Mobile Edge Computing: A Survey on Architecture

and Computation Offloading

Pavel Mach, IEEE Member, Zdenek Becvar, IEEE Member

Abstract—Technological evolution of mobile user equipments a mobile operator and the Internet. The MCC brings several
(UEs), such as smartphones or laptops, goes hand-in-hand with advantages [2]; 1) extending battery lifetime by offloading
evolution of new mobile applications. However, running compu- energy consuming computations of the applications to the
tationally demanding applications at the UEs is constrained by
limited battery capacity and energy consumption of the UEs. cloud, 2) enabling sophisticated applications to the mobile
Suitable solution extending the battery life-time of the UEs users, and 3) providing higher data storage capabilities to the
is to offload the applications demanding huge processing to a users. Nevertheless, the MCC also imposes huge additional
conventional centralized cloud (CC). Nevertheless, this option load both on radio and backhaul of mobile networks and
introduces significant execution delay consisting in delivery of
arXiv:1702.05309v2 [cs.IT] 13 Mar 2017

introduces high latency since data is sent to powerful farm

the offloaded applications to the cloud and back plus time of
the computation at the cloud. Such delay is inconvenient and of servers that are, in terms of network topology, far away
make the offloading unsuitable for real-time applications. To from the users.
cope with the delay problem, a new emerging concept, known as To address the problem of a long latency, the cloud services
mobile edge computing (MEC), has been introduced. The MEC should be moved to a proximity of the UEs, i.e., to the
brings computation and storage resources to the edge of mobile edge of mobile network as considered in newly emerged
network enabling to run the highly demanding applications
at the UE while meeting strict delay requirements. The MEC edge computing paradigm. The edge computing can be un-
computing resources can be exploited also by operators and third derstood as a specific case of the MCC. Nevertheless, in
parties for specific purposes. In this paper, we first describe the conventional MCC, the cloud services are accessed via
major use cases and reference scenarios where the MEC is the Internet connection [3] while in the case of the edge
applicable. After that we survey existing concepts integrating computing, the computing/storage resources are supposed to
MEC functionalities to the mobile networks and discuss current
advancement in standardization of the MEC. The core of this be in proximity of the UEs (in sense of network topology).
survey is, then, focused on user-oriented use case in the MEC, Hence, the MEC can offer significantly lower latencies and
i.e., computation offloading. In this regard, we divide the research jitter when compared to the MCC. Moreover, while the MCC
on computation offloading to three key areas: i) decision on is fully centralized approach with farms of computers usually
computation offloading, ii) allocation of computing resource placed at one or few locations, the edge computing is supposed
within the MEC, and iii) mobility management. Finally, we
highlight lessons learned in area of the MEC and we discuss to be deployed in fully distributed manner. On the other hand,
open research challenges yet to be addressed in order to fully the edge computing provides only limited computational and
enjoy potentials offered by the MEC. storage resources with respect to the MCC. A high level
comparison of key technical aspects of the MCC and the edge
computing is outlined in Table I.
I. I NTRODUCTION
The first edge computing concept bringing the computa-
The users’ requirements on data rates and quality of service tion/storage closer to the UEs, proposed in 2009, is cloudlet
(QoS) are exponentially increasing. Moreover, technologi- [4]. The idea behind the cloudlet is to place computers with
cal evolution of smartphones, laptops and tablets enables high computation power at strategic locations in order to
to emerge new high demanding services and applications. provide both computation resources and storage for the UEs
Although new mobile devices are more and more powerful in vicinity. The cloudlet concept of the computing ”hotspots”
in terms of central processing unit (CPU), even these may not is similar to WiFi hotspots scenario, but instead of Internet
be able to handle the applications requiring huge processing connectivity the cloudlet enables cloud services to the mobile
in a short time. Moreover, high battery consumption still users. The fact that cloudlets are supposed to be mostly
poses a significant obstacle restricting the users to fully enjoy accessed by the mobile UEs through WiFi connection is seen
highly demanding applications on their own devices. This
motivates development of mobile cloud computing (MCC)
concept allowing cloud computing for mobile users [1]. In TABLE I: High level comparison of MCC and Edge comput-
the MCC, a user equipment (UE) may exploit computing ing concepts
and storage resources of powerful distant centralized clouds Technical aspect MCC Edge computing
(CC), which are accessible through a core network (CN) of Deployment Centralized Distributed
This work has been supported by the grant of Czech Technical University Distance to the UE High Low
in Prague No. SGS17/184/OHK3/3T/13. Latency High Low
The authors are with the Department of Telecommunication Engi- Jitter High Low
neering, Faculty of Electrical Engineering, Czech Technical University Computational power Ample Limited
in Prague, Prague, 166 27 Czech Republic (email:[email protected];
[email protected]). Storage capacity Ample Limited
as a disadvantage since the UEs have to switch between Telecommunications Standards Institute (ETSI) [24]. The so-
the mobile network and WiFi whenever the cloudlet services lution outlined by ETSI is known as Mobile Edge Com-
are exploited [2]. Moreover, QoS (Quality of Service) of the puting (MEC). The standardization efforts relating the MEC
mobile UEs is hard to fulfill similarly as in case of the MCC, are driven by prominent mobile operators (e.g., DOCOMO,
since the cloudlets are not an inherent part of the mobile Vodafone, TELECOM Italia) and manufactures (e.g., IBM,
