1628 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO.

3, THIRD QUARTER 2017

Mobile Edge Computing: A Survey on Architecture

and Computation Offloading
Pavel Mach, Member, IEEE, and Zdenek Becvar, Member, IEEE

Abstract—Technological evolution of mobile user equip- huge processing in a short time. Moreover, high battery con-
ment (UEs), such as smartphones or laptops, goes hand-in-hand sumption still poses a significant obstacle restricting the users
with evolution of new mobile applications. However, running to fully enjoy highly demanding applications on their own
computationally demanding applications at the UEs is con-
strained by limited battery capacity and energy consumption devices. This motivates development of mobile cloud com-
of the UEs. A suitable solution extending the battery life-time of puting (MCC) concept allowing cloud computing for mobile
the UEs is to offload the applications demanding huge process- users [1]. In the MCC, a user equipment (UE) may exploit
ing to a conventional centralized cloud. Nevertheless, this option computing and storage resources of powerful distant cen-
introduces significant execution delay consisting of delivery of tralized clouds (CC), which are accessible through a core
the offloaded applications to the cloud and back plus time of the
computation at the cloud. Such a delay is inconvenient and makes network (CN) of a mobile operator and the Internet. The MCC
the offloading unsuitable for real-time applications. To cope with brings several advantages [2]; 1) extending battery lifetime
the delay problem, a new emerging concept, known as mobile by offloading energy consuming computations of the applica-
edge computing (MEC), has been introduced. The MEC brings tions to the cloud, 2) enabling sophisticated applications to the
computation and storage resources to the edge of mobile network mobile users, and 3) providing higher data storage capabilities
enabling it to run the highly demanding applications at the UE
while meeting strict delay requirements. The MEC computing to the users. Nevertheless, the MCC also imposes huge addi-
resources can be exploited also by operators and third parties tional load both on radio and backhaul of mobile networks and
for specific purposes. In this paper, we first describe major use introduces high latency since data is sent to powerful farm of
cases and reference scenarios where the MEC is applicable. After servers that are, in terms of network topology, far away from
that we survey existing concepts integrating MEC functionali- the users.
ties to the mobile networks and discuss current advancement in
standardization of the MEC. The core of this survey is, then, To address the problem of a long latency, the cloud ser-
focused on user-oriented use case in the MEC, i.e., computation vices should be moved to a proximity of the UEs, i.e., to the
offloading. In this regard, we divide the research on computation edge of mobile network as considered in newly emerged edge
offloading to three key areas: 1) decision on computation offload- computing paradigm. The edge computing can be understood
ing; 2) allocation of computing resource within the MEC; and as a specific case of the MCC. Nevertheless, in the conven-
3) mobility management. Finally, we highlight lessons learned
in area of the MEC and we discuss open research challenges tional MCC, the cloud services are accessed via the Internet
yet to be addressed in order to fully enjoy potentials offered connection [3] while in the case of the edge computing, the
by the MEC. computing/storage resources are supposed to be in proxim-
Index Terms—Mobile edge computing, mobile network ity of the UEs (in sense of network topology). Hence, the
architecture, computation offloading, allocation of computing MEC can offer significantly lower latencies and jitter when
resources, mobility management, standardization, use-cases. compared to the MCC. Moreover, while the MCC is fully
centralized approach with farms of computers usually placed
I. I NTRODUCTION at one or few locations, the edge computing is supposed to
be deployed in fully distributed manner. On the other hand,
HE USERS’ requirements on data rates and quality of
T service (QoS) are exponentially increasing. Moreover,
technological evolution of smartphones, laptops and tablets
the edge computing provides only limited computational and
storage resources with respect to the MCC. A high level com-
parison of key technical aspects of the MCC and the edge
enables to emerge new high demanding services and appli- computing is outlined in Table I.
cations. Although new mobile devices are more and more The first edge computing concept bringing the com-
powerful in terms of central processing unit (CPU), even putation/storage closer to the UEs, proposed in 2009, is
these may not be able to handle the applications requiring cloudlet [4]. The idea behind the cloudlet is to place com-
Manuscript received October 28, 2016; revised February 17, 2017; accepted puters with high computation power at strategic locations in
March 10, 2017. Date of publication March 15, 2017; date of current ver- order to provide both computation resources and storage for
sion August 21, 2017. This work was supported by the Czech Technical the UEs in vicinity. The cloudlet concept of the computing
University in Prague under Grant SGS17/184/OHK3/3T/13. (Corresponding
author: Pavel Mach.) “hotspots” is similar to WiFi hotspots scenario, but instead
The authors are with the Department of Telecommunication Engineering, of Internet connectivity the cloudlet enables cloud services
Faculty of Electrical Engineering, Czech Technical University in to the mobile users. The fact that cloudlets are supposed to
Prague, 166 27 Prague, Czech Republic (e-mail: [email protected];
[email protected]). be mostly accessed by the mobile UEs through WiFi connec-
Digital Object Identifier 10.1109/COMST.2017.2682318 tion is seen as a disadvantage since the UEs have to switch
1553-877X  c 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1629

TABLE I
H IGH L EVEL C OMPARISON OF MCC AND E DGE C OMPUTING C ONCEPTS Remote Heads (RRHs) to centralized baseband units (BBUs).
The BBU’s computation power is, then, pooled together into
virtualized resources that are able to serve tens, hundreds or
even thousands of RRHs. Although the computation power of
this virtualized BBU pool is exploited primarily for a cen-
tralized control and baseband processing it may also be used
for the computation offloading to the edge of the network
(see [23]).
Another concept integrating the edge computing into the
mobile network architecture is developed by newly created
(2014) industry specification group (ISG) within European
between the mobile network and WiFi whenever the cloudlet Telecommunications Standards Institute (ETSI) [24]. The solu-
services are exploited [2]. Moreover, QoS (Quality of Service) tion outlined by ETSI is known as Mobile Edge Computing
of the mobile UEs is hard to fulfill similarly as in case of (MEC). The standardization efforts relating the MEC are
the MCC, since the cloudlets are not an inherent part of driven by prominent mobile operators (e.g., DOCOMO,
the mobile network and coverage of WiFi is only local with Vodafone, TELECOM Italia) and manufactures (e.g., IBM,
limited support of mobility. Nokia, Huawei, Intel). The main purpose of ISG MEC group
The other option enabling cloud computing at the edge is to is to enable an efficient and seamless integration of the cloud
perform computing directly at the UEs through ad-hoc cloud computing functionalities into the mobile network, and to help
allowing several UEs in proximity to combine their compu- developing favorable conditions for all stakeholders (mobile
tation power and, thus, process high demanding applications operators, service providers, vendors, and users).
locally [5]–[14]. To facilitate the ad-hoc cloud, several critical Several surveys on cloud computing have been published so
challenges need to be addressed: 1) finding proper computing far. Khan et al. [3] survey MCC application models and high-
UEs in proximity while guaranteeing that processed data will light their advantages and shortcomings. In [25], a problem
be delivered back to the source UE, 2) coordination among of a heterogeneity in the MCC is tackled. The heterogene-
the computing UEs has to be enabled despite the fact that ity is understood as a variability of mobile devices, different
there are no control channels to facilitate reliable computing, cloud vendors providing different services, infrastructures,
3) the computing UEs has to be motivated to provide their platforms, and various communication medium and technolo-
computing power to other devices given the battery consump- gies. The paper identifies how this heterogeneity impacts the
tion and additional data transmission constraints, 4) security MCC and discusses related challenges. Wen et al. [26] sur-
and privacy issues. vey existing efforts on Cloud Mobile Media, which provides
A more general concept of the edge computing, when com- rich multimedia services over the Internet and mobile wire-
pared to cloudlets and ad-hoc clouds, is known as a fog less networks. All above-mentioned papers focus, in general,
computing. The fog computing paradigm (shortly often abbre- on the MCC where the cloud is not allocated specifically at
viated as Fog in literature) has been introduced in 2012 by the edge of mobile network, but it is accessed through the
Cisco to enable a processing of the applications on billions of Internet. Due to a wide potential of the MEC, there is a lot
connected devices at the edge of network [15]. Consequently, of effort both in industry and academia focusing on the MEC
the fog computing may be considered as one of key enablers in particular. Despite this fact, there is just one survey focus-
of Internet of Things (IoT) and big data applications [16] as ing primarily on the MEC [27] that, however, only briefly
it offers: 1) low latency and location awareness due to prox- surveys several research works dealing with the MEC and
imity of the computing devices to the edge of the network, presents taxonomy of the MEC by describing key attributes.
2) wide-spread geographical distribution when compared to Furthermore, Roman et al. [28] extensively surveys security
the CC; 3) interconnection of very large number of nodes issues for various edge computing concepts. On top of that,
(e.g., wireless sensors), and 4) support of streaming and Luon et al. [29] dedicate one chapter to the edge comput-
real time applications [15]. Moreover, the characteristics of ing, where applications of economic and pricing models are
the fog computing can be exploited in many other applica- considered for resource management in the edge computing.
tions and scenarios such as smart grids, connected vehicles In contrast to the above-mentioned surveys, we describe
for Intelligent Transport Systems (ITS) or wireless sensor key use cases and scenarios for the MEC (Section II). Then,
networks [17]–[20]. we survey existing MEC concepts proposed in the literature
From the mobile users’ point of view, the most notable integrating the MEC functionalities into the mobile networks
drawback of all above-mentioned edge computing concepts and we discuss standardization of the MEC (Section III).
is that QoS and QoE (Quality of Experience) for users can be After that, the core part of the paper is focused on techni-
hardly guaranteed, since the computing is not integrated into cal works dealing with computation offloading to the MEC.
an architecture of the mobile network. One concept integrat- On one hand, the computation offloading can be seen as a
ing the cloud capabilities into the mobile network is Cloud key use case from the user perspective as it enables run-
Radio Access Network (C-RAN) [21]. The C-RAN exploits ning new sophisticated applications at the UE while reducing
the idea of distributed protocol stack [22], where some lay- its energy consumption (see [30]–[36] where computation
ers of the protocol stack are moved from distributed Radio offloading to distant CC is assumed). On the other hand,

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1630 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

Fig. 1. Example of use cases and scenarios for the MEC.

the computation offloading brings several challenges, such II. U SE C ASES AND S ERVICE S CENARIOS
as selection of proper application and programming models, The MEC brings many advantages to all stakeholders, such
accurate estimation of energy consumption, efficient man- as mobile operators, service providers or users. As suggested
agement of simultaneous offloading by multiple users, or in [38] and [39], three main use case categories, depending
virtual machine (VM) migration [37]. In this respect, we on the subject to which they are profitable to, can be distin-
overview several general principles related to the computa- guished for the MEC (see Fig. 1). The next subsections discuss
tion offloading, such as offloading classification (full, partial individual use case categories and pinpoint several key service
offloading), factors influencing the offloading itself, and man- scenarios and applications.
agement of the offloading in practice (Section IV). Afterwards,
we sort the efforts within research community addressing
following key challenges regarding computation offloading A. Consumer-Oriented Services
into the MEC: The first use case category is consumer-oriented and, hence,
• A decision on the computation offloading to the MEC should be beneficial directly to the end-users. In general, the
with the purpose to determine whether the offloading is users profit from the MEC mainly by means of the com-
profitable for the UE in terms of energy consumption putation offloading, which enables running new emerging
and/or execution delay (Section V). applications at the UEs. One of the applications benefiting
• An efficient allocation of the computing resources from the computation offloading is a Web accelerated browser,
within the MEC if the computation is offloaded in order to where most of the browsing functions (Web contents evalua-
minimize execution delay and balance load of both com- tion, optimized transmission, etc.) are offloaded to the MEC;
puting resources and communication links (Section VI). see experimental results on offloading of Web accelerated
• Mobility management for the applications offloaded browser to the MEC in [40]. Moreover, face/speech recog-
to the MEC guaranteeing service continuity if the nition or image/video editing are also suitable for the MEC as
UEs exploiting the MEC roams throughout the network these require large amount of computation and storage [41].
(Section VII). Besides, the computation offloading to the MEC can be
Moreover, we summarize the lessons learned from state exploited by the applications based on augmented, assisted
of the art focused on computation offloading to the MEC or virtual reality. These applications derive additional infor-
(Section VIII) and outline several open challenges, which need mation about users’ neighborhood by performing an analysis
to be addressed to make the MEC beneficial for all stakehold- of their surroundings (e.g., visiting tourists may find points of
ers (Section IX). Finally, we summarize general outcomes and interest in his/her proximity). This may require fast responses,
draw conclusions (Section X). and/or significant amount of computing resources not available

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1631

at the UE. An applicability of the MEC for augmented real- the mobile edge. This way, the MEC application can store the
ity is shown in [42]. The authors demonstrate on a real MEC most popular content used in its geographical area. If the con-
testbed that the reduction of latency up to 88% and energy tent is requested by the users, it does not have to be transfered
consumption of the UE up to 93% can be accomplished by over the backhaul network.
the computation offloading to the MEC. Besides alleviation and optimization of the backhaul
On top of that, the users running low latency applications, network, the MEC can also help in radio network optimiza-
such as online gaming or remote desktop, may profit from tion. For example, gathering related information from the UEs
the MEC in proximity. In this case a new instance of a spe- and processing these at the edge will result in more effi-
cific application is initiated at an appropriate mobile edge cient scheduling. In addition, the MEC can also be used for
host to reduce the latency and resources requirements of the mobile video delivery optimization using throughput guidance
application at the UE. for TCP (Transmission Control Protocols). The TCP has an
inherent difficulty to adapt to rapidly varying condition on
B. Operator and Third Party Services radio channel resulting in an inefficient use of the resources.
To deal with this problem, the analytic MEC application can
The second use case category is represented by the services
provide a real-time indication on an estimated throughput to
from which operators and third parties can benefit. An example
a backend video server in order to match the application-level
of the use case profitable for the operator or third party is a
coding to the estimated throughput.
gathering of a huge amount of data from the users or sensors.
Such data is first pre-processed and analyzed at the MEC.
The pre-processed data is, then, sent to distant central servers III. MEC A RCHITECTURE AND S TANDARDIZATION
for further analysis. This could be exploited for safety and This section introduces and compares several concepts for
security purposes, such as monitoring of an area (e.g., car the computation at the edge integrated to the mobile network.
park monitoring). First, we overview various MEC solutions proposed in the
Another use case is to exploit the MEC for IoT (Internet literature that enable to bring computation close to the UEs.
of Thing) purposes [43]–[45]. Basically, IoT devices are con- Secondly, we describe the effort done within ETSI standard-
nected through various radio technologies (e.g., 3G, LTE, ization organization regarding the MEC. Finally, we compare
WiFi, etc.) using diverse communication protocols. Hence, individual existing MEC concepts (proposed in both literature
there is a need for low latency aggregation point to handle and ETSI) from several perspectives, such as MEC control or
various protocols, distribution of messages and for processing. location of the computation/storage resources.
This can be enabled by the MEC acting as an IoT gate-
way, which purpose is to aggregate and deliver IoT services A. Overview of the MEC Concept
into highly distributed mobile base stations in order to enable
In recent years, several MEC concepts with purpose to
applications responding in real time.
smoothly integrate cloud capabilities into the mobile network
The MEC can be also exploited for ITS to extend the
architecture have been proposed in the literature. This sec-
connected car cloud into the mobile network. Hence, road-
tion briefly introduces fundamental principles of small cell
side applications running directly at the MEC can receive
cloud (SCC), mobile micro cloud (MMC), fast moving
local messages directly from applications in the vehicles and
personal cloud, follow me cloud (FMC), and CONCERT.
roadside sensors, analyze them and broadcast warnings (e.g.,
Moreover, the section shows enhancements/modifications to
an accident) to nearby vehicles with very low latency. The
the network architecture necessary for implementation of each
exploitation of the MEC for car-to-car and car-to-infrastructure
MEC concept.
communications was demonstrated by Nokia and its partners
1) Small Cell Cloud (SCC): The basic idea of the
in an operator’s LTE network just recently in 2016 [46], [47].
SCC, firstly introduced in 2012 by the European project
TROPIC [48], [53], is to enhance small cells (SCeNBs), like
C. Network Performance and QoE Improvement Services microcells, picocells or femtocells, by an additional com-
The third category of use cases are those optimizing network putation and storage capabilities. The similar idea is later
performance and/or improving QoE. One such use case is to on addressed in SESAME project as well, where the cloud-
enable coordination between radio and backhaul networks. enabled SCeNBs supports the edge computing [49], [50]. The
So far, if the capacity of either backhaul or radio link is cloud-enhanced SCeNBs can pool their computation power
degraded, the overall network performance is negatively influ- exploiting network function virtualization (NFV) [51], [52]
enced as well, since the other part of the network (either radio paradigm. Because a high number of the SCeNBs is supposed
or backhaul, respectively) is not aware of the degradation. to be deployed in future mobile networks, the SCC can pro-
In this respect, an analytic application exploiting the MEC vide enough computation power for the UEs, especially for
can provide real-time information on traffic requirements of services/applications having stringent requirements on latency
the radio/backhaul network. Then, an optimization applica- (the examples of such applications are listed in Section II-A).
tion, running on the MEC, reshapes the traffic per application In order to fully and smoothly integrate the SCC concept
or re-routes traffic as required. into the mobile network architecture, a new entity, denoted
Another way to improve performance of the network is to as a small cell manager (SCM), is introduced to control the
alleviate congested backhaul links by local content caching at SCC [53]. The SCM is in charge of the management of the

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1632 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

Fig. 2. SCC architecture (MME - Mobility Management Entity, HSS - Home Subscriber Server, S-GW - Serving Gateway, P-GW - Packet Gateway).

