorithms for cloud systems,” in Proc. IEEE Utility and
500, Apr. 2016. Cloud Computing (UCC), Melbourne, Australia, Dec. 2011, pp. 81–
[69] S. M. Alamouti, “A simple transmit diversity technique for wireless 88.
communications,” IEEE J. Sel. Areas Commun., vol. 16, no. 8, pp. [93] A. Beloglazov, J. Abawajy, and R. Buyya, “Energy-aware resource
1451–1458, Aug. 1998. allocation heuristics for efficient management of data centers for cloud
[70] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, “Capacity computing,” Elsevier Future Generation Comput. Syst., vol. 28, no. 5,
limits of MIMO channels,” IEEE J. Sel. Areas Commun., vol. 51, no. 6, pp. 755–768, May 2012.
pp. 684–195, Jun. 2003. [94] K. Kumar and Y. H. Lu, “Cloud computing for mobile users: Can
[71] E. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta, “Massive MIMO offloading computation save energy?” Comput., vol. 43, no. 4, pp. 51–
for next generation wireless systems,” IEEE Commun. Mag., vol. 52, 56, Apr. 2010.
no. 2, pp. 186–195, Feb. 2014. [95] K. Kumar, J. Liu, Y.-H. Lu, and B. Bhargava, “A survey of computation
[72] J. G. Andrews, H. Claussen, M. Dohler, S. Rangan, and M. C. Reed, offloading for mobile systems,” Mobile Netw. Appl., vol. 18, no. 1, pp.
“Femtocells: Past, present, and future,” IEEE Commun. Mag., vol. 30, 129–140, Feb. 2013.
no. 3, pp. 497–508, Mar. 2012. [96] S. Barbarossa, S. Sardellitti, and P. D. Lorenzo, “Communicating while
[73] H. S. Dhillon, R. K. Ganti, F. Baccelli, and J. G. Andrews, “Modeling computing: Distributed mobile cloud computing over 5G heterogeneous
and analysis of K-tier downlink heterogeneous cellular networks,” IEEE networks,” IEEE Signal Process. Mag., vol. 31, no. 6, pp. 45–55, Nov.
Trans. Commun., vol. 30, no. 3, pp. 550–560, Mar. 2012. 2014.
 34

[97] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li, “Mobile-edge com- [119] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of radio
puting: Partial computation offloading using dynamic voltage scaling,” and computational resources for multicell mobile-edge computing,”
IEEE Trans. Commun., vol. 64, no. 10, pp. 4268–4282, Oct. 2016. IEEE Trans. Signal Info. Process. Over Networks, vol. 1, no. 2, pp.
[98] Y. H. Kao, B. Krishnamachari, M. R. Ra, and F. Bai, “Hermes: Latency 89–103, Jun. 2015.
optimal task assignment for resource-constrained mobile computing,” [120] M. Molina, O. Muñoz, A. Pascual-Iserte, and J. Vidal, “Joint scheduling
in Proc. IEEE Int. Conf. Comput. Commun. (INFOCOM), Hong Kong, of communication and computation resources in multiuser wireless
China, Apr. 2015, pp. 1894–1902. application offloading,” in Proc. IEEE Int. Symp. on Personal Indoor
[99] W. Zhang, Y. Wen, and D. O. Wu, “Collaborative task execution in and Mobile Radio Comm. (PIMRC), Washington, DC, Sep. 2014, pp.
mobile cloud computing under a stochastic wireless channel,” IEEE 1093–1098.
Trans. Wireless Commun., vol. 14, no. 1, pp. 81–93, Jan. 2015. [121] Y. Yu, J. Zhang, and K. B. Letaief, “Joint subcarrier and CPU time
[100] S. Khalili and O. Simeone, “Inter-layer per-mobile optimization of allocation for mobile edge computing,” in Proc. IEEE Global Commun.
cloud mobile computing: A message-passing approach.” [Online]. Conf. (GLOBECOM), Washington, DC, Dec. 2016, pp. 1–6.
Available: http://arxiv.org/abs/1509.01596 [122] L. Yang, J. Cao, H. Cheng, and Y. Ji, “Multi-user computation
[101] P. D. Lorenzo, S. Barbarossa, and S. Sardellitti, “Joint optimization partitioning for latency sensitive mobile cloud applications,” IEEE
of radio resources and code partitioning in mobile edge computing.” Trans. Comput., vol. 64, no. 8, pp. 2253–2266, Aug. 2015.
[Online]. Available: http://arxiv.org/abs/1307.3835v3 [123] Y. Li, L. Sun, and W. Wang, “Exploring device-to-device communica-
[102] S. E. Mahmoodi, K. P. Subbalakshmi, and V. Sagar, “Cloud offloading tion for mobile cloud computing,” in Proc. IEEE Int. Conf. Commun.
for multi-radio enabled mobile devices,” in Proc. IEEE Int. Conf. (ICC), Sydney, Australia, Jun. 2014, pp. 2239–2244.