network and coverage of WiFi is only local with limited Nokia, Huawei, Intel). The main purpose of ISG MEC group
support of mobility. is to enable an efficient and seamless integration of the cloud
The other option enabling cloud computing at the edge computing functionalities into the mobile network, and to help
is to perform computing directly at the UEs through ad- developing favorable conditions for all stakeholders (mobile
hoc cloud allowing several UEs in proximity to combine operators, service providers, vendors, and users).
their computation power and, thus, process high demanding Several surveys on cloud computing have been published
applications locally [5]-[14]. To facilitate the ad-hoc cloud, so far. In [3], the authors survey MCC application models
several critical challenges need to be addressed; 1) finding and highlight their advantages and shortcomings. In [25],
proper computing UEs in proximity while guaranteeing that a problem of a heterogeneity in the MCC is tackled. The
processed data will be delivered back to the source UE, 2) heterogeneity is understood as a variability of mobile de-
coordination among the computing UEs has to be enabled vices, different cloud vendors providing different services,
despite the fact that there are no control channels to facil- infrastructures, platforms, and various communication medium
itate reliable computing, 3) the computing UEs has to be and technologies. The paper identifies how this heterogeneity
motivated to provide their computing power to other devices impacts the MCC and discusses related challenges. The au-
given the battery consumption and additional data transmission thors in [26] survey existing efforts on Cloud Mobile Media,
constraints, 4) security and privacy issues. which provides rich multimedia services over the Internet and
A more general concept of the edge computing, when mobile wireless networks. All above-mentioned papers focus,
compared to cloudlets and ad-hoc clouds, is known as a fog in general, on the MCC where the cloud is not allocated
computing. The fog computing paradigm (shortly often ab- specifically at the edge of mobile network, but it is accessed
breviated as Fog in literature) has been introduced in 2012 by through the Internet. Due to a wide potential of the MEC,
Cisco to enable a processing of the applications on billions of there is a lot of effort both in industry and academia focusing
connected devices at the edge of network [15]. Consequently, on the MEC in particular. Despite this fact, there is just one
the fog computing may be considered as one of key enablers survey focusing primarily on the MEC [27] that, however,
of Internet of Things (IoT) and big data applications [16] as it only briefly surveys several research works dealing with the
offers: 1) low latency and location awareness due to proximity MEC and presents taxonomy of the MEC by describing
of the computing devices to the edge of the network, 2) wide- key attributes. Furthermore, the authors in [28] extensively
spread geographical distribution when compared to the CC; 3) surveys security issues for various edge computing concepts.
interconnection of very large number of nodes (e.g., wireless On top of that, the authors in [29] dedicate one chapter to the
sensors), and 4) support of streaming and real time applica- edge computing, where applications of economic and pricing
tions [15]. Moreover, the characteristics of the fog computing models are considered for resource management in the edge
can be exploited in many other applications and scenarios such computing.
as smart grids, connected vehicles for Intelligent Transport In contrast to the above-mentioned surveys, we describe
Systems (ITS) or wireless sensor networks [17]-[20]. key use cases and scenarios for the MEC (Section II). Then,
From the mobile users’ point of view, the most notable we survey existing MEC concepts proposed in the literature
drawback of all above-mentioned edge computing concepts integrating the MEC functionalities into the mobile networks
is that QoS and QoE (Quality of Experience) for users can and we discuss standardization of the MEC (Section III).
be hardly guaranteed, since the computing is not integrated After that, the core part of the paper is focused on technical
into an architecture of the mobile network. One concept works dealing with computation offloading to the MEC. On
integrating the cloud capabilities into the mobile network is one hand, the computation offloading can be seen as a key
Cloud Radio Access Network (C-RAN) [21]. The C-RAN use case from the user perspective as it enables running
exploits the idea of distributed protocol stack [22], where some new sophisticated applications at the UE while reducing its
layers of the protocol stack are moved from distributed Radio energy consumption (see, e.g., [30]-[36] where computation
Remote Heads (RRHs) to centralized baseband units (BBUs). offloading to distant CC is assumed). On the other hand,
The BBU’s computation power is, then, pooled together into the computation offloading brings several challenges, such
virtualized resources that are able to serve tens, hundreds or as selection of proper application and programming models,
even thousands of RRHs. Although the computation power accurate estimation of energy consumption, efficient manage-
of this virtualized BBU pool is exploited primarily for a ment of simultaneous offloading by multiple users, or virtual
centralized control and baseband processing it may also be machine (VM) migration [37]. In this respect, we overview
used for the computation offloading to the edge of the network several general principles related to the computation offload-
(see, for example, [23]). ing, such as offloading classification (full, partial offloading),
Another concept integrating the edge computing into the factors influencing the offloading itself, and management of
mobile network architecture is developed by newly created the offloading in practice (Section IV). Afterwards, we sort the
(2014) industry specification group (ISG) within European efforts within research community addressing following key
 Fig. 1: Example of use cases and scenarios for the MEC.