Fig. 3. MMC architecture. Fig. 4. MobiScud architecture [57].

services with a low latency. While in the SCC the computa-

computing and/or storage resources provided by the SCeNBs. tion/storage resources are provided by interworking cluster(s)
Since the SCeNBs can be switched on/off at any time (espe- of the SCeNBs, the UEs exploit the computation resources
cially if owned by the users as in case of the femtocells), the of a single MMC, which is typically connected directly to a
SCM performs dynamic and elastic management of the com- wireless base station (i.e., the eNB in the mobile network) as
putation resources within the SCC. The SCM is aware of the indicated in Fig. 3. The MMC concept does not introduce any
overall cluster context (both radio and cloud-wise) and decides control entity into the network and the control is assumed to be
where to deploy a new computation or when to migrate an fully distributed in a similar way as the VL-SCM solution for
on-going computation to optimize the service delivery for the the SCC. To this end, the MMCs are interconnected directly
end-user. The computing resources are virtualized by means of or through backhaul in order to guarantee service continuity if
Virtual Machine (VM) located at the SCeNBs. An important the UEs move within the network to enable smooth VM migra-
aspect regarding the architecture of the SCC is deployment of tion among the MMCs (see more detail on VM migration in
the SCM (see Fig. 2). The SCM may be deployed in a cen- Section VII-B).
tralized manner either as a standalone SCM located within the 3) Fast Moving Personal Cloud (MobiScud): The
RAN, close to a cluster of the SCeNBs, or as an extension to a MobiScud architecture [57] integrates the cloud services into
MME [53], [54]. Moreover, the SCM can be deployed also in a the mobile networks by means of software defined network
distributed hierarchical manner, where a local SCM (L-SCM) (SDN) [58] and NFV technologies whilst maintaining back-
or a virtual L-SCM (VL-SCM) manages the computing and ward compatibility with existing mobile network. When
storage resources of the SCeNBs’ clusters in vicinity while compared to the SCC and the MMC concepts, the cloud
a remote SCM (R-SCM), located in the CN, has resources resources in the MobiScud are not located directly at the
of all SCeNBs connected to the CN at its disposal [55] access nodes such as SCeNB or eNB, but at operator’s clouds
(see Fig. 2b). located within RAN or close to RAN (see Fig. 4). Still, these
2) Mobile Micro Clouds (MMC): The concept of the MMC clouds are assumed to be highly distributed similarly as in
has been firstly introduced in [56]. Like the SCC, also the case of the SCC and the MMC enabling the cloud service to
MMC allows users to have instantaneous access to the cloud all UEs in vicinity.

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1633

Fig. 5. The network architecture enabling FMC concept (centralized

solution).
Fig. 6. CONCERT architecture.

Analogously to the SCC, the MobiScud introduces a new

control entity, a MobiScud control (MC), which interfaces SCC as explained in Section III-A1). Note that the FMC itself
with the mobile network, SDN switches and the cloud of the may be also decentrally controlled by omitting the FMCC alto-
operator. Basically, the MC has two functionalities: 1) monitor- gether. In such a case, the DCs coordinate themselves in a
ing control plane signaling message exchange between mobile self-organizing manner.
network elements to be aware of the UEs activity (e.g., han- 5) CONCERT: A concept converging cloud and cellular
dover) and 2) orchestrating and routing data traffic within SDN systems, abbreviated as CONCERT, has been proposed in [62].
enabled transport network to facilitate the application offload- The CONCERT assumes to exploit NFV principles and SDN
ing and the VM migration if the UE moves throughout the technology like above-mentioned solutions. Hence, the com-
network. puting/storage resources, utilized by both conventional mobile
communication and cloud computing services, are presented
4) Follow Me Cloud (FMC): The key idea of the FMC
as virtual resources. The control plain is basically consisted
is that the cloud services running at distributed data cen-
of a conductor, which is a control entity managing commu-
ters (DCs) follow the UEs as they roam throughout the
nication, computing, and storage resources of the CONCERT
network [59], [60] in the same way as in the case of
architecture. The conductor may be deployed centrally or in
the MobiScud. When compared to the previous MEC con-
a hierarchical manner for better scalability as in the SCC or
cepts, the computing/storage power is moved farther from
FMC. The data plain consists of radio interface equipments
the UEs; into the CN network of the operator. Nevertheless,
(RIEs) physically representing the eNB, SDN switches, and
while previous MEC concepts assume rather centralized CN
computing resources (see Fig. 6). The computing resources are
deployment, the FMC leverages from the fact that the mobile
used both for baseband processing (similarly as in C-RAN)
operators need to decentralize their networks to cope with
and for handling an application level processing (e.g., for
growing number of the UEs. In this respect, the centralized
the application offloading). In all already described MEC
CN used in the current network deployment is assumed to be
concepts, the computation/storage resources have been fully
replaced by a distributed one as shown in Fig. 5. For a con-
distributed. The CONCERT proposes rather hierarchically
venience of the mobile operators, the DC may be located at
placement of the resources within the network to flexibly
the same place as the distributed S/P-GWs.
and elastically manage the network and cloud services. In
Similarly as the SCC and the MobiScud, the FMC intro-
this respect, local servers with a low computation power are
duces new entities into the network architecture; a DC/GW
assumed to be located directly at the physical base station
mapping entity and an FMC controller (FMCC). These can
(e.g., similarly as in the SCC or the MMC) and, if the local
be either functional entities collocated with existing network
resources are not sufficient, regional or even central servers
nodes or a software run on any DC (i.e., exploiting NFV
are exploited as indicated in Fig. 6.
principles like the SCC or MobiScud concepts). The DC/GW
mapping entity maps the DCs to the distributed S/P-GWs
according to various metrics, such as, location or hop count B. ETSI MEC
between DC and distributed CN, in static or dynamic man- Besides all above-mentioned solutions, also ETSI is cur-
ner. The FMCC manages DCs’ computation/storage resources, rently deeply involved in standardization activities in order
cloud services running on them, and decides which DC should to integrate the MEC into the mobile networks. In this
be associated to the UE using the cloud services. The FMCC regard, we briefly summarize the standardization efforts on
may be deployed either centrally (as shown in Fig. 5) or hierar- the MEC within ETSI, describe reference architecture accord-
chically [61] with global FMCC (G-FMCC) and local FMCC ing to ETSI, and contemplate various options for the MEC
(L-FMCC) for better scalability (controlled similarly as in the deployment that are considered so far.

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1634 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

management. The MEC system level management includes a

user application lifecycle management (LCM) proxy, which
mediate the requests, such as initiation, termination or relo-
cations of the UE’s application within the MEC system to
the operation support system (OSS) of the mobile operator.
Then, the OSS decides if requests are granted or not. The
granted requests are forwarded to a mobile edge orchestra-
tor. The mobile edge orchestrator is the core functionality
in the MEC system level management as it maintains over-
all view on available computing/storage/network resources and
the MEC services. In this respect, the mobile edge orchestra-
tor allocates the virtualized MEC resources to the applications
that are about to be initiated depending on the applications
requirements (e.g., latency). Furthermore, the orchestrator also
flexibly scales down/up available resources to already running
Fig. 7. MEC reference architecture [66]. applications.
The MEC system level management is interconnected with
a MEC server level management constituting a mobile edge
platform and a virtualization platform manager. The former
1) Standardization of ETSI MEC: Standardization of the one manages the life cycle of the applications, application
MEC is still in its infancy, but drafts of specifications have rules and service authorization, traffic rules, etc. The latter one
already been released by ISG MEC. The terminology used in is responsible for allocation, management and release of the
individual specifications relating to conceptual, architectural virtualized computation/storage resources provided by the vir-
and functional elements is described in [63]. The main pur- tualization infrastructure located within the MEC server. The
pose of this document is to ensure the same terminology is MEC server is an integral part of the reference architecture
used by all ETSI specifications related to the MEC. A frame- as it represents the virtualized resources and hosts the MEC
work exploited by ISG MEC for coordination and promotion applications running as the VMs on top of the virtualization
of MEC is defined in proof of concept (PoC) specification [64]. infrastructure.
The basic objectives of this document is to describe the PoC 3) Deployment Options of ETSI MEC: As already men-
activity process in order to promote the MEC, illustrate key tioned in the previous subsection, the MEC services will be
aspects of the MEC and build a confidence in viability of provided by the MEC servers, which have the computation and
the MEC technology. Further, several service scenarios that storage resources at their disposal. There are several options
should benefit from the MEC and proximity of the cloud ser- where the MEC servers can be deployed within the mobile
vices is presented in [65] (see Section II for more detail). network. The first option is to deploy the MEC server directly
Moreover, technical requirements on the MEC to guarantee at the base station similarly as in case of the SCC or the
interoperability and to promote MEC deployment are intro- MCC (see Sections III-A1 and III-A2). Note that in case of a
duced in [38]. The technical requirements are divided into legacy network deployment, such as 3G networks, the MEC
generic requirements, service requirements, requirements on servers may be deployed at 3G Radio Network Controllers
operation and management, and finally security, regulations as well [38]. The second option is to place the MEC servers
and charging requirements. at cell aggregation sites or at multi-RAT aggregation points
2) ETSI MEC Reference Architecture: The reference archi- that can be located either within an enterprise scenario (e.g.,
tecture, described by ETSI in [66], is composed of functional company) or a public coverage scenario (e.g., shopping mall,
elements and reference points allowing interaction among stadium, airport, etc.). The third option is to move the MEC
them (see Fig. 7). Basically, the functional blocks may not nec- server farther from the UEs and locate it at the edge of CN
essarily represent physical nodes in the mobile network, but analogously to the FMC (Section III-A4).
rather software entities running on the top of a virtualization Of course, selection of the MEC server deployment depends
infrastructure. The virtualization infrastructure is understood on many factors, such as, scalability, physical deployment con-
as a physical data center on which the VMs are run and the straints and/or performance criteria (e.g., delay). For example,
VMs represent individual functional elements. In this respect, the first option with fully distributed MEC servers deployment
it is assumed that some architectural features from ETSI NFV will result in very low latencies since the UEs are in proxim-
group, which runs in parallel to ETSI MEC, will be reused for ity of the eNB and, hence, in proximity of the MEC server.
the MEC reference architecture as well, since the basic idea Contrary, the UEs exploiting the MEC server located in the CN
of NFV is to virtualize all network node functions. will inevitably experience longer latencies that could prevent
As shown in Fig. 7, the MEC can be exploited either by a a use of real-time applications. An initial study determining
UE application located directly in the UE, or by third party where to optimally install the MEC servers within the mobile
customers (such as commercial enterprise) via customer fac- network with the primary objective to find a trade-off between
ing service (CFS) portal. Both the UE and the CFS portal installation costs and QoS measured in terms of latency is
interact with the MEC system through a MEC system level presented in [67] and further elaborated in [68]. Based on

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1635

TABLE II
C OMPARISON OF E XISTING MEC C ONCEPTS

these studies, it is expected that, similarly as in CONCERT

framework (see Section III-A5), the MEC servers with var-
ious computation power/storage capacities will be scattered
throughout the network. Hence, the UEs requiring only a low
computation power will be served by the local MEC servers
collocated directly with the eNB, while highly demanding
applications will be relegated to more powerful MEC servers
farther from the UEs.

C. Summary
This section mutually compares the MEC concepts proposed
in literature with the vision of the MEC developed under
ETSI. There are two common trends followed by individ- Fig. 8. Possible outcomes of computation offloading decision.
ual MEC solutions that bring cloud to the edge of mobile
network. The first trend is based on virtualization techniques
exploiting NFVs principles. The network virtualization is a are also many options where to place MEC servers offering
necessity in order to flexibly manage virtualized resources computation/storage resources to the UEs. The most proba-
provided by the MEC. The second trend is a decoupling ble course of action is that the MEC servers will be deployed
the control and data planes by taking advantage of SDN everywhere in the network to guarantee high scalability of the
paradigm, which allows a dynamic adaptation of the network computation/storage resources. The comparison of all existing
to changing traffic patterns and users requirements. The use MEC concepts is shown in Table II.
of SDN for the MEC is also in line with current trends
in mobile networks [69]–[71]. Regarding control/signaling,
the MMC and MobiScud assume fully decentralize approach IV. I NTRODUCTION TO C OMPUTATION O FFLOADING
while the SCC, FMC, and CONCERT adopt either fully cen- From the user perspective, a critical use case regarding the
tralized control or hierarchical control for better scalability and MEC is a computation offloading as this can save energy
flexibility. and/or speed up the process of computation. In general, a cru-
If we compare individual MEC concepts in terms of com- cial part regarding computation offloading is to decide whether
putation/storage resources deployment, the obvious effort is to offload or not. In the former case, also a question is how
to fully distribute these resources within the network. Still, much and what should be offloaded [41]. Basically, a decision
each MEC concept differs in the location, where the compu- on computation offloading may result in:
tation/storage resources are physically located. While the SCC, • Local execution - The whole computation is done locally
MMC and MobiScud assume to place the computation close at the UE (see Fig. 8). The offloading to the MEC is
to the UEs within RAN, the FMC solution considers integra- not performed, for example, due to unavailability of the
tion of the DCs farther away, for example, in a distributed CN. MEC computation resources or if the offloading simply
On top of that, CONCERT distributes the computation/storage does not pay off.
resources throughout the network in a hierarchical manner so • Full offloading - The whole computation is offloaded and
that a low demanding computation application are handled processed by the MEC.
locally and high demanding applications are relegated either • Partial offloading - A part of the computation is processed
to regional or central servers. Concerning ETSI MEC, there locally while the rest is offloaded to the MEC.

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1636 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

Fig. 9. An example of partial offloading for application without non-

offloadable part(s) (a) and application with non-offloadable part (b).
Fig. 10. Dependency of offloadable components [72].