Commun. (ICC), London, UK, Jun. 2015, pp. 5473–5478. [124] M. Jo, T. Maksymyuk, B. Strykhalyuk, and C.-H. Cho, “Device-
[103] D. Huang, P. Wang, and D. Niyato, “A dynamic offloading algorithm to-device-based heterogeneous radio access network architecture for
for mobile computing,” IEEE Trans. Wireless Commun., vol. 11, no. 6, mobile cloud computing,” IEEE Wireless Commun., vol. 22, no. 3, pp.
pp. 1991–1995, Jun. 2012. 50–58, Mar. 2015.
[104] J. Liu, Y. Mao, J. Zhang, and K. B. Letaief, “Delay-optimal compu- [125] Z. Sheng, C. Mahapatra, V. Leung, M. Chen, and P. Sahu, “Energy
tation task scheduling for mobile-edge computing systems,” in Proc. efficient cooperative computing in mobile wireless sensor networks,”
IEEE Int. Symp. Inf. Theory (ISIT), Barcelona, Spain, Jul 2016, pp. IEEE Trans. Cloud Comput., vol. PP, no. 99, pp. 1–13, Jul. 2015.
1451–1455. [126] J. Song, Y. Cui, M. Li, J. Qiu, and R. Buyya, “Energy-traffic trade-
[105] S. Chen, Y. Wang, and M. Pedram, “A semi-Markovian decision off cooperative offloading for mobile cloud computing,” in Proc.
process based control method for offloading tasks from mobile devices IEEE/ACM Int. Symp. Quality of Service (IWQoS), Hong Kong, China,
to the cloud,” in Proc. IEEE Global Commun. Conf. (GLOBECOM), May 2014, pp. 284–289.
Atlanta, GA, Dec 2013, pp. 2885–2890. [127] X. Cao, F. Wang, J. Xu, R. Zhang, and S. Cui, “Joint computation
[106] S.-T. Hong and H. Kim, “QoE-aware computation offloading schedul- and communication cooperation for mobile edge computing,” Apr.
ing to capture energy-latency tradeoff in mobile clouds,” in Proc. IEEE 2017. [Online]. Available: https://arxiv.org/pdf/1704.06777.pdf
Int. Conf. Sensing, Commun. Netw. (SECON), London, UK, Jun. 2016, [128] C. You and K. Huang, “Mobile cooperative computing: Energy-efficient
pp. 1–9. peer-to-peer computation offloading,” 2017. [Online]. Available:
https://arxiv.org/abs/1704.04595
[107] J. Kwak, Y. Kim, J. Lee, and S. Chong, “Dream: Dynamic resource
[129] L. Chen, S. Zhou, and J. Xu, “Computation peer offloading for
and task allocation for energy minimization in mobile cloud systems,”
energy-constrained mobile edge computing in small-cell networks,”
IEEE J. Sel. Areas Commun., vol. 33, no. 12, pp. 2510–2523, Dec
Mar. 2017. [Online]. Available: https://arxiv.org/pdf/1703.06058.pdf
2015.
[130] L. Lei, Z. Zhong, K. Zheng, J. Chen, and H. Meng, “Challenges on
[108] Z. Jiang and S. Mao, “Energy delay tradeoff in cloud offloading for
wireless heterogeneous networks for mobile cloud computing,” IEEE
multi-core mobile devices,” IEEE Access, vol. 3, pp. 2306–2316, Nov.
Wireless Commun., vol. 20, no. 3, pp. 34–44, 2013.
2015.
[131] T. Zhao, S. Zhou, X. Guo, Y. Zhao, and Z. Niu, “A cooperative
[109] D. T. Hoang, D. Niyato, and P. Wang, “Optimal admission control scheduling scheme of local cloud and Internet cloud for delay-aware
policy for mobile cloud computing hotspot with cloudlet,” in Proc. mobile cloud computing,” in Proc. IEEE Global Commun. Conf.
IEEE Wireless Commun. Networking Conf. (WCNC), Paris, France, Worshops (GC WKSHPS), San Diego, CA, Dec. 2015, pp. 1–6.
Apr. 2012, pp. 3145–3149. [132] Y. Ge, Y. Zhang, Q. Qiu, and Y.-H. Lu, “A game theoretic resource
[110] Y. Mao, J. Zhang, S. Song, and K. B. Letaief, “Power-delay tradeoff allocation for overall energy minimization in mobile cloud computing
in multi-user mobile-edge computing systems,” in Proc. IEEE Global system,” in Proc. ACM/IEEE Int. Symp. Low Power Electron. Design,
Commun. Conf. (GLOBECOM), Washington, DC, Dec. 2016, pp. 1–6. Redondo Beach, CA, Jul.-Aug. 2012, pp. 279–284.
[111] K. Wang, K. Yang, and C. Magurawalage, “Joint energy minimization [133] T. Q. Dinh, J. Tang, Q. D. La, and T. Q. Quek, “Offloading in mobile
and resource allocation in C-RAN with mobile cloud,” IEEE Trans. edge computing: Task allocation and computational frequency scaling,”
Cloud Comput., vol. PP, no. 99, pp. 1–11, Jan. 2016. Commun. ACM, vol. PP, no. 99, pp. 1–14, Apr. 2017.