challenges regarding computation offloading into the MEC: A. Consumer-oriented services

• A decision on the computation offloading to the MEC The first use case category is consumer-oriented and, hence,
with the purpose to determine whether the offloading is should be beneficial directly to the end-users. In general, the
profitable for the UE in terms of energy consumption users profit from the MEC mainly by means of the computa-
and/or execution delay (Section V). tion offloading, which enables running new emerging applica-
• An efficient allocation of the computing resources tions at the UEs. One of the applications benefiting from the
within the MEC if the computation is offloaded in order to computation offloading is a web accelerated browser, where
minimize execution delay and balance load of both com- most of the browsing functions (web contents evaluation,
puting resources and communication links (Section VI). optimized transmission, etc.) are offloaded to the MEC; see
• Mobility management for the applications offloaded experimental results on offloading of web accelerated browser
to the MEC guaranteeing service continuity if the UEs to the MEC in [40]. Moreover, face/speech recognition or
exploiting the MEC roams throughout the network (Sec- image/video editing are also suitable for the MEC as these
tion VII). require large amount of computation and storage [41].
Moreover, we summarize the lessons learned from state of the Besides, the computation offloading to the MEC can be
art focused on computation offloading to the MEC (Section exploited by the applications based on augmented, assisted or
VIII) and outline several open challenges, which need to be virtual reality. These applications derive additional information
addressed to make the MEC beneficial for all stakeholders about users’ neighborhood by performing an analysis of their
(Section IX). Finally, we summarize general outcomes and surroundings (e.g., visiting tourists may find points of interest
draw conclusions (Section X). in his/her proximity). This may require fast responses, and/or
significant amount of computing resources not available at
the UE. An applicability of the MEC for augmented reality
II. U SE CASES AND SERVICE SCENARIOS
is shown in [42]. The authors demonstrate on a real MEC
The MEC brings many advantages to all stakeholders, such testbed that the reduction of latency up to 88% and energy
as mobile operators, service providers or users. As suggested consumption of the UE up to 93% can be accomplished by
in [38][39], three main use case categories, depending on the the computation offloading to the MEC.
subject to which they are profitable to, can be distinguished for On top of that, the users running low latency applications,
the MEC (see Fig. 1). The next subsections discuss individual such as online gaming or remote desktop, may profit from
use case categories and pinpoint several key service scenarios the MEC in proximity. In this case a new instance of a
and applications. specific application is initiated at an appropriate mobile edge
host to reduce the latency and resources requirements of the scheduling. In addition, the MEC can also be used for mo-
application at the UE. bile video delivery optimization using throughput guidance
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.
The second use case category is represented by the services To deal with this problem, the analytic MEC application can
from which operators and third parties can benefit. An example provide a real-time indication on an estimated throughput to
of the use case profitable for the operator or third party is a a backend video server in order to match the application-level
gathering of a huge amount of data from the users or sensors. coding to the estimated throughput.
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 STANDARDIZATION
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 (Inter-
literature that enable to bring computation close to the UEs.
net of Thing) purposes [43]-[45]. Basically, IoT devices are
Secondly, we describe the effort done within ETSI standard-
connected 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 gateway,
which purpose is to aggregate and deliver IoT services into
highly distributed mobile base stations in order to enable A. Overview of the MEC concept
applications responding in real time. In recent years, several MEC concepts with purpose to
The MEC can be also exploited for ITS to extend the smoothly integrate cloud capabilities into the mobile network
connected car cloud into the mobile network. Hence, road- architecture have been proposed in the literature. This section
side applications running directly at the MEC can receive briefly introduces fundamental principles of small cell cloud
local messages directly from applications in the vehicles and (SCC), mobile micro cloud (MMC), fast moving personal
roadside sensors, analyze them and broadcast warnings (e.g., cloud, follow me cloud (FMC), and CONCERT. Moreover,
an accident) to nearby vehicles with very low latency. The the section shows enhancements/modifications to the network
exploitation of the MEC for car-to-car and car-to-infrastructure architecture necessary for implementation of each MEC con-
communications was demonstrated by Nokia and its partners cept.
in an operator’s LTE network just recently in 2016 [46][47]. 1) Small cell cloud (SCC): The basic idea of the SCC,
firstly introduced in 2012 by the European project TROPIC
[48][53], is to enhance small cells (SCeNBs), like microcells,
C. Network performance and QoE improvement services
picocells or femtocells, by an additional computation and
The third category of use cases are those optimizing network storage capabilities. The similar idea is later on addressed in
performance and/or improving QoE. One such use case is to SESAME project as well, where the cloud-enabled SCeNBs
enable coordination between radio and backhaul networks. So supports the edge computing [49][50]. The cloud-enhanced
far, if the capacity of either backhaul or radio link is degraded, SCeNBs can pool their computation power exploiting network
the overall network performance is negatively influenced as function virtualization (NFV) [51][52] paradigm. Because a
well, since the other part of the network (either radio or high number of the SCeNBs is supposed to be deployed in
backhaul, respectively) is not aware of the degradation. In future mobile networks, the SCC can provide enough compu-
this respect, an analytic application exploiting the MEC can tation power for the UEs, especially for services/applications
provide real-time information on traffic requirements of the having stringent requirements on latency (the examples of such
radio/backhaul network. Then, an optimization application, applications are listed in Section II-A).
running on the MEC, reshapes the traffic per application or In order to fully and smoothly integrate the SCC concept
re-routes traffic as required. into the mobile network architecture, a new entity, denoted as
Another way to improve performance of the network is to a small cell manager (SCM), is introduced to control the SCC
alleviate congested backhaul links by local content caching at [53]. The SCM is in charge of the management of the comput-
the mobile edge. This way, the MEC application can store ing and/or storage resources provided by the SCeNBs. Since
the most popular content used in its geographical area. If the SCeNBs can be switched on/off at any time (especially
the content is requested by the users, it does not have to be if owned by the users as in case of the femtocells), the SCM
transfered over the backhaul network. performs dynamic and elastic management of the computation
Besides alleviation and optimization of the backhaul net- resources within the SCC. The SCM is aware of the overall
work, the MEC can also help in radio network optimization. cluster context (both radio and cloud-wise) and decides where
For example, gathering related information from the UEs and to deploy a new computation or when to migrate an on-going
processing these at the edge will result in more efficient computation to optimize the service delivery for the end-user.
Fig. 2: SCC architecture (MME - Mobility Management Entity, HSS - Home Subscriber Server, S-GW - Serving Gateway,
P-GW - Packet Gateway).