The computation offloading, and partial offloading in par-

ticular, is a very complex process affected by different factors, offloading could be quite tricky for continuous-execution
such as users preferences, radio and backhaul connection qual- application.
ity, UE capabilities, or cloud capabilities and availability [3]. • Dependency of the offloadable parts - The last criterion
An important aspect in the computation offloading is also an for classification of application to be offloaded is a mutual
application model/type since it determines whether full or par- dependency of individual parts to be processed. The parts
tial offloading is applicable, what could be offloaded, and how. of the application can be either independent on each other
In this regard, we can classify the applications according to or mutually dependent. In the former case, all parts can
several criteria: be offloaded simultaneously and processed in parallel.
• Offloadability of application - The application enabling In the latter case, however, the application is composed
code or data partitioning and parallelization (i.e., appli- of parts (components) that need input from some oth-
cation that may be partially offloaded) can be categorized ers and parallel offloading may not be applicable. Note
into two types. The first type of the applications is the that the relationship among individual components can be
app, which can be divided into N offloadable parts that expressed by component dependency graph (CDG) or call
all can be offloaded (see Fig. 9a). Since each offloadable graph (CG) (see [34], [41], [72], [73]). The relationship
part may differ in the amount of data and required com- among the components is illustrated in Fig. 10, where the
putation, it is necessary to decide which parts should be whole application is divided into M non-offloadable parts
offloaded to the MEC. In the example given in Fig. 9a, (1st , 4th , and 6th part in Fig. 10) and N offloadable parts
1st , 2nd , 3nd , 6th , and 9th parts are processed locally while (2nd , 3rd , and 5th part in Fig. 10). In the given example,
the rest is offloaded to the MEC. Notice that in the 2nd and 3rd part can be offloaded only after execution
extreme case, this type of application may be fully of the 1st part while the 5th part can be offloaded after
offloaded to the MEC if no parts are processed by the UE. execution of the 1st − 4th parts.
The second type of the applications is always composed The other important aspect regarding computation offload-
of some non-offloadable part(s) that cannot be offloaded ing is how to utilize and manage offloading process in practice.
(e.g., user input, camera, or acquire position that needs Basically, the UE needs to be composed of a code profiler,
to be executed at the UE [72]) and M offloadable parts. system profiler, and decision engine [36]. The code profiler’s
In Fig. 9b, the UE processes the whole non-offloadable responsibility is to determine what could be offloaded (depend-
part together with 2nd , 6th , and 7th parts while the rest of ing on application type and code/data partitioned as explained
the application is offloaded to the MEC. above). Then, the system profiler is in charge of monitoring
• Knowledge on the amount of data to be processed - The various parameters, such as available bandwidth, data size to
applications can be classified according to the knowl- be offloaded or energy spent by execution of the applications
edge on the amount of data to be processed. For the locally. Finally, decision engine determines whether to offload
first type of the applications (represented, e.g., by face or not.
detection, virus scan, etc.,) the amount of data to be The next sections survey current research works focusing
processed is known beforehand. For the second type on following pivotal research topics: 1) decision on the com-
of the applications, it is not possible to estimate the putation offloading to the MEC, 2) efficient allocation of the
amount of data to be processed as these are continuous- computation resources within the MEC, and 3) mobility man-
execution application and there is no way to predict how agement for the moving users exploiting MEC services. Note
long they will be running (such as, online interactive that from now on we use explicitly the terminology according
games) [95]. It is obvious that decision on computation to ETSI standardization activities. Consequently, we use term

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1637

Fig. 11. The example of offloading decision aiming minimization of

execution delay.
Fig. 12. Computation offloading considered in [74] (CSI stands for channel
state information).
MEC server as a node providing computing/storage resources
to the UEs instead of DC, MMC, etc.
The performance of the proposed algorithm is compared to
V. D ECISION ON C OMPUTATION O FFLOADING TO MEC the local execution policy (computation done always locally),
cloud execution policy (computation performed always by the
This section surveys current research related to the decision MEC server), and greedy offloading policy (UE schedules data
on the computation offloading to the MEC. The papers are waiting in the buffer whenever the local CPU or the trans-
divided into those considering either only the full offloading mission unit is idle). The simulation results show that the
(Section V-A) or those taking into account also possibility of proposed optimal policy is able to reduce execution delay by
the partial offloading (Section V-B). up to 80% (compared to local execution policy) and roughly
up to 44% (compared to cloud execution policy) as it is able to
A. Full Offloading cope with high density of applications’ arrival. The drawback
The main objective of the works focused on the full offload- of the proposed method is that the UE requires feedback from
ing decision is to minimize an execution delay (Section V-A1), the MEC server in order to make the offloading decision, but
to minimize energy consumption at the UE while predefined the generated signaling overhead is not discussed in the paper.
delay constraint is satisfied (Section V-A2), or to find a Another idea aiming at minimization of the execution delay
proper trade-off between both the energy consumption and is introduced in [75]. When compared to the previous study,
the execution delay (Section V-A3). Mao et al. [75] also reduce application failure for the offloaded
1) Minimization of Execution Delay: One of the advan- applications. The paper considers the UE applies dynamic
tages introduced by the computation offloading to the MEC voltage and frequency scaling (DVS) [76] and energy har-
is a possibility to reduce the execution delay (D). In case vesting techniques [77] to minimize the energy consumption
the UE performs all computation by itself (i.e., no offload- during the local execution and a power control optimizing data
ing is performed), the execution delay (Dl ) represents solely transmission for the computation offloading. In this respect,
the time spent by the local execution at the UE. In case the authors propose a low-complexity Lyapunov optimization-
of the computation offloading to the MEC, the execution based dynamic computation offloading (LODCO) algorithm.
delay (Do ) incorporates three following parts: 1) transmission The LODCO makes offloading decision in each time slot
duration of the offloaded data to the MEC (Dot ), 2) compu- and subsequently allocates CPU cycles for the UE (if the
tation/processing time at the MEC (Dop ), and 3) time spent local execution is performed) or allocates transmission power
by reception of the processed data from the MEC (Dor ). The (if the computation offloading is performed). The proposed
simple example of the computation offloading decision based LODCO is able to reduce execution time by up to 64% by
solely on the execution delay is shown in Fig. 11. It could be offloading to the MEC. Furthermore, the proposal is able
observed that the UE1 performs all computation locally since to completely prevent a situation when offloaded application
the local execution delay is significantly lower than expected would be dropped.
execution delay for the computation offloading to the MEC The drawback of both above-mentioned papers is that the
(i.e., Dl < Do ). Contrary, a better alternative for the UE2 is offloading decision does not take into account energy con-
to fully offload data to the MEC as the local execution would sumption at the side of UE as fast battery depletion impose
result in notable higher execution delay (i.e., Dl > Do ). significant obstacle in contemporary networks. In [75], the
The goal to minimize execution delay is pursued by energy aspect of the UE is omitted in the decision process
Liu et al. [74]. This is accomplished by one-dimensional since the paper assumes that the UEs exploit energy harvest-
search algorithm, which finds an optimal offloading decision ing techniques. The harvesting technique, however, is not able
policy according to the application buffer queuing state, avail- to fully address energy consumption problem by itself.
able processing powers at the UE and at the MEC server, and 2) Minimization of Energy Consumption While Satisfying
characteristic of the channel between the UE and the MEC Execution Delay Constraint: The main objective of the papers
server. The computation offloading decision itself is done at surveyed in this section is to minimize the energy consumption
the UE by means of a computation offloading policy mod- at the UE while the execution delay constraint of the appli-
ule (see Fig. 12). This module decides, during each time slot, cation is satisfied. On one hand, the computation offloaded to
whether the application waiting in a buffer should be processed the MEC saves battery power of the UE since the computa-
locally or at the MEC while minimizing the execution delay. tion does not have to be done locally. On the other hand, the

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1638 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

Fig. 13. The example of computation offloading decision based on energy consumption while satisfying execution delay constraint.

UE spends certain amount of energy in order to: 1) trans- constraints. The UEs that are not allowed to offload the com-
mit offloaded data for computation to the MEC (Eot ) and putation make either local computation or stay idle. It is shown
2) receive results of the computation from the MEC (Eor ). The the offline strategy notably outperforms the online strategies
simple example of the computation offloading decision primar- in terms of the energy saving (by roughly 50%). In addition,
ily based on the energy consumption is shown in Fig. 13. In the energy consumed by individual UEs strongly depends on
the given example, the UE1 decides to perform the compu- requirements of other UEs application.
tation locally since the energy spent by the local execution Another offloading decision strategy for the multi-UEs case
(El ) is significantly lower than the energy required for trans- minimizing the energy consumption at the UEs while satisfy-
mission/reception of the offloaded data (E0 ). Contrary, the ing the maximum allowed execution delay is proposed in [81].
UE2 offloads data to the MEC as the energy required by the A decision on the computation offloading is done periodically
computation offloading is significantly lower than the energy in each time slot, during which all the UEs are divided into two
spent by the local computation. Although the overall execution groups. While the UEs in the first group are allowed to offload
delay would be lower if the UE1 offloads computation to the computation to the MEC, the UEs in the second group have to
MEC and also if the UE2 performs the local execution, the perform computation locally due to unavailable computation
delay is still below maximum allowed execution delay con- resources at the MEC (note that in the paper, the computa-
straint (i.e., Dl < Dmax ). Note that if only the execution delay tion is done at the serving SCeNB). The UEs are sorted to
would be considered for the offloading decision (as considered the groups according to the length of queue, that is, according
in Section V-A3), both UEs would unnecessarily spent more to the amount of data they need to process. After the UEs
energy. are admitted to offload the computation, joint allocation of
The computation offloading decision minimizing the energy the communication and computation resources is performed
consumption at the UE while satisfying the execution delay by finding optimal transmission power of the UEs and allo-
of the application is proposed in [78]. The optimization cation of the SCeNB’s computing resources to all individual
problem is formulated as a constrained Markov decision pro- UEs. The performance of the proposal is evaluated in terms of
cess (CMDP). To solve the optimization problem, two resource an average queue length depending on intensity of data arrival
allocation strategies are introduced. The first strategy is based and a number of antennas used at the UEs and the SCeNB.
on an online learning, where the network adapts dynamically It is shown that the more antennas is used, the less transmis-
with respect to the application running at the UE. The second sion power at the UEs is needed while still ensuring the delay
strategy is pre-calculated offline strategy, which takes advan- constraint of the offloaded computation.
tage of a certain level of knowledge regarding the application The main weak point of [81] is that it assumes only a single
(such as arrival rates measured in packets per slot, radio chan- SCeNB and, consequently, there is no interference among the
nel condition, etc.). The numerical experiments show that the UEs connected to various SCeNBs. Hence, the work in [81]
pre-calculated offline strategy is able to outperform the online is extended in [82] to the multi-cell scenario with N SCeNBs
strategy by up to 50% for low and medium arrival rates (loads). to reflect the real network deployment. Since the formulated
Since the offline resource allocation strategy proposed in [78] optimization problem in [81] is no longer convex, the authors
shows its merit, the authors devise two addition dynamic propose a distributed iterative algorithm exploiting Successive
offline strategies for the offloading [79]: deterministic offline Convex Approximation (SCA) converging to a local optimal
strategy and randomized offline strategy. It is demonstrated solution. The numerical results demonstrate that the proposed
that both offloading offline strategies can lead to significant joint optimization of radio and computational resources signifi-
energy savings comparing to the case when the computing is cantly outperforms methods optimizing radio and computation
done solely at the UE (energy savings up to 78%) or solely at separately. Moreover, it is shown that the applications with
the MEC (up to 15%). fewer amount of data to be offloaded and, at the same time,
A further extension of [79] from a single-UE to a multi- requiring high number of CPU cycles for processing are more
UEs scenario is considered in [80]. The main objective is to suitable for the computation offloading. The reason is that the
jointly optimize scheduling and computation offloading strat- energy spent by the transmission/reception of the offloaded
egy for each UE in order to guarantee QoE, fairness between data to the MEC is significantly lower than the energy sav-
the UEs, low energy consumption, and average queuing/delay ings at the UE due to the computation offloading. The work

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1639

in [82] is further extended in [83] by a consideration of multi- respect to [85] is that Chen et al. [86] assume the compu-
clouds that are associated to individual SCeNBs. The results tation can be offloaded also to the remote centralized cloud
show that with an increasing number of the SCeNBs (i.e., with (CC), if computation resources of the MEC are not sufficient.
increasing number of clouds), the energy consumption of the The computation offloading decision is done in a sequential
UE proportionally decreases. manner. In the first step, the UE decides whether to offload the
The same goal as in previous paper is achieved in [84] by application(s) to the MEC or not. If the application is offloaded
means of an energy-efficient computation offloading (EECO) to the MEC, the MEC evaluates, in the second step, if it is able
algorithm. The EECO is divided into three stages. In the first to satisfy the request or if the computation should be farther
stage, the UEs are classified according to their time and energy relayed to the CC. The problem is formulated as a non-convex
cost features of the computation to: 1) the UEs that should quadratically constrained quadratic program (QCQP), which
offload the computation to the MEC as the UEs cannot sat- is, however, NP-hard. Hence, a heuristic algorithm based on
isfy the execution latency constraint, 2) the UEs that should a semi-definite relaxation together with a novel randomiza-
compute locally as they are able to process it by itself while tion method is proposed. The proposed heuristic algorithm is
the energy consumption is below a predefined threshold, and able to significantly lower a total system cost (i.e., weighted
3) the UEs that may offload the computation or not. In the sum of total energy consumption, execution delay and costs
second stage, the offloading priority is given to the UEs from to offload and process all applications) when compared to the
the first and the third set determined by their communica- situation if the computation is done always solely at the UE
tion channels and the computation requirements. In the third (roughly up to 70%) or always at the MEC/CC (approximately
stage, the eNBs/SCeNBs allocates radio resources to the UEs up to 58%).
with respect to given priorities. The computational complex- The extension of [86] from the single-UE to the multi-
ity of the EECO is O(max(I2 + N, IK + N)), where I is the UEs scenario is presented in [87]. Since the multiple UEs
number of iterations, N stands for amount of UEs, and K repre- are assumed to be connected to the same computing node
sents the number of available channels. According to presented (e.g., eNB), the offloading decision is done jointly with the
numerical results, the EECO is able to decrease the energy allocation of computing and communication resources to all
consumption by up to 15% when compared to the computation UEs. Analogously to [86], the proposal in [87] outperforms
without offloading. Further, it is proofed that with increasing the case when computation is done always by the UE (system
computational capabilities of the MEC, the number of UEs cost decreased by up to 45%) and strategy if computation is
deciding to offload the computation increases as well. always offloaded to the MEC/CC (system cost decreased by
3) Trade-Off Between Energy Consumption and Execution up to 50%). Still, it would be useful to show the results for
Delay: The computation offloading decision for the multi-user more realistic scenario with multiple computing eNBs, where
multi-channel environment considering a trade-off between interference among the UEs attached to different eNBs would
the energy consumption at the UE and the execution delay play an important role in the offloading decision. Moreover,
is proposed in [85]. Whether the offloading decision prefers the overall complexity of the proposed solution is O(N 6 ) per
to minimize energy consumption or execution delay is deter- one iteration, which could be too high for a high number of
mined by a weighing parameter. The main objective of the UEs (N) connected to the eNB.
paper is twofold; 1) choose if the UEs should perform the
offloading to the MEC or not depending on the weighing
parameter and 2) in case of the computation offloading, select B. Partial Offloading
the most appropriate wireless channel to be used for data This subsection focuses on the works dealing with the par-
transmission. To this end, the authors present an optimal cen- tial offloading. We classify the research on works focused
tralized solution that is, however, NP-hard in the multi-user on minimization of the energy consumption at the UE while
multi-channel environment. Consequently, the authors also predefined delay constraint is satisfied (Section V-B1) and
propose a distributed computation offloading algorithm achiev- works finding a proper trade-off between both the energy
ing Nash equilibrium. Both the optimal centralized solution consumption and the execution delay (Section V-B2).
and the distributed algorithm are compared in terms of two 1) Minimization of Energy Consumption While Satisfying
performance metrics; 1) the amount of the UEs for which the Execution Delay Constraint: This section focuses on the
computation offloading to the MEC is beneficial and 2) the works aiming on minimization of the energy consumption
computation overhead expressed by a weighing of the energy while satisfying maximum allowable delay, similarly as in
consumption and the execution delay. The distributed algo- Section V-A2. Cao et al. [88] consider the application divided
rithm performs only slightly worse than the centralized one into a non-offloadable part and N offloadable parts as shown
in both above-mentioned performance metrics. In addition, in Fig. 9b. The main objective of the paper is to decide, which
the distributed algorithm significantly outperforms the cases offloadable parts should be offloaded to the MEC. The authors
when all UEs compute all applications locally and when all propose an optimal adaptive algorithm based on a combina-
UEs prefer computing at the MEC (roughly by up to 40% torial optimization method with complexity up to O(2N ). To
for 50 UEs). decrease the complexity of the optimal algorithm, also a sub-
Other algorithm for the computation offloading decision optimal algorithm is proposed reducing complexity to O(N).
weighing the energy consumption at the UE and the exe- The optimal algorithm is able to achieve up to 48% energy sav-
cution delay is proposed in [86]. The main difference with ings while the sub-optimal one performs only slightly worse

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1640 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