[112] J. Ren, G. Yu, Y. Cai, and Y. He, “Latency optimization for [134] R. Kaewpuang, D. Niyato, P. Wang, and E. Hossain, “A framework for
resource allocation in mobile-edge computation offloading.” [Online]. cooperative resource management in mobile cloud computing,” IEEE
Available: https://arxiv.org/pdf/1704.00163.pdf J. Sel. Areas Commun., vol. 31, no. 12, pp. 2685–2700, Dec. 2013.
[113] M.-H. Chen, B. Liang, and M. Dong, “Joint offloading decision and [135] R. Yu, J. Ding, S. Maharjan, S. Gjessing, Y. Zhang, and D. Tsang,
resource allocation for multi-user multi-task mobile cloud,” in Proc. “Decentralized and optimal resource cooperation in geo-distributed
IEEE Int. Conf. Commun. (ICC), Kuala Lumpur, Malaysia, May 2016, mobile cloud computing,” IEEE Trans. Emerg. Topics Comput., vol. PP,
pp. 1–6. no. 99, pp. 1–13, Sep. 2015.
[114] M.-H. Chen, B. Liang, and D. Ming, “Joint offloading and resource [136] M. S. Elbamby, M. Bennis, and W. Saad, “Proactive edge
allocation for computation and communication in mobile cloud with computing in latency-constrained fog networks.” [Online]. Available:
computing access point,” in Proc. IEEE Int. Conf. Comput. Commun. https://arxiv.org/pdf/1704.06749.pdf
(INFOCOM), Atlanta, GA, USA, Apr. 2017, pp. 1863–1871. [137] S. Wang, R. Urgaonkar, T. He, M. Zafer, K. Chan, and K. K. Leung,
[115] X. Chen, “Decentralized computation offloading game for mobile cloud “Mobility-induced service migration in mobile micro-clouds,” in Proc.
computing,” IEEE Trans. Parallel Distrib. Syst., vol. 26, no. 4, pp. 974– IEEE Military Commun. Conf. (MILCOM), Baltimore, MD, Oct. 2014,
983, Apr. 2016. pp. 835–840.
[116] M.-H. Chen, B. Liang, and M. Dong, “Multi-user mobile cloud [138] R. Urgaonkar, S. Wang, T. He, M. Zafer, K. Chan, and K. K. Leung,
offloading game with computing access point,” in Proc. IEEE Int. Conf. “Dynamic service migration and workload scheduling in edge-clouds,”
Cloud Networking (Cloudnet), Pisa, Italy, Oct. 2016, pp. 64–69. Performance Evaluation, vol. 91, pp. 205–228, 2015.
[117] X. Ma, C. Lin, X. Xiang, and C. Chen, “Game-theoretic analysis of [139] M.-H. Chen, M. Dong, and B. Liang, “Joint offloading decision and
computation offloading for cloudlet-based mobile cloud computing,” in resource allocation for mobile cloud with computing access point,”
Proc. ACM Int. Conf. Modeling, Anal. and Simulation of Wireless and in Proc. IEEE Int. Conf. Accoustic, Speech, and Signal Processing
Mobile Syst. (MSWiM), Cancun, Mexico, Nov. 2015, pp. 271–278. (ICASSP), Shanghai, China, Mar. 2016, pp. 3516–3520.
[118] S. Guo, B. Xiao, Y. Yang, and Y. Yang, “Energy-efficient dynamic [140] O. Munoz, A. P-Iserte, and J. Vidal, “Optimization of radio and
offloading and resource scheduling in mobile cloud computing,” in computational resources for energy efficiency in latency-constrained
Proc. IEEE Int. Conf. Comput. Commun. (INFOCOM), San Francisco, application offloading,” IEEE Trans. Veh. Techn., vol. 64, no. 10, pp.
CA, Apr. 2016, pp. 1–9. 497–508, Oct. 2015.
 35

[141] S. Wang and S. Dey, “Modeling and characterizing user experience in [164] Y. Gao, H. Guan, Z. Qi, Y. Hou, and L. Liu, “A multi-objective
a cloud server based mobile gaming approach,” in Proc. IEEE Global ant colony system algorithm for virtual machine placement in cloud
Commun. Conf. (GLOBECOM), Honolulu, HI, Nov.-Dec. 2009, pp. computing,” J. Comput. Syst. Sci., vol. 79, no. 8, pp. 1230–1242, Dec.
1–7. 2013.
[142] S. Wang, M. Zafer, and K. K. Leung, “Online placement of multi- [165] J. L. Lucas-Simarro, R. Moreno-Vozmediano, R. S. Montero, and I. M.
component applications in edge computing environments,” IEEE Ac- Llorente, “Scheduling strategies for optimal service deployment across
cess, vol. 5, Feb. 2017. multiple clouds,” Future Generation Comput. Syst., vol. 29, no. 6, pp.
[143] Y. Shi, J. Zhang, B. O’Donoghue, and K. B. Letaief, “Large-scale 1431–1441, Aug. 2013.
convex optimization for dense wireless cooperative networks,” IEEE [166] H. Rheingold, Virtual Reality: Exploring the Brave New Technologies.