The computing resources are virtualized by means of Virtual fully distributed in a similar way as the VL-SCM solution for
Machine (VM) located at the SCeNBs. An important aspect the SCC. To this end, the MMCs are interconnected directly
regarding the architecture of the SCC is deployment of the or through backhaul in order to guarantee service continuity
SCM (see Fig. 2). The SCM may be deployed in a centralized if the UEs move within the network to enable smooth VM
manner either as a standalone SCM located within the RAN, migration among the MMCs (see more detail on VM migration
close to a cluster of the SCeNBs, or as an extension to a in Section VII-B).
MME [53][54]. Moreover, the SCM can be deployed also in a 3) Fast moving personal cloud (MobiScud): The MobiScud
distributed hierarchical manner, where a local SCM (L-SCM) architecture [57] integrates the cloud services into the mobile
or a virtual L-SCM (VL-SCM) manages the computing and networks by means of software defined network (SDN) [58]
storage resources of the SCeNBs’ clusters in vicinity while a and NFV technologies whilst maintaining backward compati-
remote SCM (R-SCM), located in the CN, has resources of all bility with existing mobile network. When compared to the
SCeNBs connected to the CN at its disposal [55] (see Fig. 2b). SCC and the MMC concepts, the cloud resources in the
2) Mobile micro clouds (MMC): The concept of the MMC MobiScud are not located directly at the access nodes such as
has been firstly introduced in [56]. Like the SCC, also the SCeNB or eNB, but at operator’s clouds located within RAN
MMC allows users to have instantaneous access to the cloud or close to RAN (see Fig. 4). Still, these clouds are assumed
services with a low latency. While in the SCC the computa- to be highly distributed similarly as in case of the SCC and
tion/storage resources are provided by interworking cluster(s) the MMC enabling the cloud service to all UEs in vicinity.
of the SCeNBs, the UEs exploit the computation resources Analogously to the SCC, the MobiScud introduces a new
of a single MMC, which is typically connected directly to a control entity, a MobiScud control (MC), which interfaces
wireless base station (i.e., the eNB in the mobile network) as with the mobile network, SDN switches and the cloud of
indicated in Fig. 3. The MMC concept does not introduce any the operator. Basically, the MC has two functionalities: 1)
control entity into the network and the control is assumed to be monitoring control plane signaling message exchange between

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

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

B. ETSI MEC
Besides all above-mentioned solutions, also ETSI is cur-
rently deeply involved in standardization activities in order to
integrate the MEC into the mobile networks. In this regard,
we briefly summarize the standardization efforts on the MEC
within ETSI, describe reference architecture according to
ETSI, and contemplate various options for the MEC deploy-
ment that are considered so far.
1) Standardization of ETSI MEC: Standardization of the
MEC is still in its infancy, but drafts of specifications have
Fig. 5: The network architecture enabling FMC concept (cen- already been released by ISG MEC. The terminology used in
tralized solution). individual specifications relating to conceptual, architectural
and functional elements is described in [63]. The main purpose the operation support system (OSS) of the mobile operator.
of this document is to ensure the same terminology is used Then, the OSS decides if requests are granted or not. The
by all ETSI specifications related to the MEC. A framework granted requests are forwarded to a mobile edge orchestrator.
exploited by ISG MEC for coordination and promotion of The mobile edge orchestrator is the core functionality in
MEC is defined in proof of concept (PoC) specification [64]. the MEC system level management as it maintains overall
The basic objectives of this document is to describe the PoC view on available computing/storage/network resources and
activity process in order to promote the MEC, illustrate key the MEC services. In this respect, the mobile edge orchestrator
aspects of the MEC and build a confidence in viability of allocates the virtualized MEC resources to the applications
the MEC technology. Further, several service scenarios that that are about to be initiated depending on the applications
should benefit from the MEC and proximity of the cloud requirements (e.g., latency). Furthermore, the orchestrator also
services is presented in [65] (see Section II for more detail). flexibly scales down/up available resources to already running
Moreover, technical requirements on the MEC to guarantee applications.
interoperability and to promote MEC deployment are intro- The MEC system level management is interconnected with
duced in [38]. The technical requirements are divided into a MEC server level management constituting a mobile edge
generic requirements, service requirements, requirements on platform and a virtualization platform manager. The former
operation and management, and finally security, regulations one manages the life cycle of the applications, application
and charging requirements. rules and service authorization, traffic rules, etc. The latter
2) ETSI MEC reference architecture: The reference archi- one is responsible for allocation, management and release of
tecture, described by ETSI in [66], is composed of functional the virtualized computation/storage resources provided by the
elements and reference points allowing interaction among virtualization infrastructure located within the MEC server.
them (see Fig. 7). Basically, the functional blocks may not The MEC server is an integral part of the reference architecture
necessarily represent physical nodes in the mobile network, but as it represents the virtualized resources and hosts the MEC
rather software entities running on the top of a virtualization applications running as the VMs on top of the virtualization
infrastructure. The virtualization infrastructure is understood infrastructure.
as a physical data center on which the VMs are run and the 3) Deployment options of ETSI MEC: As already men-
VMs represent individual functional elements. In this respect, tioned in the previous subsection, the MEC services will be
it is assumed that some architectural features from ETSI NFV provided by the MEC servers, which have the computation and
group, which runs in parallel to ETSI MEC, will be reused for storage resources at their disposal. There are several options
the MEC reference architecture as well, since the basic idea where the MEC servers can be deployed within the mobile
of NFV is to virtualize all network node functions. network. The first option is to deploy the MEC server directly
As shown in Fig. 7, the MEC can be exploited either by at the base station similarly as in case of the SCC or the MCC
a UE application located directly in the UE, or by third (see Section III-A1 and Section III-A2). Note that in case of
party customers (such as commercial enterprise) via customer a legacy network deployment, such as 3G networks, the MEC
facing service (CFS) portal. Both the UE and the CFS portal servers may be deployed at 3G Radio Network Controllers
interact with the MEC system through a MEC system level as well [38]. The second option is to place the MEC servers
management. The MEC system level management includes a at cell aggregation sites or at multi-RAT aggregation points
user application lifecycle management (LCM) proxy, which that can be located either within an enterprise scenario (e.g.,
mediate the requests, such as initiation, termination or relo- company) or a public coverage scenario (e.g., shopping mall,
cations of the UE’s application within the MEC system to stadium, airport, etc.). The third option is to move the MEC
server farther from the UEs and locate it at the edge of CN
analogously to the FMC (Section III-A4).
Of course, selection of the MEC server deployment depends
on many factors, such as, scalability, physical deployment
constraints and/or performance criteria (e.g., delay). For ex-
ample, the first option with fully distributed MEC servers
deployment will result in very low latencies since the UEs
are in proximity of the eNB and, hence, in proximity of the
MEC server. Contrary, the UEs exploiting the MEC server
located in the CN will inevitably experience longer latencies
that could prevent a use of real-time applications. An initial
study determining where to optimally install the MEC servers
within the mobile network with the primary objective to find
a trade-off between installation costs and QoS measured in
terms of latency is presented in [67] and further elaborated
in [68]. Based on these studies, it is expected that, similarly
as in CONCERT framework (see Section III-A5), the MEC
servers with various computation power/storage capacities will
Fig. 7: MEC reference architecture [66]. be scattered throughout the network. Hence, the UEs requiring
 TABLE II: Comparison of existing MEC concepts.
MEC con- Control en- Control manner Control placement Computation/storage placement
cept tity
SCC SCM Centralized, decentralized hierar- In RAN (e.g., at eNB) or in CN SCeNB, cluster of SCeNBs
chical (depending on SCM type (e.g., SCM collocated with MME)
and placement)
MMC - Decentralized MMC (eNB) eNB
MobiScud MC Decentralized Between RAN and CN Distributed cloud within RAN or
close to RAN
FMC FMCC Centralized, decentralized Collocated with existing node (e.g., DC close or collocating with dis-
hierarchical (option with node in CN) or run as software on tributed CN
hierarchical FMCC), decentralized DC
(option without FMC controller)
CONCERT Conductor Centralized, decentralized hierar- N/A (it could be done in the same eNB (RIE), regional and central
chical manner as in FMC concept) servers
ETSI MEC Mobile edge Centralized N/A (the most feasible option is to eNB, aggregation point, edge of
orchestrator place control into CN CN