(up to 47% energy savings). Moreover, it is shown that increas- computation offloading to the MEC. Contrary, if the UE has
ing SINR between the UE and the serving eNBs leads to more a lower priority than the threshold, it offloads only minimum
prominent energy savings. amount of computation to satisfy the application latency con-
The minimization of the energy consumption while satis- straints. Since the optimal joint allocation of communication
fying the delay constrains of the whole application is also and computation resources is of a high complexity, the authors
the main objective of [72]. Contrary to [88] the application also propose a sub-optimal allocation algorithm, which decou-
in [72] is supposed to be composed of several atomic parts ples communication and computation resource allocation. The
dependable on each other, i.e., some parts may be processed simulation results indicate this simplification leads to negligi-
only after execution of other parts as shown in Fig. 10 in bly higher total energy consumption of the UE when compared
Section IV. The authors formulate the offloading problem as to the optimal allocation. The paper is further extended in [92],
0 − 1 programming model, where 0 stands for the applica- where You et al. show that OFDMA access enables roughly
tion offloading and 1 represents the local computation at the ten times higher energy savings achieved by the UEs com-
UE. Nevertheless, the optimal solution is of a high complex- paring to TDMA system due to higher granularity of radio
ity as there exists 2N possible solutions to this problem (i.e., resources.
O(2N N 2 )). Hence, the heuristic algorithm exploiting Binary In all above-mentioned papers on partial offloading, the
Particle Swarm Optimizer (BPSO) [89] is proposed to reduce minimization of UE’s energy consumption depends on the
the complexity to O(G.K.N 2 ), where G is the number of iter- quality of radio communication channel and transmission
ations, and K is the number of particles. The BPSO algorithm power of the UE. Contrary, in [93], the minimization of
is able to achieve practically the same results as the high com- energy consumption while satisfying execution delay of the
plex optimal solution in terms of the energy consumption. application is accomplished through DVS technique. In this
Moreover, the partial offloading results in more significant respect, the authors propose an energy-optimal partial offload-
energy savings with respect to the full offloading (up to 25% ing scheme that forces the UE adapt its computing power
energy savings at the UE). depending on maximal allowed latency of the application
A drawback of both above papers focusing in detail on the (LMAX ). In other words, the objective of the proposed scheme
partial computation offloading is the assumption of only single is to guarantee that the actual latency of the application is
UE in the system. Hence, Zhao et al. [90] address the partial always equal to LMAX . As a consequence, the energy con-
offloading decision problem for the multi-UEs scenario. With sumption is minimized while perceived QoS by the users is
respect to [72] and [88], the application to be offloaded does not negatively affected.
not contain any non-offloadable parts and, in some extreme 2) Trade-Off Between Energy Consumption and Execution
cases, the whole application may be offloaded if profitable Delay: A trade-off analysis between the energy consumption
(i.e., the application is structured as illustrated in Fig. 9a. The and the execution delay for the partial offloading decision is
UEs are assumed to be able to determine whether to parti- delivered in [94]. Similarly as in [90], the application to be
tion the application and how many parts should be offloaded offloaded contains only offloadable parts and in extreme case,
to the MEC. The problem is formulated as a nonlinear con- the full offloading may occur (as explained in Section V-B).
straint problem of a high complexity. As a consequence, it is The offloading decision considers the following parameters:
simplified to the problem solvable by linear programming and 1) total number of bits to be processed, 2) computational capa-
resulting in the complexity O(N) (N is the number of UEs bilities of the UE and the MEC, 3) channel state between
performing the offloading). If the optimal solution applying the UE and the serving SCeNB that provides access to the
exhaustive search is used, 40% energy savings are achieved MEC, and 4) energy consumption of the UE. The computa-
when compared to the scenario with no offloading. In case of tion offloading decision is formulated as a joint optimization
the heuristic low complex algorithm, 30% savings are observed of communication and computation resources allocation. The
for the UEs. The disadvantage of the proposal is that it assumes simulation results indicate that the energy consumption at
the UEs in the system have the same channel quality and the UE decreases with increasing total execution time. This
all of them are of the same computing capabilities. These decrease, however, is notable only for small execution time
assumptions, however, are not realistic for the real network. duration. For a larger execution time, the gain in the energy
A multi-UEs scenario is also assumed in [91], where savings is inconsequential. Moreover, the authors show the
You and Huang assume TDMA based system where time is offloading is not profitable if the communication channel is
divided into slots with duration of T seconds. During each slot, of a low quality since a high amount of energy is spent to
the UEs may offload a part of their data to the MEC according offload the application. In such situation, the whole appli-
to their channel quality, local computing energy consump- cation is preferred to be processed locally at the UE. With
tion, and fairness among the UEs. In this regard, an optimal an intermediate channel quality, a part of the computation
resource allocation policy is defined giving higher priority to is offloaded to the MEC as this results in energy savings.
those UEs that are not able to meet the application latency con- Finally, if the channel is of a high quality, the full offloading
straints if the computation would be done locally. After that, is preferred since the energy consumption for data transmis-
the optimal resource allocation policy with threshold based sion is low while the savings accomplished by the computation
structure is proposed. In other words, the optimal policy makes offloading are high.
a binary offloading decision for each UE. If the UE has a The study in [95] provides more in-depth theoretical anal-
priority higher than a given threshold, the UE performs full ysis on trade-off between the energy consumption and the

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1641

TABLE III
C OMPARISON OF I NDIVIDUAL PAPERS A DDRESSING C OMPUTATION O FFLOADING D ECISIONS

latency of the offloaded applications preliminarily handled The main drawback in [94] and [95] is that these papers con-
in [94]. Moreover, the authors further demonstrate that a sider only the single-UE scenario. A trade-of analysis between
probability of the computation offloading is higher for good the energy consumption at the UE and the execution delay
channel quality. With higher number of antennas (4x2 MIMO for the multi-UEs scenario is delivered in [96]. In case of
and 4x4 MIMO is assumed), the offloading is done more often the multi-UEs scenario, the whole joint optimization process
and the energy savings at the UE are more significant when proposed in [95] has to be further modified since both com-
compared to SISO or MISO (up to 97% reduction of energy munication and computation resources provided by the MEC
consumption for 4x4 MIMO antenna configuration). Note that are shared among multiple UEs. In the paper, it is proven that
the same conclusion is also reached, e.g., in [81] and [82]. with more UEs in the system, it takes more time to offload the

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1642 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

application and it also lasts longer to process the application

in the MEC. The reason for this phenomenon is quite obvious
since less radio and computational resources remains for each
UE. Still, up to 90% of energy savings may be accomplished
in multi-UE scenario.
A trade-off between the power consumption and the execu-
tion delay for the multi-UEs scenario is also tackled in [97].
The authors formulate a power consumption minimization
problem with application buffer stability constraints. In this
regard, the online algorithm based on Lyapunov optimiza-
tion is proposed to decide on optimal CPUs frequency for
those UEs performing the local execution and to allocate
transmission power and bandwidth to the UEs offloading the Fig. 14. An example of allocation of computing resources within the MEC.
application to the MEC. The proposed algorithm is able to con-
trol the power consumption and the execution delay depending
on the selected priority. The paper also demonstrates that the
use of the MEC for the computation offloading is able to bring
up to roughly 90% reduction in the power consumption while
the execution delay is reduced approximately by 98%.

C. Summary of Works Focusing on Computation

Offloading Decision
A comparison of individual computation offloading strate-
gies is illustrated in Table III. The majority of computation Fig. 15. Allocation of computation resources according to [98].
offloading decision algorithms aims to minimize the energy
consumption at the UE (EUE ) while satisfying the execu-
tion delay (D) acceptable by the offloaded application or to
focusing on allocation of the computation resources at 1) a sin-
find/analyse a trade-off between these two metrics. The papers
gle computing node (Section VI-A) and 2) multiple computing
indicate up to 90% energy savings achievable by the compu-
nodes (Section VI-B).
tation offloading to the MEC and execution delay may be
reduced even up to 98%. Besides, all the papers evaluate
the proposed solutions mostly by means of simulation (only
A. Allocation of Computation Resources at a Single Node
several studies perform analytical evaluations).
The maximization of the amount of the applications served
by the MEC while satisfying the delay requirements of the
offloaded applications is the main objective in [98]. The
VI. A LLOCATION OF C OMPUTING R ESOURCES decision where the individual applications should be placed
If a decision on the full or partial offloading of an applica- depends on the applications priorities (derived from the appli-
tion to the MEC is taken (as discussed in previous section), a cation’s delay requirements, i.e., the application with a low
proper allocation of the computation resources has to be done. delay requirements has higher priority) and availability of the
Similarly as in case of the computation offloading decision, the computing resources at the MEC. The basic principle for the
selection of computation placement is influenced by the abil- allocation of computation resources is depicted in Fig. 15.
ity of the offloaded application to be parallelized/partitioned. The offloaded applications are firstly delivered to the local
If the parallelization/partitioning of the application is not scheduler within the MEC. The scheduler checks if there is a
possible, only one physical node may be allocated for the computing node with sufficient computation resources. If there
computing since the application cannot be split into several is a computing node with enough available resources, the VM
parts (in Fig. 14, the UE1 offloads whole application to the is allocated at the node. Then the application is processed at
eNB as this application cannot be partitioned). In the opposite this MEC node, and finally sent back to the UE (see Fig. 15).
case, the offloaded application may be processed by resources However, if the computation power provided by the MEC
distributed over several computing nodes (in Fig. 14, the appli- server is not sufficient, the scheduler delegates the applica-
cation offloaded by the UE2 is partitioned and processed by tion to the distant CC. In order to maximize the amount of
all three eNBs). applications processed in the MEC while satisfying their delay
This section surveys the papers addressing the problem of requirements, the authors propose a priority based cooperation
a proper allocation of the computing resources for the appli- policy, which defines several buffer thresholds for each priority
cations that are going to be offloaded to the MEC (or in some level. Hence, if the buffer is full, the applications are sent to
cases to the CC, if the MEC computing resources are not the CC. The optimal size of the buffer thresholds is found by
sufficient). We categorize the research in this area into papers means of low-complexity recursive algorithm. The proposed

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1643

Fig. 16. An example of the VM allocation at single computing SCeNB

according to [100].
Fig. 17. An example of allocation of computation resources for individual
UEs according to [101].

cooperation policy is able to increase the probability of the

application completion within the tolerated delay by 25%. B. Allocation of Computation Resources at Multiple
When compared to the previous paper, the general objective Nodes (Federated Clouds)
of [99] is to minimize not only the execution delay but also When compared to the previous section, the allocation of
the power consumption at the MEC. The paper considers a computation resources at multiple computing nodes is con-
hot spot area densely populated by the UEs, which are able sidered here. The papers are split into subsections according
to access several MEC servers through nearby eNBs. To that to the main objective: 1) minimize execution delay and/or
end, an optimal policy is proposed using equivalent discrete power consumption of computing nodes (Section VI-B1)
MDP framework. However, this method results in a high com- and 2) balance both communication and computing loads
munication overhead and high computational complexity with (Section VI-B1).
increasing number of the MEC servers. Hence, this problem is 1) Minimization of Execution Delay and/or Power
overcome by developing an application assignment index pol- Consumption of Computing Nodes: The minimization of
icy. In this respect, each eNB calculates its own index policy the execution delay by allocation of computing resources
according to the state of its computing resources. Then, this provided by the cluster of SCeNBs while avoiding to use
index policy is broadcasted by all eNBs and the UE is able the CC is proposed in [101]. The cluster formation is done
to select the most suitable MEC server in order to minimize by means of a cooperative game approach, where monetary
both execution delay and power consumption. According to incentives are given to the SCeNBs if they perform the
the results, the index policy is in the worst case by 7% more computation for the UEs attached to other SCeNBs. The
costly than optimal policy in terms of system cost (note that coalition among the SCeNBs is formed for several time slots
system cost represents weighted execution delay and power and then new coalitions may be created. The allocation of
consumptions of the MEC). computation resources is done as shown in Fig. 17. Firstly, the
The minimization of the execution delay of the offloaded serving SCeNB tries to serve their UEs on its own since this
application is also the main goal in [100]. Nonetheless, with results in the shortest communication delay (e.g., in Fig. 17
respect to [98] and [99], the other objectives are to minimize SCeNB1 allocates the computation resources to the UE1 and
both communication and computing resource overloading and the UE2, etc.). Only if the SCeNB is not able to process the
the VM migration cost (note that in [100], the computing application on its own, it is forwarded to all SCeNBs in the
nodes are represented by the SCeNBs and the VM migration same cluster (in Fig. 17, the computation for the UE3 is done
may be initiated due to the SCeNBs shutdown). The whole at the SCeNB2 and the SCeNB3). The numerical results show
problem is formulated as the VM allocation at the SCeNB that the proposed scheme is able to reduce the execution
and solved by means of MDP. An example of the VM alloca- delay by up to 50% when compared to the computation only
tion according to [100] is shown in Fig. 16, where the VM for at the serving SCeNB and by up to 25% comparing to the
the UE1 is allocated at the serving SCeNB1 while the UE2 has scenario when all SCeNBs in the system participate in the
allocated the VM at the neighbouring SCeNB3. The SCeNB3 computation. Unfortunately, the proposed approach does not
is preferred because of a high quality backhaul resulting in a address a problem of forming new coalitions and its impact
low transmission delay of the offloaded data. The simulations on currently processed applications.
show that with higher VM migration cost, the VM is preferred The selection of computing nodes can significantly influence
to be allocated at the serving SCeNB (i.e., the SCeNB closest not only the execution delay, as considered in [101], but also
to the UE) if this SCeNB has enough computation power. the power consumption of the computing nodes. Hence, the
The main disadvantage of all above-mentioned approaches main objective of [102] is to analyze an impact of the cluster
is that these do not consider more computing nodes within size (i.e., the amount of the SCeNBs performing computing)
the MEC for single application in order to further decrease its on both execution latency of the offloaded application and the
execution delay. power consumption of the SCeNBs. The analysis is done for

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1644 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

different backhaul topologies (ring, tree, full mesh) and tech-

nologies (fiber, microwave, LTE). The authors demonstrate
that full mesh topology combined with fiber or microwave
connection is the most profitable in terms of execution latency
(up to 90% execution delay reduction). Contrary, a fiber back-
haul in ring topology results in the lowest power consumption.
Moreover, the paper shows that an increasing number of the
computing SCeNBs does not always shorten execution delay.
Quite the opposite, if a lot of SCeNBs process the offloading
applications and the transmission delay becomes longer than
the computing delay at the SCeNBs, the execution delay may
be increased instead. Besides, with an increasing number of
the computing SCeNBs, power consumption increases as well.
Consequently, a proper cluster formation and the SCeNBs Fig. 18. An example of application and physical graph according to [107] (FD
selection play a crucial part in system performance. - Face detection, IPFE - Image processing and feature extraction, FR - Face
recognition, D - Database).
The problem to find an optimal formation of the clusters
of SCeNBs for computation taking into account both execu-
tion delay and power consumption of the computing nodes is
addressed in [103]. The paper proposes three different cluster- Similar as in [104], the multi-UE cluster allocation is
ing strategies. The first clustering strategy selects the SCeNBs assumed in [105], but the cluster formation is done jointly
in order to minimize execution delay. Since all SCeNBs in the with the UEs scheduling. The proposed resource allocation
system model are assumed to be one hop away (i.e., full mesh process is split into two steps similarly as proposed in [101].
topology is considered), basically all SCeNBs are included in In the first step, labeled as local computational resource allo-
the computation resulting in up to 22% reduction of execu- cation, each SCeNB allocates its computational resources to
tion delay. This is due to the fact that the computation gain their own UEs according to specific scheduling rules, such
(and, thus, increase in the offloaded application processing) is as application latency constraint, computation load or min-
far greater than the transmission delay. The objective of the imum required computational capacity. In the second step,
second clustering strategy is to minimize overall power con- labelled as establishment of computing clusters, the computa-
sumption of the cluster. In this case, only the serving SCeNB is tion clusters are created for each UE that cannot be served by
preferred to compute, thus, any computation at the neighbour- its serving SCeNB. The authors propose three algorithm real-
ing SCeNBs is suppressed to minimize power consumption of izations differing in applications prioritization (e.g., earliest
the SCeNBs (up to 61% reduction of power consumption is deadline first or according to computation size of applica-
observed). This, however, increases overall latency and high tion) and the objective (minimization of power consumption
variations of the computation load. The last clustering strat- or execution latency similarly as, e.g., [104]). The simulations
egy aims to minimize the power consumption of each SCeNB illustrate that there could be found the algorithm realiza-
in the cluster, since the power consumptions of the individual tion resulting in the users satisfaction ratio above 95% while
SCeNBs is highly imbalanced in the second strategy. keeping a moderate power consumption of all computing
While in [103] the optimal clustering of the SCeNBs is nodes.
done only for single UE, the multi UEs scenario is assumed 2) Balancing of Communication and Computation Load:
in [104]. When compared to the previous paper, whenever the In the previous section, the allocation of computing resources
UE is about to offload data for the computation, the computing is done solely with purpose to minimize the execution delay
cluster is assigned to it. Consequently, each UE has assigned and/or the power consumption of the computing nodes.
different cluster size depending on the application and the This could, however, result in unequal load distribution
UE’s requirements. The core idea of the proposal is to jointly among individual computing nodes and backhaul overload-
compute clusters for all active users’ requests simultaneously ing. The balancing of communication and computation load
to being able efficiently distribute computation and communi- of the SCeNBs while satisfying the delay requirement of the
cation resources among the UEs and to achieve higher QoE. offloaded application is addressed in [106]. To this end, an
The main objective is to minimize the power consumption of Application Considering Algorithm (ACA) selecting suitable
the clusters while guaranteeing required execution delay for SCeNBs according to the current computation and com-
each UE. The joint clusters optimization is able to significantly munication load of the SCeNBs is proposed. The ACA
outperform the successive cluster optimization (allocation of exploits knowledge of the offloaded application’s require-
the clusters are done subsequently for each UE), the static ments (i.e., the number of bytes to be transferred and the
clustering (equal load distribution among SCeNBs) and no maximum latency acceptable by the application/user). The
clustering (computation is done only by the serving SCeNB) selection of the SCeNBs for the computation is done in a
in terms of the users’ satisfaction ratio (up to 95% of UEs is static way prior to the offloading to avoid expensive VMs
satisfied). On the other hand, the average power consumption migration. The performance evaluation is done for two back-
is significantly higher when compared to “no clustering” and hauls, low throughput ADSL and high quality gigabit pas-
“successive clusters optimization” scenarios. sive optical network (GPON). The proposed ACA algorithm