Trans. Signal Process., vol. 63, no. 18, pp. 4729–4743, Sep. 2013. New York, NY, USA: Simon & Schuster Adult Publishing Group, 1991.
[144] C. Vallati, A. Virdis, E. Mingozzi, and G. Stea, “Mobile-edge com- [167] S. Wang and S. Dey, “Modeling and characterizing user experience in
puting come home connecting things in future smart homes using a cloud server based mobile gaming approach,” in Proc. IEEE Global
LTE device-to-device communications,” IEEE Consum. Electron. Mag., Commun. Conf. (GLOBECOM), Honolulu, HI, Nov. 2009, pp. 1–7.
vol. 5, no. 4, pp. 77–83, Oct. 2016. [168] T. X. Tran, P. Pandey, A. Hajisami, and D. Pompili, “Collaborative
[145] T. H. Luan, L. Gao, Z. Li, Y. Xiang, G. Wei, and L. Sun, “Fog multi-bitrate video caching and processing in mobile-edge computing
computing: Focusing on mobile users at the edge.” [Online]. Available: networks,” in Proc. IEEE/IFIP Conf. Wireless On-demand Network
https://arxiv.org/pdf/1502.01815v3.pdf Systems and Services (WONS), Jackson Hole, WY, USA, Feb. 2017,
[146] L. Tong, Y. Li, and W. Gao, “A hierarchical edge cloud architecture pp. 1–8. [Online]. Available: https://arxiv.org/pdf/1612.01436v2.pdf
for mobile computing,” in Proc. IEEE Int. Conf. Comput. Commun.
[169] E. Baştuǧ, M. Bennis, M. Kountouris, and M. Debbah, “Cache-enabled
(INFOCOM), San Francisco, CA, USA, Apr. 2016, pp. 1–9.
small cell networks: Modeling and tradeoffs,” EURASIP J. Wireless
[147] G. Kirby, A. Dearle, A. Macdonald, and A. Fernandes, “An
Commun. Networking, vol. 2015, no. 1, pp. 1–11, Feb. 2015.
approach to ad hoc cloud computing.” [Online]. Available: https:
//arxiv.org/pdf/1002.4738v1.pdf [170] Y. Cui, Y. Wu, and D. Jiang, “Analysis and optimization of caching
[148] T. T. Huu, C. K. Tham, and D. Niyato, “A stochastic workload and multicasting in large-scale cache-enabled information-centric net-
distribution approach for an ad-hoc mobile cloud,” in Proc. IEEE Int. works,” in Proc. IEEE Global Commun. Conf. (GLOBECOM), San
Conf. Cloud Comput. Techn. Sci. (CloudCom), Singapore, Dec. 2014, Diego, CA, Dec. 2015, pp. 1–7.
pp. 174–181. [171] V. Suryaprakash, J. Møller, and G. Fettweis, “On the modeling and
[149] D. M. Shila, W. Shen, Y. Cheng, X. Tian, and X. Shen, “AMCloud: analysis of heterogeneous radio access networks using a poisson cluster
Toward a secure autonomic mobile ad hoc cloud computing system,” process,” IEEE Trans. Wireless Commun., vol. 14, no. 2, pp. 1035–
IEEE Wireless Commun., vol. PP, no. 99, pp. 1–8, Oct. 2016. 1047, Feb. 2015.
[150] X. Hou, Y. Li, M. Chen, D. Wu, D. Jin, and S. Chen, “Vehicular fog [172] D. Lopez-Perez, I. Guvenc, and X. Chu, “Mobility management
computing: A viewpoint of vehicles as the infrastructures,” IEEE Trans. challenges in 3GPP heterogeneous networks,” IEEE Commun. Mag.,
Veh. Techn., vol. 65, no. 6, pp. 3860–3873, Jun. 2016. vol. 50, no. 12, pp. 70–78, 2012.
[151] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and [173] A. Damnjanovic, J. Montojo, Y. Wei, T. Ji, T. Luo, M. Vajapeyam,
M. Franceschetti, “Stochastic geometry and random graphs for the T. Yoo, O. Song, and D. Malladi, “A survey on 3GPP heterogeneous
analysis and design of wireless networks,” IEEE J. Sel. Areas Commun., networks,” IEEE Wireless Commun., vol. 18, no. 3, pp. 10–21, Mar.
vol. 27, no. 7, pp. 1029–1046, Sep. 2009. 2011.
[152] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to [174] M. Kassar, B. Kervella, and G. Pujolle, “An overview of vertical
coverage and rate in cellular networks,” IEEE Trans. Commun., vol. 59, handover decision strategies in heterogeneous wireless networks,”
no. 11, pp. 3122–3134, Nov. 2011. ELSEVIER Comput. Commun., vol. 31, no. 10, pp. 2607–2620, Oct.
[153] M. Haenggi, Stochastic Geometry for Wireless Networks. New York, 2008.