only a low computation power will be served by the local IV. I NTRODUCTION TO COMPUTATION OFFLOADING
MEC servers collocated directly with the eNB, while highly From the user perspective, a critical use case regarding the
demanding applications will be relegated to more powerful MEC is a computation offloading as this can save energy
MEC servers farther from the UEs. and/or speed up the process of computation. In general, a
crucial part regarding computation offloading is to decide
C. Summary whether to offload or not. In the former case, also a question
This section mutually compares the MEC concepts proposed is how much and what should be offloaded [41]. Basically, a
in literature with the vision of the MEC developed under ETSI. decision on computation offloading may result in:
There are two common trends followed by individual MEC • Local execution - The whole computation is done locally
solutions that bring cloud to the edge of mobile network. at the UE (see Fig. 8). The offloading to the MEC is not
The first trend is based on virtualization techniques exploiting performed, for example, due to unavailability of the MEC
NFVs principles. The network virtualization is a necessity in computation resources or if the offloading simply does not
order to flexibly manage virtualized resources provided by the pay off.
MEC. The second trend is a decoupling the control and data • Full offloading - The whole computation is offloaded and
planes by taking advantage of SDN paradigm, which allows a processed by the MEC.
dynamic adaptation of the network to changing traffic patterns • Partial offloading - A part of the computation is processed
and users requirements. The use of SDN for the MEC is also in locally while the rest is offloaded to the MEC.
line with current trends in mobile networks [69]-[71]. Regard- The computation offloading, and partial offloading in par-
ing control/signaling, the MMC and MobiScud assume fully ticular, is a very complex process affected by different factors,
decentralize approach while the SCC, FMC, and CONCERT such as users preferences, radio and backhaul connection
adopt either fully centralized control or hierarchical control quality, UE capabilities, or cloud capabilities and availability
for better scalability and flexibility. [3]. An important aspect in the computation offloading is also
If we compare individual MEC concepts in terms of com- an application model/type since it determines whether full or
putation/storage resources deployment, the obvious effort is partial offloading is applicable, what could be offloaded, and
to fully distribute these resources within the network. Still, how. In this regard, we can classify the applications according
each MEC concept differs in the location, where the compu- to several criteria:
tation/storage resources are physically located. While the SCC,
MMC and MobiScud assume to place the computation close to
the UEs within RAN, the FMC solution considers integration
of the DCs farther away, for example, in a distributed CN.
On top of that, CONCERT distributes the computation/storage
resources throughout the network in a hierarchical manner so
that a low demanding computation application are handled
locally and high demanding applications are relegated either
to regional or central servers. Concerning ETSI MEC, there
are also many options where to place MEC servers offering
computation/storage resources to the UEs. The most probable
course of action is that the MEC servers will be deployed
everywhere in the network to guarantee high scalability of the
computation/storage resources. The comparison of all existing
MEC concepts is shown in Table II. Fig. 8: Possible outcomes of computation offloading decision.
 • Knowledge on the amount of data to be processed - The
applications can be classified according to the knowledge
on the amount of data to be processed. For the first type
of the applications (represented, e.g., by face detection,
virus scan, etc.,) the amount of data to be processed is
known beforehand. For the second type of the applications,
it is not possible to estimate the amount of data to be
processed as these are continuous-execution application and
there is no way to predict how long they will be running
(such as, online interactive games) [95]. It is obvious that
decision on computation offloading could be quite tricky
for continuous-execution application.
• Dependency of the offloadable parts - The last criterion
for classification of application to be offloaded is a mutual
Fig. 9: An example of partial offloading for application dependency of individual parts to be processed. The parts
without non-offloadable part(s) (a) and application with non- of the application can be either independent on each other
offloadable part (b). or mutually dependent. In the former case, all parts can
be offloaded simultaneously and processed in parallel. In
the latter case, however, the application is composed of
• Offloadability of application - The application enabling parts (components) that need input from some others and
code or data partitioning and parallelization (i.e., applica- parallel offloading may not be applicable. Note that the
tion that may be partially offloaded) can be categorized relationship among individual components can be expressed
into two types. The first type of the applications is the app, by component dependency graph (CDG) or call graph
which can be divided into N offloadable parts that all can (CG) (see, e.g., [34][41][72][73]). The relationship among
be offloaded (see Fig. 9a). Since each offloadable part may the components is illustrated in Fig. 10, where the whole
differ in the amount of data and required computation, it application is divided into M non-offloadable parts (1st , 4th ,
is necessary to decide which parts should be offloaded to and 6th part in Fig. 10) and N offloadable parts (2nd , 3rd ,
the MEC. In the example given in Fig. 9a, 1st , 2nd , 3nd , and 5th part in Fig. 10). In the given example, 2nd and 3rd
6th , and 9th parts are processed locally while the rest is part can be offloaded only after execution of the 1st part
offloaded to the MEC. Notice that in the extreme case, this while the 5th part can be offloaded after execution of the
type of application may be fully offloaded to the MEC if 1st - 4th parts.
no parts are processed by the UE. The second type of the The other important aspect regarding computation offload-
applications is always composed of some non-offloadable ing is how to utilize and manage offloading process in practice.
part(s) that cannot be offloaded (e.g., user input, camera, or Basically, the UE needs to be composed of a code profiler,
acquire position that needs to be executed at the UE [72]) system profiler, and decision engine [36]. The code profiler’s
and M offloadable parts. In Fig. 9b, the UE processes the responsibility is to determine what could be offloaded (depend-
whole non-offloadable part together with 2nd , 6th , and 7th ing on application type and code/data partitioned as explained
parts while the rest of the application is offloaded to the above). Then, the system profiler is in charge of monitoring
MEC. various parameters, such as available bandwidth, data size to
be offloaded or energy spent by execution of the applications
locally. Finally, decision engine determines whether to offload
or not.
The next sections survey current research works focus-
ing on following pivotal research topics: 1) decision on the
computation offloading to the MEC, 2) efficient allocation of
the computation resources within the MEC, and 3) mobility
management for the moving users exploiting MEC services.
Note that from now on we use explicitly the terminology
according to ETSI standardization activities. Consequently, we
use term MEC server as a node providing computing/storage
resources to the UEs instead of DC, MMC, etc.