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1645

TABLE IV
C OMPARISON OF I NDIVIDUAL PAPERS A DDRESSING A LLOCATION OF C OMPUTATION R ESOURCES FOR
A PPLICATION /DATA A LREADY D ECIDED TO B E O FFLOADED

is able to satisfy 100% of the UEs as long as number two heuristic approaches in terms of resource utilization by
of offloaded tasks per second is up to 6. Moreover, the roughly 10%.
paper shows that tasks parallelization helps to better balance
computation load.
The main objective to balance the load (both communication C. Summary of Works Dealing With Allocation of
and computation) among physical computing nodes and, at the Computing Resources
same time, to minimize the resource utilization of each phys- The comparison of individual methods addressing alloca-
ical computing node (i.e., reducing sum resource utilization) tion of the computation resources within the MEC is shown
is also considered in [107]. The overall problem is formu- in Table IV. The main objective of the studies dealing with the
lated as a placement of application graph onto a physical allocation of computation resources is to minimize the execu-
graph. The former represents the application where nodes in tion delay of the offloaded application (D). In other words the
graph correspond to individual components of the applica- aim is to ensure QoS to the UEs in order to fully exploit prox-
tion and edges to the communication requirements between imity of the MEC with respect to the computing in faraway
them. The latter represents physical computing system, where CC. Moreover, several studies also focus on minimization of
the nodes in graph are individual computing devices and the energy consumption of computing nodes (EC ). In addi-
edges stands for the capacity of the communication links tion, some limited effort has been focused on balancing of
between them (see the example of application and physi- computing and communication load to more easily satisfy the
cal graphs in Fig. 18 for the face recognition application). requirements on execution delay and/or to minimize overall
The authors firstly propose the algorithm finding the optimal resources utilization.
solution for the linear application graph and, then, more gen- A common drawback of all proposed solutions is that only
eral online approximation algorithms. The numerical results simulations are provided to demonstrate proposed solutions
demonstrate that the proposed algorithm is able to outperform for allocation of MEC computing resources. Moreover, all

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1646 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

papers disregard mobility of the UEs. Of course, if the UEs

are fixed, individual proposal yield a satisfactory execution
delay and/or power consumption at the computing nodes.
Nevertheless, if the UE moves far away from the computing
nodes, this could result in significant QoS degradation due to
long transmission latency and extensive users dissatisfaction.
This issue is addressed in the subsequent section targeting
mobility management for the MEC.

VII. M OBILITY M ANAGEMENT FOR MEC

In the conventional mobile cellular networks, a mobility of
users is enabled by handover procedure when the UE changes Fig. 19. Principle of CaPC according to [108] and [109].
the serving eNB/SCeNB as it roams throughout the network
to guarantee the service continuity and QoS. Analogously, if
the UE offloads computation to the MEC, it is important to control is able to successfully deliver only roughly 80% of
ensure the service continuity. In fact, there are several options offloaded applications.
how to cope with the mobility of UEs. The first option, appli- The main disadvantage of the CaPC presented in [109] is
cable only for the UEs with a low mobility (e.g., within a that the time when the fine adjustment of the transmission
room), is to adapt transmission power of the eNB/SCeNB dur- power is triggered (t ) is the same for all SCeNBs and UEs
ing the time when the offloaded application is processed by the independently on the channel quality (i.e., SINR). As a con-
MEC (Section VII-A). If the UE performs handover to the new sequence, the CaPC may be triggered too late when sufficient
serving eNB/SCeNB despite of the power control, the service SINR cannot be guaranteed in due time to successfully deliver
continuity may be guarantee either by the VM migration (i.e., the offloaded application back to the UE. This problem is
the process during which the VM run at the current comput- addressed in [109], where the t is set individually for each
ing node(s) is migrated to another, more suitable, computing UEs depending on its current channel quality. The proposed
node(s) as discussed in Section VII-B) or by selection of a algorithm finds t by iterative process when t is adapted
new communication path between the UE and the computing after each application is successfully delivered back to the UE.
node (Section VII-C). This way, the amount of successfully delivered applications is
increased up to 98%, as demonstrated by simulations.
A. Power Control
In case when the UEs’ mobility is low and limited, e.g.,
when the UEs are slowly moving inside a building, a proper B. VM Migration
setting of the transmission power of the serving and/or If the UEs mobility is not limited, as considered in
neighboring SCeNBs can help to guarantee QoS. This is Section VII-A, and power control is no longer sufficient to
considered in [108], where Mach and Becvar propose a cloud- keep the UE at the same serving eNB/SCeNB, a possibility to
aware power control (CaPC) algorithm helping to manage the initiate the VM migration should be contemplated in order to
offloading of real-time applications with strict delay require- guarantee the service continuity and QoS requirements. On one
ments. The main objective of the CaPC is to maximize the hand, the VM migration has its cost (CostM ) representing the
amount of the offloaded applications processed by the MEC time required for the VM migration and backhaul resources
with a given latency constrain. This is achieved by an adap- spent by transmission of the VM(s) between the computing
tation of the transmission power of the SCeNBs so that the nodes. On the other hand, there is a gain if the VM migration
handover to a new SCeNB is avoided if possible (see the basic is initiated (GainM ) since the UE can experience lower latency
principle in Fig. 19 where the moving UE remains connected (data is processed in UE’s vicinity) and backhaul resources do
to the same SCeNB as its transmission power is increased). not have to be allocated for transmission of the computation
The CaPC is composed of coarse and fine settings of the results back to the UE.
SCeNBs transmission power. The purpose of the coarse set- A preliminary analysis how the VM migration influences
ting is to find an optimal default transmission power Pt,def , performance of the UE is tackled in [110]. The authors
which is applied if all of the UEs attached to the SCeNB are describe analytical model based on Markov chains. Without
idle. Setting of the Pt,def depends on the power level received the VM migration, a probability that the UE is connected to
by the serving SCeNB from the most interfering neighbor- the optimal MEC decreases with increasing number of hops
ing SCeNB and the interference generated by the eNBs. The between the UE and the eNB, where the service is initially
fine setting consists in a short-term adaptation of the SCeNB’s placed. This also results in increasing delay. Contrary, the con-
transmission power when the UE would not be able to receive nection of the UE to the optimal MEC server results in the
the offloaded application from the cloud due to low SINR. If lowest delay but at the high cost of the migration. The rea-
the CaPC is utilized, up to 95% applications computed at the son for this phenomenon is that the VM migration should be
SCeNBSs are successfully delivered back to the UE with satis- ideally initiated after each handover performed by the UE to
fying delay. Contrary, a conventional, non-cloud-aware, power keep minimum delay.

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1647

as the offset). The main objective of the paper is to minimize

the overall sum cost by the optimal VM migration decision
(i.e., the VM migration is performed if CostM < GainM as
explained earlier). The authors proof the existence of the opti-
mal threshold policy and propose an iterative algorithm in
order to find the optimal thresholds for the VM migration. The
time complexity of the algorithm is O(|M|N), where M and
N is the maximum negative and positive offset, respectively.
The performed results proof the optimal threshold policy is
able to always outperform “never migrate” or “always migrate”
strategies in terms of the sum cost.
The main drawback of [111] and [113] is that these assume
Fig. 20. VM migration principle according to [111]. simple 1D mobility model. More general setting for the VM
migration is contemplated in [114], where 2D mobility and
real mobility traces are assumed. The authors formulate a
While the previous paper is more general and focused on sequential decision making problem for the VM migration
preliminary analysis regarding the VM migration, the main using MDP and define algorithm for finding optimal policy
objective of [111] is to design a proper decision policy deter- with the complexity O(N 3 ), where N is the number of states
mining whether to initiate the VM migration or not. As (note that the state is defined as the number of hops between
discussed above, there is a trade-off between the migration the eNB to which the UE is connected and the location of the
cost (CostM ) and migration gain (GainM ). The authors formu- MEC server analogously to [113]). Since the proposed optimal
late the VM migration policy as a Continuous Time MDP VM migration strategy is too complex, the authors propose
(CTMDP) and they try to find an optimal threshold pol- an approximation of the underlying state space by defining
icy when the VM migration is initiated. Consequently, after the space as a distance between the UE and the MEC server
each handover to the new eNB, the optimal threshold pol- where the service is running. In this case, the time complexity
icy decides whether the VM migration should be initiated or is reduced to O(N 2 ). As demonstrated by the numerical eval-
not. An example of this principle is shown in Fig. 20, where uations, the proposed migration strategy is able to decrease
the UE exploiting the MEC1 moves from the eNB1 to the sum cost by roughly 35% compared to both never and always
eNBn. While the conventional radio handover is performed migrate strategy.
whenever the UE crosses cell boundaries, the VM migration The VM migration process may be further improved by a
is initiated after handover to the eNBn is performed since mobility prediction as demonstrated in [115]. The proposed
CostM < GainM . Simulations show, that the proposed opti- scheme is able to: 1) estimate in advance a throughput that
mal policy always achieves the maximum expected gain if user can receive from individual MEC servers as it roams
compared to never migrate strategy (i.e., the computation is throughout the network, 2) estimate time windows when the
located still at the same MEC) and the scheme when the VM user perform handover, and 3) and VM migration manage-
migration is performed after a specific number of handovers ment scheme selecting the optimal MEC servers according to
(10 handovers is set in the paper). offered throughput. The simulation results demonstrate that
A proper trade-off between VM migration cost (CostM ) and the proposed scheme is able to decrease latency by 35% with
VM migration gain (GainM ) is also studied in [112]. The paper respect to scheme proposed in [111]. Nonetheless, a disad-
proposes a Profit Maximization Avatar Placement (PRIMAL) vantage of the proposal is that it requires huge amount of
strategy deciding whether the VM should be migrated or information in order to predict the throughput. Moreover, the
not. Since the PRIMAL problem is NP-hard, the authors use paper does not consider the cost of migration itself.
Mixed-Integer Quadratic Programming tool to find the heuris- In [116], the VM migration decision process is further
tic solution. The proposed solution is able to significantly enhanced by the mechanism predicting future migration cost
reduce execution delay when compared to the situation with with specified upper bound on a prediction error. The main
no migration (roughly by 90%) while reducing the migra- objective of the paper is, similarly as in [113] and [114], to
tion cost approximately by 40%. When compared to [111], minimize the sum cost over a given time. First, the authors
Ksentini et al. also show the influence of α parameter weighing propose an offline algorithm for finding the optimal placement
CostM and GainM . Basically, with increasing α, the migration sequence for a specific look-ahead window size T, which rep-
cost is decreasing (i.e., migration is not done so frequently), resents the time to which the cost prediction is done. For the
but at the cost of higher execution delay. offline algorithm, an arrival and a departure of the applica-
An optimal threshold policy for the VM migration is also tions offloaded to the MEC are assumed to be exactly known.
considered in [113]. The problem is again formulated as the The time complexity of the algorithm is O(M 2 T), where M
MDP and the VM migration is initiated always if the state of stands for the number of MEC serves in the system. The VM
the UE is bounded by a particular set of thresholds. The state migration is strongly dependent on the size of T. If T is too
of the UE is defined as the number of hops between the eNB to large, the future predicted values may be far away from the
which the UE is connected and the location of the MEC server actual values and, thus, the VM migration far from the opti-
where the computing service is running (in the paper labelled mal. Contrary if T is too short, a long term effect of the service

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1648 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

placement is not considered. As a result, also a binary search

algorithm finding the optimal window size is proposed in the
paper. The proposed offline algorithm is able to reduce cost
by 25% (compared to never migrate strategy) and by 32%
(compared to always migrate strategy). Although the simu-
lation results are demonstrated for the multi-UEs scenario,
the problem is formulated only for the single-UE. Hence, the
paper is further extended in [117] for the multi-UEs offload-
ing K applications to the MEC. Similarly as in [116], the
problem is solved by the offline algorithm with complexity
of O(M K2 T). Since the offline algorithm is of high complex-
Fig. 21. An example of path selection algorithm proposed in [122] (cR x and
ity and impractical for real systems, the paper also propose CxB stands for capacity of radio links and backhaul links, respectively).
an online approximation algorithm reducing the complexity to
O(M 2 KT). The proposed online algorithm outperforms never
migrate and always migrate strategies by approximately 32% A proper VM migration may not result only in an exe-
and 50%, respectively. cution delay reduction, but it can also increase throughput
So far, all the studies focusing on the VM migration do not of the system as demonstrated in [120]. The paper pro-
consider an impact on a workload scheduling, i.e., how the poses a protocol architecture for cloud access optimization
VM migration would be affected by a load of individual MEC (PACAO), which is based on Locator/Identifier Separation
servers. As suggested in [118], the problem of the VM migra- Protocol (LISP) [121]. If the user is experiencing latency or
tion and scheduling of the MEC workloads should be done jitter above maximum tolerated threshold, the VM migration
jointly. Although the problem could be formulated as a sequen- to a new MEC server is initiated. The selection of the new
tial decision making problem in the framework of MDPs (like MEC server, which is about to host VM of the user is based
in above studies) it would suffer from several drawbacks, on the required computing power and availability of resources
such as, 1) extensive knowledge of the statistics of the users at the MEC servers. The proposal is evaluated by means of
mobility and request arrival process is impractical, 2) problem both experiments on real testbed and simulations. The results
can is computationally challenging, and 3) any change in show that the system throughput is increased by up to 40%
the mobility and arrival statistics would require re-computing when compared to the case without VM migration.
the optimal solution. Hence, the main contribution of [118]
is a development of a new methodology overcoming these
drawbacks inspired by Lyapunov optimization framework. The C. Path Selection and/or VM Migration
authors propose online control algorithm making decision on The VM migration is not a convenient option when a huge
where the application should be migrated so that the over- amount of data needs to be migrated among the comput-
all transmission and reconfiguration costs are minimized. The ing nodes and the whole process may take minutes or even
complexity of the algorithm is O(M!/(M − K)!), where M is hours [119]. Even if the migration process lasts few sec-
the number of MEC servers and K is the amount of applica- onds, real-time applications cannot be offloaded to the MEC.
tions host by the MEC. By means of proposed optimization Moreover, the load imposed on backhaul links may be too sig-
framework, the reconfiguration cost is reduced when com- nificant. In such cases, finding and optimizing new paths for
pared to always migrate strategy (by 7%) and never migrate delivery of the computed data from the MEC are a more viable
strategy (by 26%). option. This eventuality is considered in [122], where the path
While the main objective of the previous papers focusing selection algorithm for a delivery of the offloaded data from
on the VM migration is to make a proper decision on whether the cluster of computing SCeNBs to the UE is proposed. The
to migrate or not, the main aim of Ha et al. [119] is to mini- main objective of the path selection algorithm is to minimize
mize the VM migration time when the migration is about to be transmission delay taking into account quality of both radio
performed. This is accomplished by a compression algorithm and backhaul links. Moreover, the authors enable to enforce
reducing the amount of transmission data during the migration handover to new serving SCeNB to minimize the transmission
itself. On one hand, if the compression rate of the algorithm delay. An example of the data delivery is shown in Fig. 21,
is low, more data has to be transmitted, but the compression where three SCeNBs are computing the application offloaded
itself is shorter in terms of time. On the other hand, a higher by the UE. The data processed by the serving SCeNB2 are
compression rate results in a significant reduction of the trans- received by the UE directly while data from the SCeNB3 are
mitted data during the VM migration, but the compression delivered to the UE through the CN and the serving SCeNB2.
takes significant amount of time. Hence, the paper proposes Finally, the UE performs handover to the SCeNB1 and, then,
a dynamic adaptation of the compression rate depending on it receives results from the SCeNB3 directly via radio link.
the current backhaul available bandwidth and the processing The complexity of the proposed algorithm is O(mn ), where
load of the MEC. The paper presents extensive experiments m is the number of UEs and n the amount of the SCeNBs
on real system showing that the dynamic adaptation during in cluster. The proposed algorithm is able to reduce transmis-
the VM migration is able to cope with changing of available sion delay by up to 9% with respect to a case when the UE
bandwidth capacity. receives all data from the same serving SCeNB. In [123], the