NY, USA: Cambridge University Press, 2012. [175] C. Wang, Y. Li, and D. Jin, “Mobility-assisted opportunistic computa-
[154] C. Li, J. Zhang, J. G. Andrews, and K. B. Letaief, “Success probability tion offloading,” IEEE Commun. Lett., vol. 18, no. 10, pp. 1779–1782,
and area spectral efficiency in multiuser MIMO HetNets,” IEEE Trans. Oct. 2014.
Commun., vol. 64, no. 4, pp. 1544–1556, Apr. 2016. [176] Y. Zhang, D. Niyato, and P. Wang, “Offloading in mobile cloudlet
[155] N. Vastardis and K. Yang, “An enhanced community-based mobility systems with intermittent connectivity,” IEEE Trans. Mobile Comput.,
model for distributed mobile social networks,” J. Ambient Intelligence vol. 14, no. 12, pp. 2516–2529, Dec. 2015.
and Humanized Comput., vol. 5, no. 1, pp. 65–75, Feb. 2014. [177] K. Lee and I. Shin, “User mobility model based computation offloading
[156] CISCO, “Cisco visual networking index: Global mobile decision for mobile cloud,” J. Comput. Sci. Eng., vol. 9, no. 3, pp. 155–
data traffic forecast update, 2015-2020,” White paper, 2016. 162, Sep. 2015.
[Online]. Available: http://www.cisco.com/c/en/us/solutions/collateral/ [178] M. R. Rahimi, N. Venkatasubramanian, and A. V. Vasilakos, “Music:
service-provider/visual-networking-index-vni/vni-forecast-qa.pdf Mobility-aware optimal service allocation in mobile cloud computing,”
[157] X. Wang, M. Chen, T. Taleb, A. Ksentini, and V. C. Leung, “Cache in Proc. IEEE Int. Conf. Cloud Comput. (CLOUD), Santa Clara
in the air: Exploiting content caching and delivery techniques for 5G Marriott, CA, Jun. 2013, pp. 75–82.
systems,” IEEE Commun. Mag., vol. 52, no. 2, pp. 131–139, Feb. 2014. [179] A. Prasad, P. Lundén, M. Moisio, M. A. Uusitalo, and Z. Li, “Efficient
[158] E. Bastug, M. Bennis, and M. Debbah, “Living on the edge: The role mobility and traffic management for delay tolerant cloud data in 5G
of proactive caching in 5G wireless networks,” IEEE Commun. Mag., networks networks,” in Proc. IEEE Int. Symp. on Personal Indoor and
vol. 52, no. 8, pp. 82–89, Aug. 2014. Mobile Radio Comm. (PIMRC), 2015, pp. 1740–1745.
[159] N. Golrezaei, K. Shanmugam, A. G. Dimakis, A. F. Molisch, and
[180] R. Wang, X. Peng, J. Zhang, and K. B. Letaief, “Mobility-aware
G. Caire, “Femtocaching: Wireless video content delivery through
caching for content-centric wireless networks: Modeling and method-
distributed caching helpers,” in Proc. IEEE Int. Conf. Compt. Commun.
ology,” IEEE Commun. Mag., vol. 54, no. 8, pp. 77–83, Aug. 2016.
(INFOCOM), Orlando, FL, Mar. 2012, pp. 1107–1115.
[160] A. S. Gomes, B. Sousa, D. Palma, V. Fonseca, Z. Zhao, E. Monteiro, [181] S.-W. Ko, K. Huang, S.-L. Kim, and H. Chae, “Online prefetching
T. Braun, P. Simoes, and L. Cordeiro, “Edge caching with mobility for mobile computation offloading,” 2016. [Online]. Available:
prediction in virtualized LTE mobile networks,” Future Generation https://arxiv.org/abs/1608.04878
Comput. Syst., vol. 70, pp. 148–162, 2017. [182] K. Doppler, M. Rinne, C. Wijting, C. B. Ribeiro, and K. Hugl, “Device-
[161] L. Yang, J. Cao, G. Liang, and X. Han, “Cost aware service placement to-device communication as an underlay to LTE-advanced networks,”
and load dispatching in mobile cloud systems,” IEEE Trans. Comput., IEEE Commun. Mag., vol. 47, no. 12, pp. 42–49, Dec 2009.
vol. 65, no. 5, pp. 1440–1452, May 2016. [183] C.-A. Chen, M. Won, R. Stoleru, and G. G. Xie, “Energy-efficient fault-
[162] J. Tordsson, R. S. Montero, R. Moreno-Vozmediano, and I. M. Llorente, tolerant data storage and processing in mobile cloud,” IEEE Trans.
“Cloud brokering mechanisms for optimized placement of virtual Cloud Comput., vol. 3, no. 1, pp. 28–41, Jan. 2015.
machines across multiple providers,” Future Generation Comput. Syst., [184] C. A. Chen, R. Stoleru, and G. G. Xie, “Energy-efficient and fault-
vol. 28, no. 2, pp. 358–367, Feb. 2012. tolerant mobile cloud storage,” in Proc. IEEE Int. Conf. Could Net-
[163] B. Li, J. Li, J. Huai, T. Wo, Q. Li, and L. Zhong, “EnaCloud: An working (Cloudnet), Pisa, Italy, Oct. 2016, pp. 51–57.
energy-saving application live placement approach for cloud computing [185] D. Satria, D. Park, and M. Jo, “Recovery for overloaded mobile edge
environments,” in Proc. IEEE Int. Conf. Cloud Comput. (CLOUD), computing,” Future Generation Comput. Syst., vol. 70, pp. 138–147,
Bangalore, India, Sep. 2009, pp. 17–24. 2017.