V. D ECISION ON COMPUTATION OFFLOADING TO MEC

This section surveys current research related to the decision
on the computation offloading to the MEC. The papers are
divided into those considering either only the full offloading
(Section V-A) or those taking into account also possibility of
Fig. 10: Dependency of offloadable components [72]. the partial offloading (Section V-B).
 (see Fig. 12). This module decides, during each time slot,
whether the application waiting in a buffer should be processed
locally or at the MEC while minimizing the execution delay.
The performance of the proposed algorithm is compared to
the local execution policy (computation done always locally),
cloud execution policy (computation performed always by the
MEC server), and greedy offloading policy (UE schedules
Fig. 11: The example of offloading decision aiming minimiza- data waiting in the buffer whenever the local CPU or the
tion of execution delay. transmission unit is idle). The simulation results show that the
proposed optimal policy is able to reduce execution delay by
up to 80% (compared to local execution policy) and roughly
A. Full offloading up to 44% (compared to cloud execution policy) as it is able to
The main objective of the works focused on the full offload- cope with high density of applications’ arrival. The drawback
ing decision is to minimize an execution delay (Section V-A1), of the proposed method is that the UE requires feedback from
to minimize energy consumption at the UE while predefined the MEC server in order to make the offloading decision, but
delay constraint is satisfied (Section V-A2), or to find a the generated signaling overhead is not discussed in the paper.
proper trade-off between both the energy consumption and Another idea aiming at minimization of the execution de-
the execution delay (Section V-A3). lay is introduced in [75]. When compared to the previous
1) Minimization of execution delay: One of the advantages study, the authors in [75] also reduce application failure
introduced by the computation offloading to the MEC is a for the offloaded applications. The paper considers the UE
possibility to reduce the execution delay (D). In case the applies dynamic voltage and frequency scaling (DVS) [76]
UE performs all computation by itself (i.e., no offloading and energy harvesting techniques [77] to minimize the energy
is performed), the execution delay (Dl ) represents solely consumption during the local execution and a power control
the time spent by the local execution at the UE. In case optimizing data transmission for the computation offload-
of the computation offloading to the MEC, the execution ing. In this respect, the authors propose a low-complexity
delay (Do ) incorporates three following parts: 1) transmission Lyapunov optimization-based dynamic computation offloading
duration of the offloaded data to the MEC (Dot ), 2) compu- (LODCO) algorithm. The LODCO makes offloading decision
tation/processing time at the MEC (Dop ), and 3) time spent in each time slot and subsequently allocates CPU cycles for
by reception of the processed data from the MEC (Dor ). The the UE (if the local execution is performed) or allocates trans-
simple example of the computation offloading decision based mission power (if the computation offloading is performed).
solely on the execution delay is shown in Fig. 11. It could be The proposed LODCO is able to reduce execution time by up
observed that the UE1 performs all computation locally since to 64% by offloading to the MEC. Furthermore, the proposal
the local execution delay is significantly lower than expected is able to completely prevent a situation when offloaded
execution delay for the computation offloading to the MEC application would be dropped.
(i.e., Dl < Do ). Contrary, a better alternative for the UE2 is The drawback of both above-mentioned papers is that the
to fully offload data to the MEC as the local execution would offloading decision does not take into account energy con-
result in notable higher execution delay (i.e., Dl > Do ). sumption at the side of UE as fast battery depletion impose
The goal to minimize execution delay is pursued by the au- significant obstacle in contemporary networks. In [75], the
thors in [74]. This is accomplished by one-dimensional search energy aspect of the UE is omitted in the decision process
algorithm, which finds an optimal offloading decision policy since the paper assumes that the UEs exploit energy harvesting
according to the application buffer queuing state, available techniques. The harvesting technique, however, is not able to
processing powers at the UE and at the MEC server, and fully address energy consumption problem by itself.
characteristic of the channel between the UE and the MEC 2) Minimization of energy consumption while satisfying
server. The computation offloading decision itself is done at execution delay constraint: The main objective of the papers
the UE by means of a computation offloading policy module surveyed in this section is to minimize the energy con-
sumption at the UE while the execution delay constraint of
the application is satisfied. On one hand, the computation
offloaded to the MEC saves battery power of the UE since
the computation does not have to be done locally. On the
other hand, the UE spends certain amount of energy in order
to: 1) transmit offloaded data for computation to the MEC
(Eot ) and 2) receive results of the computation from the MEC
(Eor ). The simple example of the computation offloading
decision primarily based on the energy consumption is shown
in Fig. 13. In the given example, the UE1 decides to perform
the computation locally since the energy spent by the local
Fig. 12: Computation offloading considered in [74] (CSI stands execution (El ) is significantly lower than the energy required
for channel state information). for transmission/reception of the offloaded data (E0 ). Contrary,
Fig. 13: The example of computation offloading decision based on energy consumption while satisfying execution delay
constraint.