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1649

TABLE V
C OMPARISON OF I NDIVIDUAL PAPERS F OCUSING ON M OBILITY M ANAGEMENT IN MEC

algorithm’s complexity is further decreased to O(I n ), where I an algorithm for the dynamic VM migration and the path
is the set of SCeNBs with sufficient radio/backhaul link qual- selection algorithm proposed in [123] further enhanced by
ity. It is shown that the proposed path selection algorithm is consideration of a mobility prediction. The first algorithm
able to reduce transmission delay by 54%. decides whether the VM migration should be initiated
The path selection algorithm contemplated or not based on the mobility prediction and the com-
in [122] and [123] may not be sufficient if the UE is too putation/communication load of the eNB(s). The second
far away from the computing location since increased trans- algorithm, then, finds the most suitable route for downloading
mission delay may result in QoS reduction notwithstanding. the offloaded data with the mobility prediction outcomes
Hence, Plachy et al. [124] suggest a cooperation between taken into account. The complexity of the first algorithm is

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1650 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

O(|Z||I|τ ) and the complexity of the second algorithm equals time, sending only small amount of data [82]. The rea-
to O(|I|τ ), where Z is the number of eNBs with sufficient son is that the energy spent by transmission/reception of
channel quality and computing capacity, and τ stands for the the offloaded computing is small while the energy savings
size of the prediction window. The proposed algorithm is achieved by the computation offloading are significant.
able reducing the average offloading time by 27% comparing Contrary, the applications that need to offload a lot
to the situation when the VM migration is performed after of data should be computed locally as the offloading
each conventional handover and by roughly 10% with respect simply does not pay off due to huge amount of energy
to [123]. spent by the offloading and high offloading delays.
• If the computing capacities at the MEC are fairly
D. Summary of Works Focused on Mobility Management limited, the probability to offload data for processing
is lowered. This is due to the fact that the probabilities
A comparison of the studies addressing the mobility issues
of the offloading and local processing are closely related
for the MEC is shown in Table V. As it can be observed
to the computation power available at the MEC.
from Table V, the majority of works so far focuses on the
• With more UEs in the system, the application offload-
VM migration. Basically, the related papers try to find an
ing as well as its processing at the MEC last
optimal decision policy whether the VM migration should be
longer [96]. Consequently, if there is high amount of UEs
initiated or not to minimize overall system cost (up to 32%
in the system, the local processing may be more prof-
and up to 50% reduction of average cost is achieved com-
itable, especially if the minimization of execution delay
pared to never and always migrate options, respectively [117]).
is the priority (such is the case of real-time applications).
Moreover, some papers aim to find a proper trade-off between
• The energy savings achieved by the computation
VM migration cost and VM migration gain [112], minimizing
offloading is strongly related to the radio access
execution delay [115], minimizing VM migration time [119],
technology used at radio link. To be more specific,
or maximizing overall throughput [120].
OFDMA enables significantly higher energy savings of
From Table V can be further observed that all papers dealing
the UEs than TDMA due to higher granularity of radio
with the VM migration assume the computation is done by a
resources [92].
single computing node. Although this option is less complex,
• The partial offloading can save significantly more
the parallel computation by more nodes should not be entirely
energy at the UE when compared to the full offload-
neglected as most of the papers focusing on the allocation of
ing [72]. Nevertheless, in order to perform the partial
computing resources assume multiple computing nodes (see
offloading, the application has to enable paralleliza-
Section VI-B).
tion/partitioning. Hence, the energy savings accom-
plished by computation offloading is also strongly
VIII. L ESSONS L EARNED related to the application type and the way how the
This section summarizes lessons learned from the state of code of the application is written.
the art focusing on computation offloading into the MEC. We From the surveyed papers focused on allocation of comput-
again address all three key items: decision on computation ing resources, the following key facts are learned:
offloading, allocation of computing resources, and mobility • The allocation of computation resources is strongly
management. related to the type of the application being offloaded
From the surveyed papers dealing with the decision on com- in a sense that only applications allowing paralleliza-
putation offloading, following key observations are derived: tion/partitioning may be distributed to multiple computing
• If the channel quality between the UE and its serv- nodes. Obviously, a proper parallelization and code
ing station is low, it is profitable to compute rather partitioning of the offloaded application can result
locally [95]. The main reason is that the energy spent in shorter execution delays as multiple nodes may
by the transmission/reception of the offloaded data is too pool their computing resources (up to 90% reduction
expensive in terms of the energy consumption at the UE. of execution delay when compared to single computing
Contrary, with increasing quality of the channel, it is node). On the other hand, the allocation of computation
better to delegate the computation to the MEC since resources for parallelized applications is significantly
the energy cost required for transmission/reception of the more complex.
offloaded data is reduced and it is easily outweighed • An increase in the number of computing nodes does
by the energy saving due to the remote computation. not have to result always in a reduction in the
Consequently, the computation can be offloaded more execution delay [102]. On the contrary, if the commu-
frequently if MIMO is exploited as it improves channel nication delay becomes predominant over the compu-
quality. Moreover, it is efficient to exploit connection tation delay, the overall execution delay may be even
through SCeNBs for the offloading as the SCeNBs are increased. Hence, a proper trade-off between the number
supposed to serve fewer users in proximity providing high of computing nodes and execution delay needs to be care-
channel quality and more available radio resources. fully considered when allocating computing resources to
• The most suitable applications for offloading are those offloaded data.
requiring high computational power (i.e., high com- • If the backhaul is of a low quality, it is mostly
putational demanding applications) and, at the same preferred to perform the computation locally by the

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1651

serving node (e.g., SCeNB/eNB) since the distribution algorithms should be employed with purpose to find
of data for computing is too costly in terms of the trans- the optimal path for delivery of the offloaded data
mission latency. Contrary, a high quality backhaul is back to the UEs while computing is done by the same
a prerequisite for an efficient offloading to multiple node(s) (i.e., without VM migration) [123]. However, if
computing nodes. the UE moves too far away from the computation
• The execution delay of the offloaded application placement, more robust mobility management based
depends not only on the backhaul quality, but also on on joint VM migration and path selection should be
a backhaul topology (e.g., mesh, ring, tree, etc.) [102]. adopted [124].
The mesh topology is the most advantageous in terms of
the execution delay since all computing nodes are con- IX. O PEN R ESEARCH C HALLENGES
nected directly and distribution of the offloaded data for AND F UTURE W ORK
computing is more convenient. On the other hand, mesh As shown in the previous sections, the MEC has attracted
topology would require huge investment in the backhaul. a lot of attention in recent years due to its ability to signif-
Finally, after surveying the papers addressing mobility icantly reduce energy consumption of the UEs while, at the
issues in the MEC, we list following key findings: same time, enabling real-time application offloading because
• There are several options of the UE’s mobility manage-
of proximity of computing resources to the users. Despite this
ment if the data/application is offloaded to the MEC. In fact the MEC is still rather immature technology and there are
cases of the low mobility, the power control at the many challenges that need to be addressed before its imple-
SCeNBs/eNBs side can be sufficient to handle mobility mentation into mobile network to be beneficial. This section
(up to 98% of offloaded applications can be successfully discusses several open research challenges not addressed by
delivered back to the UE [109]). This is true as long as the the current researcher.
adaption of transmission power enables keeping the UE at
the same serving station during the computation offload- A. Distribution and Management of MEC Resources
ing. However, if the UE performs handover, the power
control alone is not sufficient and the VM migration or In Section III, we have discussed several possible options
new communication path selection may be necessary to for placement of the computing nodes enabling the MEC
comply with requirements of offloaded applications in within the mobile network architecture. To guarantee ubiq-
terms of latency. uitous MEC services for all users wanting to utilize the MEC,
• A decision on VM migration depends strongly on three
the MEC servers and the computation/storage resource should
metrics: be distributed throughout whole network. Consequently, the
1) The VM migration cost (CostM ) representing the individual options where to physically place the MEC servers
time required for the service migration and the back- should complement each other in a hierarchical way. This will
haul resources spent by the transmission of VM(s) allow efficient usage of the computing resources while respect-
between the computing nodes. ing QoS and QoE requirements of the users. In this context,
2) The VM migration gain (GainM ) is the gain con- an important challenge is to find an optimal way where to
stituting delay reduction (data are computed in physically place the computation depending on expected users
proximity of the UE) and saving of the backhaul demands while, at the same time, consider related CAPEX and
resources (data does not have to be sent through OPEX (as initially tackled in [67] and [68]).
several nodes). Another missing topic in the literature is a design of effi-
3) The computing load of the node(s) to which the cient control procedures for proper management of the MEC
VM is reallocated since, in some situations, the opti- resources. This includes design of signalling messages, their
mal computing node for the VM migration may be exchange and optimization in terms of signalling overhead.
unavailable due to its high computation load. The control messages should be able to deliver status informa-
• The VM migration is impractical if huge amount of
tion, such as load of individual computing nodes and quality of
data needs to be transmitted between the comput- wireless/backhaul links in order to efficiently orchestrate com-
ing nodes and/or if the backhaul resources between puting resources within the MEC. There is a trade-off between
VMs are inadequate since it may take minutes or even high signalling overhead related to frequent exchange of the
hours to migrate whole VM. This is obviously too long status information and an impact on the MEC performance
for real-time services and it also implies significant load due to aging of the status information if these are exchanged
on backhaul, especially if the VM migration would need rarely. This trade-off have to be carefully analysed and effi-
to be performed frequently. Note that time consuming cient signalling mechanisms need to be proposed to ensure that
migration goes against the major benefit of the MEC, the control entities in the MEC have up to date information at
i.e., low latency resulting in suitability of the offloading their disposal while the cost to obtain them is minimized.
for real-time services.
• The minimization of the VM migration time can B. Offloading Decision
be done by reduction of the amount of migrated The offloading decision plays a crucial part as it basi-
data [119]. Nonetheless, even this option is not enough cally determines whether the computation would be performed
for real-time services. Thus, various path selection locally, remotely or jointly in both locations as discussed in

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1652 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

Section V. All papers focusing on the offloading decision con- D. Mobility Management
sider only the energy consumption at the side of the UE. So far, the works focusing on mobility management and
However, to be in line with future green networking, also particularly on the VM migration consider mostly a scenario
the energy consumption at the MEC (including computation when only a single computing node (SCeNB or eNB) makes
as well as related communication) should be further taken computation for each UE. Hence, the challenge is how to
into account during the decision. Moreover, all papers dealing efficiently handle the VM migration procedure when appli-
with the offloading decision assume strictly static scenarios, cation is offloaded to several computing nodes. Moreover, the
i.e., the UEs are not moving before and during the offload- VM migration impose high load on the backhaul and leads
ing. Nevertheless, the energy necessary for transmission of to high delay, which makes it unsuitable for real-time appli-
the offloaded data can be significantly changed even during cations. Hence, new advanced techniques enabling very fast
offloading if channel quality drops due to low movement or VM migration in order of milliseconds should be developed.
fading. This can result in the situation when the offloading However, this alternative is very challenging due to commu-
may actually increase the energy consumption and/or exe- nication limits between computing nodes. Therefore, more
cution delay comparing to local computation. Hence, it is realistic challenge is how to pre-migrate the computation in
necessary to propose new advanced methods for the offloading advance (e.g., based on some prediction techniques) so that
decision, for instance, exploiting various prediction techniques there would be no service disruption observed by the users.
on the UEs mobility and channel quality during the offload- Despite of above-mentioned suggestions potentially reduc-
ing to better estimate how much the offloading will cost for ing VM migration time, stand-alone VM migration may
varying conditions. be unsuitable for real-time applications notwithstanding.
Besides, current papers focusing on the partial offloading Consequently, it is important to aim majority of research effort
decision disregard the option to offload individual parts to towards a cooperation of the individual techniques for mobility
multiple computing nodes. Multiple computing nodes enables management. In this regard, dynamic optimization and joint
higher flexibility and increases a probability that the offload- consideration of all techniques (such as power control, VM
ing to the MEC will be efficient for the UE (in terms of both migration, compression of migrated data, and/or path selec-
energy consumption and execution delay). Of course, a sig- tion) should be studied more closely in order to enhance QoE
nificant challenge in this scenario belongs to consideration of for the UEs and to optimize overall system performance for
backhaul between the MEC servers and ability to reflect their moving users.
varying load and parameters during the offloading decision.

C. Allocation of Computing Resources E. Traffic Paradigm Imposed by Coexistence of Offloaded

The studies addressing the problem of an efficient alloca- Data and Conventional Data
tion of the computing resources for the application offloaded Current research dealing with the decision on computation
to the MEC do not consider dynamicity of the network. To be offloading, allocation of computing resources and mobility
more precise, the computing nodes (e.g., SCeNBs, eNB) are management mostly neglects the fact that conventional data
selected in advance before the application is offloaded to the not offloaded to the MEC, such as VoIP, HTTP, FTP, machine
MEC and then the same computing node(s) is (are) assumed type communication, video streaming, etc., has to be transmit-
to process the offloaded application (at least as long as the UE ted over radio and backhaul links in parallel to the offloaded
is relatively static and does not perform handover among cells data. Hence, whenever any application is being offloaded to
as considered in Section VII). However, if some additional the MEC, it is necessary to jointly allocate/schedule commu-
computing resources are freed while given application is pro- nication resources both for the offloaded data to the MEC
cessed at the MEC, these resources could be also allocated and the conventional data (i.e., data not exploiting MEC) in
for it in order to farther speed up the offloaded computing. order to guarantee QoS and QoE. Especially, if we consider
Hence, a dynamic allocation of the computing resources dur- the fact that the offloaded data represents additional load on
ing processing of the offloaded applications in the MEC is an already resource starving mobile cellular networks. The effi-
interesting research challenge to be addressed in the future. cient scheduling of the communication resources may also
So far all the studies focusing on the allocation of com- increase the amount of data offloaded to the MEC because of
puting resources assume a “flat” MEC architecture in a sense more efficient utilization radio and backhaul communication
that the MEC computing nodes are equally distributed and links.
of the same computing power. In this respect, it would be Besides, the offloading reshapes conventional perception of
interesting to consider more hierarchical placement of the uplink/downlink utilization as the offloading is often more
computing nodes within the MEC. More specifically, com- demanding in terms of the uplink transmission (offloading
puting resources should be distributed within the network as from the UE to the MEC). The reason for this is that
described in Section III-B3 (e.g., cluster of SCeNBs, eNBs, many applications require delivering of large files/data to
aggregation points or even at the edge of CN). A hierarchi- the MEC for processing (e.g., image/video/voice recognition,
cal MEC placement should result in a better distribution of file scanning, etc.) while the results delivered to the UE are
the computing load and a lower execution delay experienced of significantly lower volume. This paradigm motivates for
by the users since the use of distant CC can be more easily rethinking and reshaping research effort from sole downlink
avoided. to the mixed downlink and uplink in the future.