 36

[186] S. Chaisiri, B.-S. Lee, and D. Niyato, “Optimization of resource [210] T. Han and N. Ansari, “On optimizing green energy utilization for
provisioning cost in cloud computing,” IEEE Trans. Serv. Comput., cellular networks with hybrid energy supplies,” IEEE Trans. Wireless
vol. 5, no. 2, pp. 164–177, Apr. 2012. Commun., vol. 12, no. 8, pp. 3872–3882, 2013.
[187] Y. Zhang, J. Yan, and X. Fu, “Reservation-based resource scheduling [211] Y. Mao, J. Zhang, and K. B. Letaief, “Grid energy consumption and
and code partition in mobile cloud computing,” in Proc. IEEE Int. Conf. QoS tradeoff in hybrid energy supply wireless networks,” IEEE Trans.
Compt. Commun. Workshops (INFOCOM WKSHPS), San Francisco, Wireless Commun., vol. 15, no. 5, pp. 3573 – 3586, May 2016.
CA, Apr. 2016, pp. 962–967. [212] W. C. Brown, “The history of power transmission by radio waves,”
[188] X. Jin, F. Zhang, A. V. Vasilakos, and Z. Liu, “Green data centers: IEEE Trans. Microw. Theory Techn., vol. 32, no. 9, pp. 1230–1242,
A survey, perspectives, and future directions.” [Online]. Available: Sep. 1984.
http://arxiv.org/abs/1608.00687 [213] H. Ju and R. Zhang, “Throughput maximization in wireless powered
[189] X. Sun and N. Ansari, “Green cloudlet network: A distributed green communication networks,” IEEE Trans. Wireless Commun., vol. 13,
mobile cloud network,” IEEE Netw., to appear. no. 1, pp. 418–428, Jan. 2014.
[190] L. A. Barroso and U. Holzle, “The case for energy-proportional [214] K. Huang and V. K. N. Lau, “Enabling wireless power transfer
computing,” Comput., vol. 40, no. 12, pp. 651–664, Dec 2007. in cellular networks: Architecture, modeling and deployment,” IEEE
[191] M. Lin, A. Wierman, L. L. H. Andrew, and E. Thereska, “Dynamic Trans. Wireless Commun., vol. 13, no. 2, pp. 902–912, May 2014.
right-sizing for power-proportional data centers,” IEEE/ACM Trans. [215] Z. Chang, J. Gong, Y. Li, Z. Zhou, T. Ristaniemi, G. Shi, Z. Han, and
Netw., vol. 21, no. 5, pp. 1378–1391, Oct. 2013. Z. Niu, “Energy efficient resource allocation for wireless power transfer
[192] M. Lin, Z. Liu, A. Wierman, and L. L. H. Andrew, “Online algorithms enabled collaborative mobile clouds,” IEEE J. Sel. Areas Commun.,
for geographical load balancing,” in Proc. IEEE Int. Green Comput. vol. 34, no. 12, pp. 3438–3450, Dec. 2016.
Conf. (IGCC), San Jose, CA, Jun 2012, pp. 1–10. [216] R. Roman, J. Lopez, and M. Mambo, “Mobile edge computing, fog et
[193] H. Xu, C. Feng, and B. Li, “Temperature aware workload management al.: A survey and analysis of security threats and challenges,” Elsevier
in geo-distributed data centers,” IEEE Trans. Parallel Distrib. Syst., Future Generation Comput. Syst., vol. PP, no. 99, Nov. 2016.
vol. 26, no. 6, pp. 1743–1753, Jun. 2015. [217] S. Yi, Z. Qin, and Q. Li, “Security and privacy issues of fog computing:
[194] A. Beloglazov and R. Buyya, “Energy efficient resource management in A survey,” in Proc. Int. Conf. Wireless Algorithms, Systems, and
virtualized cloud data centers,” in Proc. IEEE/ACM Int. Conf. Cluster, Applications (WASA), Qufu, China, Aug. 2015, pp. 1–10.
Cloud and Grid Computing (CCGrid), Melbourne, Australia, May [218] M. M. Fouda, Z. M. Fadlullah, N. Kato, R. Lu, and X. S. Shen, “A
2010, pp. 826–831. lightweight message authentication scheme for smart grid communi-
[195] X. Li, J. Wu, S. Tang, and S. Lu, “Let’s stay together: Towards traffic cations,” IEEE Trans. Smart Grid, vol. 2, no. 4, pp. 675–685, Dec.
aware virtual machine placement in data centers,” in Proc. IEEE Int. 2011.