the UE2 offloads data to the MEC as the energy required by the satisfying the maximum allowed execution delay is proposed
computation offloading is significantly lower than the energy in [81]. A decision on the computation offloading is done
spent by the local computation. Although the overall execution periodically in each time slot, during which all the UEs are
delay would be lower if the UE1 offloads computation to divided into two groups. While the UEs in the first group
the MEC and also if the UE2 performs the local execution, are allowed to offload computation to the MEC, the UEs in
the delay is still below maximum allowed execution delay the second group have to perform computation locally due
constraint (i.e., Dl < Dmax ). Note that if only the execution to unavailable computation resources at the MEC (note that in
delay would be considered for the offloading decision (as the paper, the computation is done at the serving SCeNB). The
considered in Section V-A3), both UEs would unnecessarily UEs are sorted to the groups according to the length of queue,
spent more energy. that is, according to the amount of data they need to process.
The computation offloading decision minimizing the energy After the UEs are admitted to offload the computation, joint
consumption at the UE while satisfying the execution delay allocation of the communication and computation resources
of the application is proposed in [78]. The optimization prob- is performed by finding optimal transmission power of the
lem is formulated as a constrained Markov decision process UEs and allocation of the SCeNB’s computing resources to all
(CMDP). To solve the optimization problem, two resource individual UEs. The performance of the proposal is evaluated
allocation strategies are introduced. The first strategy is based in terms of an average queue length depending on intensity of
on an online learning, where the network adapts dynamically data arrival and a number of antennas used at the UEs and the
with respect to the application running at the UE. The second SCeNB. It is shown that the more antennas is used, the less
strategy is pre-calculated offline strategy, which takes advan- transmission power at the UEs is needed while still ensuring
tage of a certain level of knowledge regarding the application the delay constraint of the offloaded computation.
(such as arrival rates measured in packets per slot, radio The main weak point of [81] is that it assumes only a single
channel condition, etc.). The numerical experiments show that SCeNB and, consequently, there is no interference among the
the pre-calculated offline strategy is able to outperform the UEs connected to various SCeNBs. Hence, the work in [81]
online strategy by up to 50% for low and medium arrival is extended in [82] to the multi-cell scenario with N SCeNBs
rates (loads). Since the offline resource allocation strategy to reflect the real network deployment. Since the formulated
proposed in [78] shows its merit, the authors devise two optimization problem in [81] is no longer convex, the authors
addition dynamic offline strategies for the offloading [79]: propose a distributed iterative algorithm exploiting Successive
deterministic offline strategy and randomized offline strategy. Convex Approximation (SCA) converging to a local optimal
It is demonstrated that both offloading offline strategies can solution. The numerical results demonstrate that the proposed
lead to significant energy savings comparing to the case when joint optimization of radio and computational resources signifi-
the computing is done solely at the UE (energy savings up to cantly outperforms methods optimizing radio and computation
78%) or solely at the MEC (up to 15%). separately. Moreover, it is shown that the applications with
A further extension of [79] from a single-UE to a multi-UEs fewer amount of data to be offloaded and, at the same time,
scenario is considered in [80]. The main objective is to jointly requiring high number of CPU cycles for processing are more
optimize scheduling and computation offloading strategy for suitable for the computation offloading. The reason is that the
each UE in order to guarantee QoE, fairness between the energy spent by the transmission/reception of the offloaded
UEs, low energy consumption, and average queuing/delay data to the MEC is significantly lower than the energy savings
constraints. The UEs that are not allowed to offload the at the UE due to the computation offloading. The work in
computation make either local computation or stay idle. It [82] is further extended in [83] by a consideration of multi-
is shown the offline strategy notably outperforms the online clouds that are associated to individual SCeNBs. The results
strategies in terms of the energy saving (by roughly 50%). show that with an increasing number of the SCeNBs (i.e., with
In addition, the energy consumed by individual UEs strongly increasing number of clouds), the energy consumption of the
depends on requirements of other UEs application. UE proportionally decreases.
Another offloading decision strategy for the multi-UEs The same goal as in previous paper is achieved in [84] by
case minimizing the energy consumption at the UEs while means of an energy-efficient computation offloading (EECO)
algorithm. The EECO is divided into three stages. In the to satisfy the request or if the computation should be farther
first stage, the UEs are classified according to their time and relayed to the CC. The problem is formulated as a non-convex
energy cost features of the computation to: 1) the UEs that quadratically constrained quadratic program (QCQP), which
should offload the computation to the MEC as the UEs cannot is, however, NP-hard. Hence, a heuristic algorithm based on
satisfy the execution latency constraint, 2) the UEs that should a semi-definite relaxation together with a novel randomization
compute locally as they are able to process it by itself while method is proposed. The proposed heuristic algorithm is able
the energy consumption is below a predefined threshold, and to significantly lower a total system cost (i.e., weighted sum of
3) the UEs that may offload the computation or not. In the total energy consumption, execution delay and costs to offload
second stage, the offloading priority is given to the UEs from and process all applications) when compared to the situation
the first and the third set determined by their communication if the computation is done always solely at the UE (roughly
channels and the computation requirements. In the third stage, up to 70%) or always at the MEC/CC (approximately up to
the eNBs/SCeNBs allocates radio resources to the UEs with 58%).
respect to given priorities. The computational complexity of The extension of [86] from the single-UE to the multi-
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 are assumed to be connected to the same computing node