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1653

F. Concept Validation [7] A. Mtibaa, A. Fahim, K. A. Harras, and M. H. Ammar, “Towards

resource sharing in mobile device clouds: Power balancing across
As shown in Sections V–VII, the MEC concept is analyzed mobile devices,” in Proc. ACM SIGCOMM Workshop Mobile Cloud
and novel algorithms and proposed solutions are validated typ- Comput., Hong Kong, 2013, pp. 51–56.
ically by numerical analysis or by simulations. In addition, [8] A. Mtibaa, K. A. Harras, and A. Fahim, “Towards computational
offloading in mobile device clouds,” in Proc. IEEE Int. Conf. Cloud
majority of work assume rather simple, and sometimes unre- Comput. Technol. Sci., Bristol, U.K., 2013, pp. 331–338.
alistic, scenarios for simplification of the problem. Although [9] T. Nishio, R. Shinkuma, T. Takahashi, and N. B. Mandayam, “Service-
these are a good starting point in uncovering MEC potentials, oriented heterogeneous resource sharing for optimizing service latency
in mobile cloud,” in Proc. 1st Int. Workshop Mobile Cloud Comput.
it is important to validate key principles and findings by means Netw., Bengaluru, India, 2013, pp. 19–26.
of simulations under more complex and realistic situations and [10] K. Habak, M. Ammar, K. A. Harras, and E. Zegura, “Femto clouds:
scenarios such as, e.g., [114]–[118] where at least real world Leveraging mobile devices to provide cloud service at the edge,” in
Proc. IEEE Int. Conf. Cloud Comput., New York, NY, USA, 2015,
user mobility traces are considered for evaluation and propos- pp. 9–16.
als on VM migration. At the same time, massive trials and [11] F. Liu et al., “Gearing resource-poor mobile devices with powerful
further experiments in emulated networks (like initially pro- clouds: Architectures, challenges, and applications,” IEEE Wireless
Commun., vol. 20, no. 3, pp. 14–22, Jun. 2013.
vided in [42]) or real networks (similar to those just recently [12] I. Yaqoob et al., “Mobile ad hoc cloud: A survey,” Wireless Commun.
performed by Nokia [46]) are mandatory to move the MEC Mobile Comput., vol. 16, no. 16, pp. 2572–2589, 2016.
concept closer to the reality. [13] C. Ragona, F. Granelli, C. Fiandrino, D. Kliazovich, and P. Bouvry,
“Energy-efficient computation offloading for wearable devices and
smartphones in mobile cloud computing,” in Proc. IEEE Glob.
Commun. Conf. (GLOBECOM), San Diego, CA, USA, 2015, pp. 1–6.
X. C ONCLUSION [14] W. Zhang, Y. Wen, J. Wu, and H. Li, “Toward a unified elastic comput-
ing platform for smartphones with cloud support,” IEEE Netw., vol. 27,
The MEC concept brings computation resources close to no. 5, pp. 34–40, Sep./Oct. 2013.
the UEs, i.e., to the edge of mobile network. This enables to [15] F. Bonomi, R. Milito, J. Zhu, and S. Addepalli, “Fog computing and its
offload highly demanding computations to the MEC in order role in the Internet of Things,” in Proc. MCC Workshop Mobile Cloud
to cope with stringent requirements of applications on latency Comput., Helsinki, Finland, 2012, pp. 13–16.
[16] J. Zhu et al., “Improving Web sites performance using edge servers in
(e.g., real time applications) and to reduce energy consump- fog computing architecture,” in Proc. IEEE Int. Symp. Service Oriented
tion at the UE. Although the research on the MEC gains its Syst. Eng., San Francisco, CA, USA, 2013, pp. 320–323.
momentum, as reflected in this survey after all, the MEC [17] I. Stojmenovic and S. Wen, “The fog computing paradigm: Scenarios
and security issues,” in Proc. Federated Conf. Comput. Sci. Inf. Syst.,
itself is still immature and highly unproved technology. In Warsaw, Poland, 2014, pp. 1–8.
this regard, the MEC paradigm introduces several critical chal- [18] I. Stojmenovic, “Fog computing: A cloud to the ground support
lenges waiting to be addressed to the full satisfaction of all for smart things and machine-to-machine networks,” in Proc. Aust.
Telecommun. Netw. Appl. Conf. (ATNAC), 2014, pp. 117–122.
involved parties such as mobile operators, service providers, [19] T. H. Luan et al. (2016). Fog Computing: Focusing on Mobile Users
and users. The alpha and the omega of current research regard- at the Edge. [Online]. Available:https://arxiv.org/abs/1502.01815
ing the MEC is how to guarantee service continuity in highly [20] M. Yannuzzi, R. Milito, R. Serral-Gracia, D. Montero, and
M. Nemirovsky, “Key ingredients in an IoT recipe: Fog computing,
dynamic scenarios. This part is lacking in terms of research cloud computing, and more fog computing,” in Proc. Int. Workshop
and is one of the blocking point to enroll the MEC concept. Comput.-Aided Model. Design Commun. Links Netw. (CAMAD),
Moreover, recent research validates solution mostly under very Athens, Greece, 2014, pp. 325–329.
[21] A. Checko et al., “Cloud RAN for mobile networks—A technology
simplistic scenarios and by means of simulations or analytical overview,” IEEE Commun. Surveys Tuts., vol. 17, no. 1, pp. 405–426,
evaluations. Nevertheless, to demonstrate the expected val- 1st Quart., 2015.
ues introduced by the MEC, real tests and trials under more [22] D. Kliazovich and F. Granelli, “Distributed protocol stacks: A frame-
work for balancing interoperability and optimization,” in Proc. IEEE
realistic assumptions are further required. Int. Conf. Commun. Workshop (ICC), 2008, pp. 241–245.
[23] J. Cheng, Y. Shi, B. Bai, and W. Chen, “Computation offloading in
cloud-RAN based mobile cloud computing system,” in Proc. IEEE Int.
R EFERENCES Conf. Commun. (ICC), Kuala Lumpur, Malaysia, 2016, pp. 1–6.
[24] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young, “Mobile
[1] H. T. Dinh, C. Lee, D. Niyato, and P. Wang, “A survey of mobile edge computing—A key technology towards 5G,” ETSI, Sophia
cloud computing: Architecture, applications, and approaches,” Wireless Antipolis, France, White Paper, vol. 11, 2015.
Commun. Mobile Comput., vol. 13, no. 18, pp. 1587–1611, 2013. [25] Z. Sanaei, S. Abolfazli, A. Gani, and R. Buyya, “Heterogeneity
[2] S. Barbarossa, S. Sardellitti, and P. Di Lorenzo, “Communicating while in mobile cloud computing: Taxonomy and open challenges,” IEEE
computing: Distributed mobile cloud computing over 5G heteroge- Commun. Surveys Tuts., vol. 16, no. 1, pp. 369–392, 1st Quart., 2014.
neous networks,” IEEE Signal Process. Mag., vol. 31, no. 6, pp. 45–55, [26] Y. Wen, X. Zhu, J. J. P. C. Rodrigues, and C. W. Chen, “Cloud mobile
Nov. 2014. media: Reflections and outlook,” IEEE Trans. Multimedia, vol. 16,
[3] A. R. Khan, M. Othman, S. A. Madani, and S. U. Khan, “A survey of no. 4, pp. 885–902, Jun. 2014.
mobile cloud computing application models,” IEEE Commun. Surveys [27] A. Ahmed and E. Ahmed, “A survey on mobile edge computing,” in
Tuts., vol. 16, no. 1, pp. 393–413, 1st Quart., 2014. Proc. IEEE Int. Conf. Intell. Syst. Control (ISCO), Coimbatore, India,
[4] M. Satyanarayanan, P. Bahl, R. Caceres, and N. Davies, “The case for 2016, pp. 1–8.
VM-based cloudlets in mobile computing,” IEEE Pervasive Comput., [28] R. Roman, J. Lopez, and M. Mambo, “Mobile edge computing,
vol. 8, no. 4, pp. 14–23, Oct./Dec. 2009. Fog et al.: A survey and analysis of security threats and challenges,”
[5] C. Shi, V. Lakafosis, M. H. Ammar, and E. W. Zegura, “Serendipity: Future Gener. Comput. Syst., to be published.
Enabling remote computing among intermittently connected mobile [29] N. C. Luong, P. Wang, D. Niyato, Y. Wen, and Z. Han, “Resource
devices,” in Proc. ACM Int. Symp. Mobile Ad Hoc Netw. Comput., management in cloud networking using economic analysis and pricing
Hilton Head Island, SC, USA, 2012, pp. 145–154. models: A survey,” IEEE Commun. Surveys Tuts., to be published.
[6] U. Drolia et al., “The case for mobile edge-clouds,” in Proc. IEEE [30] E. Cuervo et al., “MAUI: Making smartphones last longer with code
10th Int. Conf. Ubiquitous Intell. Comput. IEEE 10th Int. Conf. Auton. offload,” in Proc. Int. Conf. Mobile Syst. Appl. Services (Mobysis),
Trusted Comput., Vietri sul Mare, Italy, 2013, pp. 209–215. San Francisco, CA, USA, 2010, pp. 49–62.

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1654 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

[31] B.-G. Chun, S. Ihm, P. Maniatis, M. Naik, and A. Patti, “CloneCloud: [55] Z. Becvar et al., “Distributed architecture of 5G mobile networks
Elastic execution between mobile device and cloud,” in Proc. Eur. Conf. for efficient computation management in mobile edge computing,”
Comput. Syst. (Eurosys), Salzburg, Austria, 2011, pp. 301–314. in Chapter in 5G Radio Access Network (RAN)—Centralized RAN,
[32] S. Kosta, A. Aucinas, P. Hui, R. Mortier, and X. Zhang, “ThinkAir: Cloud-RAN and Virtualization of Small Cells, H. Venkataraman and
Dynamic resource allocation and parallel execution in the cloud for R. Trestian, Eds. Boca Raton, FL, USA: Taylor and Francis Group,
mobile code offloading,” in Proc. IEEE INFOCOM, Orlando, FL, USA, Mar. 2017.
2012, pp. 945–953. [56] S. Wang et al., “Mobile micro-cloud: Application classification, map-
[33] W. Zhang, Y. Wen, and D. O. Wu, “Energy-efficient scheduling policy ping, and deployment,” in Proc. Annu. Fall Meeting ITA (AMITA),
for collaborative execution in mobile cloud computing,” in Proc. IEEE New York, NY, USA, Oct. 2013, pp. 1–7.
INFOCOM, Turin, Italy, 2013, pp. 190–194. [57] K. Wang et al., “MobiScud: A fast moving personal cloud in the
[34] W. Zhang, Y. Wen, and D. O. Wu, “Collaborative task execution in mobile network,” in Proc. Workshop All Things Cellular Oper. Appl.
mobile cloud computing under a stochastic wireless channel,” IEEE Challenge, London, U.K., 2015, pp. 19–24.
Trans. Wireless Commun., vol. 14, no. 1, pp. 81–93, Jan. 2015. [58] A. Manzalini et al., “Towards 5G software-defined ecosystems:
[35] Y. Wen, W. Zhang, and H. Luo, “Energy-optimal mobile application Technical challenges, business sustainability and policy issues,” white
execution: Taming resource-poor mobile devices with cloud clones,” in paper, 2014.
Proc. IEEE INFOCOM, Orlando, FL, USA, 2012, pp. 2716–2720. [59] T. Taleb and A. Ksentini, “Follow me cloud: Interworking federated
[36] H. Flores et al., “Mobile code offloading: From concept to practice and clouds and distributed mobile networks,” IEEE Netw., vol. 27, no. 5,
beyond,” IEEE Commun. Mag., vol. 53, no. 3, pp. 80–88, Mar. 2015. pp. 12–19, Sep./Oct. 2013.
[37] L. Jiao et al., “Cloud-based computation offloading for mobile devices: [60] T. Taleb, A. Ksentini, and P. A. Frangoudis, “Follow-me cloud: When
State of the art, challenges and opportunities,” in Proc. Future Netw. cloud services follow mobile users,” IEEE Trans. Cloud Comput., to
Mobile Summit, Lisbon, Portugal, 2013, pp. 1–11. be published.
[38] ETSI, “Mobile edge computing (MEC): Technical requirements,” [61] A. Aissioui, A. Ksentini, and A. Gueroui, “An efficient elastic dis-
V1.1.1, Mar. 2016. tributed SDN controller for follow-me cloud,” in Proc. IEEE Int. Conf.
[39] M. T. Beck, M. Werner, S. Feld, and T. Schimper, “Mobile edge com- Wireless Mobile Comput. Netw. Commun. (WiMob), Abu Dhabi, UAE,
puting: A taxonomy,” in Proc. Int. Conf. Adv. Future Internet (AFIN), 2015, pp. 876–881.
Lisbon, Portugal, Nov. 2014, pp. 48–54. [62] J. Liu, T. Zhao, S. Zhou, Y. Cheng, and Z. Niu, “CONCERT: A cloud-
[40] N. Takahashi, H. Tanaka, and R. Kawamura, “Analysis of process based architecture for next-generation cellular systems,” IEEE Wireless
assignment in multi-tier mobile cloud computing and application to Commun., vol. 21, no. 6, pp. 14–22, Dec. 2014.
edge accelerated Web browsing,” in Proc. IEEE Int. Conf. Mobile [63] ETSI, “Mobile edge computing (MEC): Terminology,” V1.1.1,
Cloud Comput. Services Eng., San Francisco, CA, USA, 2015, Mar. 2016.
pp. 233–234. [64] ETSI, “Mobile edge computing (MEC): Proof of concept Framework,”
[41] Y. Zhang, H. Liu, L. Jiao, and X. Fu, “To offload or not to offload: V1.1.1, Mar. 2016.
An efficient code partition algorithm for mobile cloud computing,” in [65] ETSI, “Mobile edge computing (MEC): Service scenarios” V1.1.1,
Proc. 1st Int. Conf. Cloud Netw. (CLOUDNET), Paris, France, 2012, Mar. 2016.
pp. 80–86.
[66] ETSI, “Mobile edge computing (MEC): Framework and Reference
[42] J. Dolezal, Z. Becvar, and T. Zeman, “Performance evaluation of com-
Architecture,” V1.1.1, Mar. 2016.
putation offloading from mobile device to the edge of mobile network,”
[67] A. Ceselli, M. Premoli, and S. Secci, “Cloudlet network design
in Proc. IEEE Conf. Stand. Commun. Netw. (CSCN), Berlin, Germany,
optimization,” in Proc. IFIP Netw., Toulouse, France, 2015, pp. 1–9.
2016, pp. 1–7.
[68] A. Ceselli, M. Premoli, and S. Secci, “Mobile edge cloud network
[43] O. Salman, I. Elhajj, A. Kayssi, and A. Chehab, “Edge computing
design optimization,” IEEE/ACM Trans. Netw., to be published.
enabling the Internet of Things,” in Proc. IEEE 2nd World Forum
Internet Things (WF IoT), Milan, Italy, 2015, pp. 603–608. [69] D. Kreutz et al., “Software-defined networking: A comprehensive
[44] S. Abdelwahab, B. Hamdaoui, M. Guizani, and T. Znati, “REPLISOM: survey,” Proc. IEEE, vol. 103, no. 1, pp. 14–76, Jan. 2015.
Disciplined tiny memory replication for massive IoT devices in LTE [70] N. A. Jagadeesan and B. Krishnamachari, “Software-defined
edge cloud,” IEEE Internet Things J., vol. 3, no. 3, pp. 327–338, networking paradigms in wireless networks: A survey,” ACM Comput.
Jun. 2016. Surveys, vol. 47, no. 2, Jan. 2015, Art. no. 27.
[45] X. Sun and N. Ansari, “EdgeIoT: Mobile edge computing for the [71] X. Jin, L. E. Li, L. Vanbever, and J. Rexford, “SoftCell: Scalable
Internet of Things,” IEEE Commun. Mag., vol. 54, no. 12, pp. 22–29, and flexible cellular core network architecture,” in Proc. ACM Conf.
Dec. 2016. Emerg. Netw. Exp. Technol. (CoNEXT), Santa Barbara, CA, USA, 2013,
[46] NOKIA: Multi-Access Edge Computing. Accessed on pp. 163–174.
Mar. 20, 2017. [Online]. Available: https://networks.nokia.com/ [72] M. Deng, H. Tian, and B. Fan, “Fine-granularity based application
solutions/multi-access-edge-computing offloading policy in cloud-enhanced small cell networks,” in Proc.
[47] NOKIA: Mobile Edge Computing. Accessed on Mar. 20, 2017. IEEE Int. Conf. Commun. Workshops (ICC), Kuala Lumpur, Malaysia,
[Online]. Available: http://resources.alcatel-lucent.com/asset/200546 2016, pp. 638–643.
[48] (2012). FP7 European Project, Distributed Computing, Storage and [73] S. E. Mahmoodi, R. N. Uma, and K. P. Subbalakshmi, “Optimal joint
Radio Resource Allocation Over Cooperative Femtocells (TROPIC). scheduling and cloud offloading for mobile applications,” IEEE Trans.
[Online]. Available: http://www.ict-tropic.eu/ Cloud Comput., to be published.
[49] (2015). H2020 European Project, Small cEllS coordinAtion for [74] J. Liu, Y. Mao, J. Zhang, and K. B. Letaief, “Delay-optimal com-
Multi-tenancy and edge services (SESAM). [Online]. Available: putation task scheduling for mobile-edge computing systems,” in
http://www.sesame-h2020-5g-ppp.eu/ Proc. IEEE Int. Symp. Inf. Theory (ISIT), Barcelona, Spain, 2016,
[50] I. Giannoulakis et al., “The emergence of operator-neutral small cells pp. 1451–1455.
as a strong case for cloud computing at the mobile edge,” Trans. Emerg. [75] Y. Mao, J. Zhang, and K. B. Letaief, “Dynamic computation offloading
Telecommun. Technol., vol. 27, no. 9, pp. 1152–1159, 2016. for mobile-edge computing with energy harvesting devices,” IEEE J.
[51] M. Chiosi et al., “Network functions virtualisation: An introduction, Sel. Areas Commun., vol. 34, no. 12, pp. 3590–3605, Dec. 2016.
benefits, enablers, challenges & call for action,” Introductory white [76] W. Zhang et al., “Energy-optimal mobile cloud computing under
paper, 2012. stochastic wireless channel,” IEEE Trans. Wireless Commun., vol. 12,
[52] ETSI, “Network function virtualisation (NFV): Architectural frame- no. 9, pp. 4569–4581, Sep. 2013.
work,” V1.1.1, Oct. 2013. [77] S. Ulukus et al., “Energy harvesting wireless communications:
[53] F. Lobillo et al., “An architecture for mobile computation offloading A review of recent advances,” IEEE J. Sel. Areas Commun., vol. 33,
on cloud-enabled LTE small cells,” in Proc. Workshop Cloud Technol. no. 3, pp. 360–381, Mar. 2015.
Energy Efficiency Mobile Commun. Netw. (IEEE WCNCW), Istanbul, [78] M. Kamoun, W. Labidi, and M. Sarkiss, “Joint resource allocation and
Turkey, 2014, pp. 1–6. offloading strategies in cloud enabled cellular networks,” in Proc. IEEE
[54] M. A. Puente, Z. Becvar, M. Rohlik, F. Lobillo, and E. C. Strinati, Int. Conf. Commun. (ICC), London, U.K., 2015, pp. 5529–5534.
“A seamless integration of computationally-enhanced base sta- [79] W. Labidi, M. Sarkiss, and M. Kamoun, “Energy-optimal resource
tions into mobile networks towards 5G,” in Proc. IEEE Veh. scheduling and computation offloading in small cell networks,” in
Technol. Conf. Workshops (IEEE VTC Spring), Glasgow, U.K., 2015, Proc. Int. Conf. Telecommun. (ICT), Sydney, NSW, Australia, 2015,
pp. 1–5. pp. 313–318.