Conf. Comput. Commun. (INFOCOM), Toronto, Canada, Apr. 2014, [219] A. M. Y. Ahmed and D. Qian, “An optimization of security and
pp. 1842–1850. trust management in distributed systems,” in Proc. IEEE Int. Advance
[196] L. Chen and H. Shen, “Consolidating complementary VMs with Computing Conf. (IACC), Ghaziabad, India, Feb. 2013, pp. 120–126.
spatial/temporal-awareness in cloud datacenters,” in Proc. IEEE Int. [220] X. Huang, Y. Xiang, E. Bertino, J. Zhou, and L. Xu, “Robust multi-
Conf. Comput. Commun. (INFOCOM), Apr. 2014, pp. 1033–1041. factor authentication for fragile communications,” IEEE Trans. De-
[197] Z. Han, H. Tan, G. Chen, R. Wang, Y. Chen, and F. C. M. Lau, “Dy- pendable Secure Comput., vol. 11, no. 6, pp. 568–581, Nov.-Dec. 2014.
namic virtual machine management via approximate Markov decision [221] M. C. Gorantla, C. Boyd, and J. M. G. Nieto, “Attribute-based
process,” in Proc. IEEE Int. Conf. Comput. Commun. (INFOCOM), authenticated key exchange,” in Proc. Australian Conf. Info. Security
San Francisco, CA, Apr. 2016, pp. 1–9. and Privacy (ACISP), Sydney, Australia, Jul. 2010, pp. 1–25.
[198] S. Sudevalayam and P. Kulkarni, “Energy harvesting sensor nodes: [222] H. M. Pimentel, S. Kopp, M. A. S. Jr., R. M. Silveira, and G. Bressan,
Survey and implications,” IEEE Commun. Surveys Tuts., vol. 13, no. 3, “OCP: A protocol for secure communication in federated content
pp. 443–461, 3rd Quater 2011. networks,” Compt. Commun., vol. 68, pp. 47–60, Sep. 2015.
[199] S. Ulukus, A. Yener, E. Erkip, O. Simeone, M. Zorzi, P. Grover, and [223] M. Liyanage, A. B. Abro, M. Ylianttila, and A. Gurtov, “Opportuni-
K. Huang, “Energy harvesting wireless communications: A review of ties and challenges of software-defined mobile networks in network
recent advances,” IEEE J. Sel. Areas Commun., vol. 33, no. 3, pp. security,” IEEE Security Privacy, vol. 14, no. 4, pp. 34–44, Jul. 2016.
360–381, Mar. 2015. [224] W. Yang and C. Fung, “A survey on security in network functions
[200] Y. Mao, Y. Luo, J. Zhang, and K. B. Letaief, “Energy harvesting small virtualization,” in Proc. IEEE NetSoft Conf. Workshops (NetSoft),
cell networks: Feasibility, deployment, and operation,” IEEE Commun. Seoul, Korea, Jun. 2016, pp. 15–19.
Mag., vol. 53, no. 6, pp. 94–101, Jun 2015. [225] B. Liang, “Mobile edge computing,” in Key Technologies for 5G
[201] CNN, “Battery life concerns mobile users,” Sep. 2005. [Online]. Avail- Wireless Systems, V. W. S. Wong, R. Schober, D. W. K. Ng, and
able: http://edition.cnn.com/2005/TECH/ptech/09/22/phone.study/ L.-C. Wang, Eds. Cambridge University Press, 2017. [Online].
[202] J. Xu and S. Ren, “Online learning for offloading and autoscaling in Available: http://www.comm.utoronto.ca/∼liang/publications/Chapter
renewable-powered mobile edge computing,” in Proc. IEEE Global MEC 2016.pdf
Commun. Conf. (GLOBECOM), Washington, DC, Dec. 2016, pp. 1–6. [226] Alcatel-Lucent, “Providing security in NFV: Challenges and opportu-
[203] Y. Mao, J. Zhang, and K. B. Letaief, “Dynamic computation offloading nities.” [Online]. Available: http://www.tmcnet.com/tmc/whitepapers/
for mobile-edge computing with energy harvesting devices,” IEEE J. documents/whitepapers/2014/10172-providing-security-nfv.pdf
Sel. Areas Commun., vol. 34, no. 12, pp. 3590–3605, Dec. 2016. [227] C. Wang, K. Ren, and J. Wang, “Secure optimization computation
[204] C. Chen, B. He, and X. Tang, “Green-aware workload scheduling in outsourcing in cloud computing: A case study of linear programming,”
geographically distributed data centers,” in Proc. IEEE Int. Conf. Cloud IEEE Trans. Comput., vol. 65, no. 1, pp. 216–229, Jan. 2016.
Comput. Tech. Sci. (CloudCom), Taipei, Taiwan, Nov. 2012, pp. 82–89. [228] R. Gennaro, G. Craig, and P. Bryan, “Non-interactive verifiable com-
[205] C. Dong, F. Kong, X. Liu, and H. Zeng, “Green power analysis for puting: Outsourcing computation to untrusted workers,” in Annu. Conf.
geographical load balancing based datacenters,” in Proc. IEEE Int. Adv. Cryptol., Santa Barbara, CA, USA, Aug. 2010, pp. 465–482.
Green Comput. Conf. (IGCC), Arlington, VA, Jun. 2013, pp. 1–8. [229] ETSI, “Executive briefing - mobile edge computing (MEC) initiative.”
[206] X. Sun, N. Ansari, and Q. Fan, “Green energy aware avatar migration [Online]. Available: https://portal.etsi.org/portals/0/tbpages/mec/docs/
strategy in green cloudlet networks,” in Proc. IEEE Int. Conf. Cloud mec%20executive%20brief%20v1%2028-09-14.pdf
Comput. Tech. Sci. (CloudCom), Vancouver, Canada, Nov. 2015, pp. [230] ——, “Mobile-edge computing (MEC): Terminology.” [Online].
139–146. Available: http://www.etsi.org/deliver/etsi gs/MEC/001 099/001/01.
[207] T. Chen, Y. Zhang, X. Wang, and G. B. Giannakis, “Robust workload 01.01 60/gs MEC001v010101p.pdf
and energy management for sustainable data centers,” IEEE J. Sel. [231] ——, “Mobile-edge computing (MEC): Framework and reference
Areas Commun., vol. 34, no. 3, pp. 651–664, Mar. 2016. architecture.” [Online]. Available: http://www.etsi.org/deliver/etsi gs/
[208] Y. Luo, J. Zhang, and K. B. Letaief, “Transmit power minimization MEC/001 099/003/01.01.01 60/gs MEC003v010101p.pdf
for wireless networks with energy harvesting relays,” IEEE Trans. [232] ——, “Mobile-edge computing (MEC): Service scenarios.”
Commun., vol. 64, no. 3, pp. 987–1000, Mar. 2016. [Online]. Available: http://www.etsi.org/deliver/etsi gs/MEC/001 099/
[209] J. Gong, S. Zhou, and Z. Niu, “Optimal power allocation for energy 002/01.01.01 60/gs MEC002v010101p.pdf
harvesting and power grid coexisting wireless communication systems,” [233] S. Antipolis, “ETSI first mobile edge computing proof of concepts at
IEEE Trans. Commun., vol. 61, no. 7, pp. 3040 – 3049, Jul. 2013. MEC World Congress,” Sep. 2016.
 37

[234] N. Sprecher, J. Friis, R. Dolby, and J. Reister,

“Edge computing prepares for a multi-access future,” Sep.
2016. [Online]. Available: http://www.telecomtv.com/articles/mec/
edge-computing-prepares-for-a-multi-access-future-13986/
[235] 3GPP, “Technical specification group services and system
aspects; system architecture for the 5g systems; stage 2
(release 15),” 3GPP TS 23.501 V0.4.0, Apr. 2017. [On-
line]. Available: https://portal.3gpp.org/desktopmodules/Specifications/
SpecificationDetails.aspx?specificationId=3144
[236] A. Anjum, T. Abdullah, M. Tariq, Y. Baltaci, and N. Antonopoulos,
“Video stream analysis in clouds: An object detection and classification
framework for high performance video analytics,” IEEE Trans. Cloud
Comput., vol. PP, no. 99, pp. 1–1, Jan. 2016.
[237] Intel, “Intel mobile edge computing technology improves the
augmented reality experience,” Sep. 2016. [Online]. Available:
https://www.youtube.com
[238] P. Papadimitratos, A. L. Fortelle, K. Evenssen, R. Brignolo, and
S. Cosenza, “Vehicular communication systems: Enabling technologies,
applications, and future outlook on intelligent transportation,” IEEE
Commun. Mag., vol. 47, no. 11, pp. 84–95, Nov. 2009.
[239] N. Lu, N.Cheng, N. Zhang, X. Shen, and J. W. Mark, “Connected
vehicles: Solutions and challenges,” IEEE Internet Things J., vol. 1,
no. 4, pp. 289–299, Apr. 2015.
[240] E. Uhlemann, “Introducing connected vehicles,” IEEE Veh. Techn.
Mag., vol. 10, no. 1, pp. 23–31, Jan. 2015.
[241] NOKIA, “UTM infrastructure and connected society,” 2016.
[Online]. Available: http://rpas-civops.com/wp-content/uploads/2016/
11/S7.2 Nokia DE V1.pdf
[242] Saguna, “Executive briefing - mobile edge computing (MEC)
initiative.” [Online]. Available: http://www.saguna.net/site/assets/
files/1723/saguna intel mec wp.pdf?utm source=homepage?utm
medium=button3?utm campaign=Q1-2016
[243] Ericsson, “Policy control function in 5G,” Jan. 2017.
[Online]. Available: http://portal.3gpp.org/ngppapp/CreateTdoc.aspx?
mode=view&contributionId=756820#
[244] N. Alliance, “Description of network slicing concept,” Jan. 2016.
[Online]. Available: https://www.ngmn.org/uploads/media/160113
Network Slicing v1 0.pdf
[245] A. Nakao, “Network softwarization and slicing: Ongoing developments
in standard developing organizations,” Oct.-Nov. 2016. [Online].
Available: http://cscn2016.ieee-cscn.org/document.pdf