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 MACH AND BECVAR: MEC: SURVEY ON ARCHITECTURE AND COMPUTATION OFFLOADING 1655

[80] W. Labidi, M. Sarkiss, and M. Kamoun, “Joint multi-user resource [101] S. M. S. Tanzil, O. N. Gharehshiran, and V. Krishnamurthy, “Femto-
scheduling and computation offloading in small cell networks,” in Proc. cloud formation: A coalitional game-theoretic approach,” in Proc. IEEE
IEEE Int. Conf. Wireless Mobile Comput. Netw. Commun. (WiMob), Glob. Commun. Conf. (GLOBECOM), San Diego, CA, USA, 2015,
Abu Dhabi, UAE, 2015, pp. 794–801. pp. 1–6.
[81] S. Barbarossa, S. Sardellitti, and P. Di Lorenzo, “Joint allocation of [102] J. Oueis, E. Calvanese-Strinati, A. De Domenico, and S. Barbarossa,
computation and communication resources in multiuser mobile cloud “On the impact of backhaul network on distributed cloud computing,”
computing,” in Proc. IEEE Workshop Signal Process. Adv. Wireless in Proc. IEEE Wireless Commun. Netw. Conf. Workshops (WCNCW),
Commun. (SPAWC), Darmstadt, Germany, 2013, pp. 26–30. Istanbul, Turkey, 2014, pp. 12–17.
[82] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of [103] J. Oueis, E. C. Strinati, and S. Barbarossa, “Small cell clustering for
radio and computational resources for multicell mobile cloud com- efficient distributed cloud computing,” in Proc. IEEE Annu. Int. Symp.
puting,” in Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Pers. Indoor Mobile Radio Commun. (PIMRC), Washington, DC, USA,
Commun. (SPAWC), Toronto, ON, Canada, 2014, pp. 354–358. 2014, pp. 1474–1479.
[83] S. Sardellitti, S. Barbarossa, and G. Scutari, “Distributed mobile cloud [104] J. Oueis, E. C. Strinati, S. Sardellitti, and S. Barbarossa, “Small cell
computing: Joint optimization of radio and computational resources,” clustering for efficient distributed fog computing: A multi-user case,” in
in Proc. IEEE Globecom Workshops (GC Wkshps), Austin, TX, USA, Proc. IEEE Veh. Technol. Conf. (VTC Fall), Boston, MA, USA, 2015,
2014, pp. 1505–1510. pp. 1–5.
[84] K. Zhang et al., “Energy-efficient offloading for mobile edge computing [105] J. Oueis, E. C. Strinati, and S. Barbarossa, “The fog balancing: Load
in 5G heterogeneous networks,” IEEE Access, vol. 4, pp. 5896–5907, distribution for small cell cloud computing,” in Proc. IEEE 81st Veh.
2016. Technol. Conf. (VTC Spring), Glasgow, U.K., 2015, pp. 1–6.
[85] X. Chen, L. Jiao, W. Li, and X. Fu, “Efficient multi-user computation [106] M. Vondra and Z. Becvar, “QoS-ensuring distribution of computation
offloading for mobile-edge cloud computing,” IEEE/ACM Trans. Netw., load among cloud-enabled small cells,” in Proc. IEEE Int. Conf. Cloud
vol. 24, no. 5, pp. 2795–2808, Oct. 2016. Netw. (CloudNet), 2014, pp. 197–203.
[86] M.-H. Chen, B. Liang, and M. Dong, “A semidefinite relax- [107] S. Wang, M. Zafer, and K. K. Leung, “Online placement of
ation approach to mobile cloud offloading with computing access multi-component applications in edge computing environments,” IEEE
point,” in Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Access, vol. 5, pp. 2514–2533, 2017.
Commun. (SPAWC), Stockholm, Sweden, 2015, pp. 186–190. [108] P. Mach and Z. Becvar, “Cloud-aware power control for cloud-enabled
[87] M.-H. Chen, M. Dong, and B. Liang, “Joint offloading decision and small cells,” in Proc. IEEE Globecom Workshops (GC Wkshps), Austin,
resource allocation for mobile cloud with computing access point,” TX, USA, 2014, pp. 1038–1043.
in Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), [109] P. Mach and Z. Becvar, “Cloud-aware power control for real-time
Pudong, China, 2016, pp. 3516–3520. application offloading in mobile edge computing,” Trans. Emerg.
[88] S. Cao, X. Tao, Y. Hou, and Q. Cui, “An energy-optimal offloading Telecommun. Technol., vol. 27, no. 5, pp. 648–661, 2016.
algorithm of mobile computing based on HetNets,” in Proc. Int. Conf. [110] T. Taleb and A. Ksentini, “An analytical model for follow me cloud,” in
Connected Veh. Expo (ICCVE), Shenzhen, China, 2015, pp. 254–258. Proc. IEEE Glob. Commun. Conf. (GLOBECOM), Atlanta, GA, USA,
[89] J. Kennedy and R. C. Eberhart, “A discrete binary version of the par- 2013, pp. 1291–1296.
ticle swarm algorithm,” in Proc. IEEE Int. Conf. Syst. Man Cybern., [111] A. Ksentini, T. Taleb, and M. Chen, “A Markov decision process-based
Orlando, FL, USA, 1997, pp. 4104–4108. service migration procedure for follow me cloud,” in Proc. IEEE Int.
[90] Y. Zhao, S. Zhou, T. Zhao, and Z. Niu, “Energy-efficient task offloading Conf. Commun. (ICC), Sydney, NSW, Australia, 2014, pp. 1350–1354.
for multiuser mobile cloud computing,” in Proc. IEEE/CIC Int. Conf. [112] X. Sun and N. Ansari, “PRIMAL: PRofIt Maximization Avatar
Commun. China (ICCC), Shenzhen, China, 2015, pp. 1–5. pLacement for Mobile edge computing,” in Proc. IEEE Int. Conf.
[91] C. You and K. Huang, “Multiuser resource allocation for mobile- Commun. (ICC), Kuala Lumpur, Malaysia, 2016, pp. 1–6.
edge computation offloading,” in Proc. IEEE Glob. Commun. [113] S. Wang et al., “Mobility-induced service migration in mobile micro-
Conf. (GLOBECOM), Washington, DC, USA, 2016, pp. 1–6. clouds,” in Proc. IEEE Mil. Commun. Conf., Baltimore, MD, USA,
[92] C. You, K. Huang, H. Chae, and B.-H. Kim, “Energy-efficient 2014, pp. 835–840.
resource allocation for mobile-edge computation offloading,” [114] S. Wang et al., “Dynamic service migration in mobile edge-clouds,” in
IEEE Trans. Wireless Commun., vol. 16, no. 3, pp. 1397–1411, Proc. Netw. Conf. (IFIP Networking), Toulouse, France, 2015, pp. 1–9.
Mar. 2017. [115] A. Nadembega, A. S. Hafid, and R. Brisebois, “Mobility prediction
[93] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li, “Mobile-edge model-based service migration procedure for follow me cloud to
computing: Partial computation offloading using dynamic voltage support QoS and QoE,” in Proc. IEEE Int. Conf. Commun. (ICC),
scaling,” IEEE Trans. Commun., vol. 64, no. 10, pp. 4268–4282, Kuala Lumpur, Malaysia, 2016, pp. 1–6.
Oct. 2016. [116] S. Wang et al., “Dynamic service placement for mobile micro-clouds
[94] O. Muñoz, A. Pascual-Iserte, and J. Vidal, “Joint allocation of radio and with predicted future costs,” in Proc. IEEE Int. Conf. Commun. (ICC),
computational resources in wireless application offloading,” in Proc. London, U.K., 2015, pp. 5504–5510.
Future Netw. Mobile Summit, Lisbon, Portugal, 2013, pp. 1–10. [117] S. Wang et al., “Dynamic service placement for mobile micro-clouds
[95] O. Muñoz, A. Pascual-Iserte, and J. Vidal, “Optimization of radio and with predicted future costs,” IEEE Trans. Parallel Distrib. Syst., vol. 28,
computational resources for energy efficiency in latency-constrained no. 4, pp. 1002–1016, Apr. 2017.
application offloading,” IEEE Trans. Veh. Technol., vol. 64, no. 10, [118] R. Urgaonkar et al., “Dynamic service migration and workload schedul-
pp. 4738–4755, Oct. 2015. ing in edge-clouds,” Perform. Eval., vol. 91, pp. 205–228, Sep. 2015.
[96] O. Muñoz, A. Pascual-Iserte, J. Vidal, and M. Molina, “Energy-latency [119] K. Ha et al., “Adaptive VM handoff across cloudlets,” Dept. Comput.
trade-off for multiuser wireless computation offloading,” in Proc. IEEE Sci., Carnegie Mellon Univ., Pittsburgh, PA, USA, Tech. Rep.
Wireless Commun. Netw. Conf. Workshops (WCNCW), Istanbul, Turkey, CMU-CS-15-113, Jun. 2015.
2014, pp. 29–33. [120] S. Secci, P. Raad, and P. Gallard, “Linking virtual machine mobility
[97] Y. Mao, J. Zhang, S. H. Song, and K. B. Letaief, “Power-delay tradeoff to user mobility,” IEEE Trans. Netw. Service Manag., vol. 13, no. 4,
in multi-user mobile-edge computing systems,” in Proc. IEEE Glob. pp. 927–940, Dec. 2016.
Commun. Conf. (GLOBECOM), Washington, DC, USA, Dec. 2016, [121] D. Farinacci, V. Fuller, D. Meyer, and D. Lewis, “The locator/ID sepa-
pp. 1–6. ration protocol (LISP),” Internet Eng. Task Force, Fremont, CA, USA,
[98] T. Zhao, S. Zhou, X. Guo, Y. Zhao, and Z. Niu, “A cooperative schedul- RFC 6830, 2013.
ing scheme of local cloud and Internet cloud for delay-aware mobile [122] Z. Becvar, J. Plachy, and P. Mach, “Path selection using handover in
cloud computing,” in Proc. IEEE Globecom Workshops (GC Wkshps), mobile networks with cloud-enabled small cells,” in Proc. IEEE Int.
San Diego, CA, USA, 2015, pp. 1–6. Symp. Pers. Indoor Mobile Radio Commun. (PIMRC), Washington, DC,
[99] X. Guo, R. Singh, T. Zhao, and Z. Niu, “An index based task assign- USA, 2014, pp. 1480–1485.
ment policy for achieving optimal power-delay tradeoff in edge cloud [123] J. Plachy, Z. Becvar, and P. Mach, “Path selection enabling user mobil-
systems,” in Proc. IEEE Int. Conf. Commun. (ICC), Kuala Lumpur, ity and efficient distribution of data for computation at the edge of
Malaysia, 2016, pp. 1–7. mobile network,” Comput. Netw., vol. 108, pp. 357–370, Oct. 2016.
[100] V. Di Valerio and F. L. Presti, “Optimal virtual machines allocation [124] J. Plachy, Z. Becvar, and E. C. Strinati, “Dynamic resource allo-
in mobile femto-cloud computing: An MDP approach,” in Proc. IEEE cation exploiting mobility prediction in mobile edge computing,” in
Wireless Commun. Netw. Conf. Workshops (WCNCW), Istanbul, Turkey, Proc. IEEE Int. Symp. Pers. Indoor Mobile Radio Commun. (PIMRC),
2014, pp. 7–11. Valencia, Spain, 2016, pp. 1–6.

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.
 1656 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 3, THIRD QUARTER 2017

Pavel Mach (M’10) received the M.Sc. and Ph.D. Zdenek Becvar (M’10) received the M.Sc. and
degrees from the Czech Technical University in, Ph.D. degrees in telecommunication engineering
Prague, Czech Republic, in 2006 and 2010, respec- from the Czech Technical University, Prague,
tively, where he is currently a Senior Researcher Czech Republic, in 2005 and 2010, respectively,
with the Department of Telecommunication where he is currently an Associate Professor with
Engineering. He has been actively involved in the Department of Telecommunication Engineering.
several national and international projects founded From 2006 to 2007, he joined Sitronics Research and
by European Commission. In 2015, he has joined Development Center, Prague, focusing on speech
5G mobile research laboratory funded at Czech quality in VoIP. Furthermore, he was involved
Technical University focusing on key aspects and in research activities of Vodafone Research and
challenges related to future mobile networks and Development Center, Czech Technical University in
emerging wireless technologies. He has co-authored over 50 papers in Prague in 2009. He was on internships with Budapest Politechnic, Hungary in
international journals and conferences. His research interests include radio 2007, CEA-Leti, France, in 2013, and EURECOM, France, in 2016. In 2013,
resource management in emerging wireless technologies, device-to-device he became a representative of the Czech Technical University in ETSI and
communication, cognitive radio, and mobile edge computing. 3GPP standardization organizations. In 2015, he founded 5G Mobile Research
Laboratory, CTU, Prague, focusing on research toward 5G mobile networks
and beyond. He is a member of over 15 program committees at international
conferences or workshops and he has published three book chapters and over
60 conference or journal papers. He works on development of solutions for
future mobile networks (5G and beyond) with special focus on optimization
of radio resource management, mobility support, device-to-device communi-
cation, self-optimization, power control, architecture of radio access network,
and small cells.

Authorized licensed use limited to: Maulana Abul Kalam Azad Univ of Tech - Kolkata. Downloaded on May 21,2024 at 08:59:51 UTC from IEEE Xplore. Restrictions apply.