represents the number of available channels. According to (e.g., eNB), the offloading decision is done jointly with the
presented numerical results, the EECO is able to decrease allocation of computing and communication resources to all
the energy consumption by up to 15% when compared to the UEs. Analogously to [86], the proposal in [87] outperforms
computation without offloading. Further, it is proofed that with the case when computation is done always by the UE (system
increasing computational capabilities of the MEC, the number cost decreased by up to 45%) and strategy if computation is
of UEs 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- more realistic scenario with multiple computing eNBs, where
user multi-channel environment considering a trade-off be- interference among the UEs attached to different eNBs would
tween the energy consumption at the UE and the execution play an important role in the offloading decision. Moreover,
delay is proposed in [85]. Whether the offloading decision the overall complexity of the proposed solution is O(N 6 ) per
prefers to minimize energy consumption or execution delay one iteration, which could be too high for a high number of
is determined by a weighing parameter. The main objective UEs (N ) connected to the eNB.
of the paper is twofold; 1) choose if the UEs should perform
the offloading to the MEC or not depending on the weighing
B. Partial offloading
parameter and 2) in case of the computation offloading, select
the most appropriate wireless channel to be used for data This subsection focuses on the works dealing with the
transmission. To this end, the authors present an optimal partial offloading. We classify the research on works focused
centralized solution that is, however, NP-hard in the multi- on minimization of the energy consumption at the UE while
user multi-channel environment. Consequently, the authors predefined delay constraint is satisfied (Section V-B1) and
also propose a distributed computation offloading algorithm works finding a proper trade-off between both the energy
achieving Nash equilibrium. Both the optimal centralized consumption and the execution delay (Section V-B2).
solution and the distributed algorithm are compared in terms 1) Minimization of energy consumption while satisfying
of two performance metrics; 1) the amount of the UEs for execution delay constraint: This section focuses on the works
which the computation offloading to the MEC is beneficial and aiming on minimization of the energy consumption while
2) the computation overhead expressed by a weighing of the satisfying maximum allowable delay, similarly as in Sec-
energy consumption and the execution delay. The distributed tion V-A2. In [88], the authors consider the application divided
algorithm performs only slightly worse than the centralized into a non-offloadable part and N offloadable parts as shown
one in both above-mentioned performance metrics. In addition, in Fig. 9b. The main objective of the paper is to decide,
the distributed algorithm significantly outperforms the cases which offloadable parts should be offloaded to the MEC.
when all UEs compute all applications locally and when all The authors propose an optimal adaptive algorithm based on
UEs prefer computing at the MEC (roughly by up to 40% for a combinatorial optimization method with complexity up to
50 UEs). O(2N ). To decrease the complexity of the optimal algorithm,
Other algorithm for the computation offloading decision also a sub-optimal algorithm is proposed reducing complexity
weighing the energy consumption at the UE and the execution to O(N ). The optimal algorithm is able to achieve up to
delay is proposed in [86]. The main difference with respect 48% energy savings while the sub-optimal one performs only
to [85] is that the authors in [86] assume the computation slightly worse (up to 47% energy savings). Moreover, it is
can be offloaded also to the remote centralized cloud (CC), shown that increasing SINR between the UE and the serving
if computation resources of the MEC are not sufficient. The eNBs leads to more prominent energy savings.
computation offloading decision is done in a sequential man- The minimization of the energy consumption while satis-
ner. In the first step, the UE decides whether to offload the fying the delay constrains of the whole application is also
application(s) to the MEC or not. If the application is offloaded the main objective of [72]. Contrary to [88] the application
to the MEC, the MEC evaluates, in the second step, if it is able in [72] is supposed to be composed of several atomic parts
dependable on each other, i.e., some parts may be processed allocation. The simulation results indicate this simplification
only after execution of other parts as shown in Fig. 10 in leads to negligibly higher total energy consumption of the UE
Section IV. The authors formulate the offloading problem as when compared to the optimal allocation. The paper is further
0 − 1 programming model, where 0 stands for the application extended in [92], where the authors show that OFDMA access
offloading and 1 represents the local computation at the UE. enables roughly ten times higher energy savings achieved by
Nevertheless, the optimal solution is of a high complexity the UEs comparing to TDMA system due to higher granularity
as there exists 2N possible solutions to this problem (i.e., of radio resources.
O(2N N 2 )). Hence, the heuristic algorithm exploiting Binary In all above-mentioned papers on partial offloading, the min-
Particle Swarm Optimizer (BPSO) [89] is proposed to reduce imization of UE’s energy consumption depends on the quality
the complexity to O(G.K.N 2 ), where G is the number of of radio communication channel and transmission power of
iterations, and K is the number of particles. The BPSO the UE. Contrary, in [93], the minimization of energy con-
algorithm is able to achieve practically the same results as sumption while satisfying execution delay of the application
the high complex optimal solution in terms of the energy is accomplished through DVS technique. In this respect, the
consumption. Moreover, the partial offloading results in more authors propose an energy-optimal partial offloading scheme
significant energy savings with respect to the full offloading that forces the UE adapt its computing power depending on
(up to 25% energy savings at the UE). maximal allowed latency of the application (LMAX ). In other
A drawback of both above papers focusing in detail on the words, the objective of the proposed scheme is to guarantee
partial computation offloading is the assumption of only single that the actual latency of the application is always equal
UE in the system. Hence, in [90], the authors address the to LMAX . As a consequence, the energy consumption is
partial offloading decision problem for the multi-UEs scenario. minimized while perceived QoS by the users is not negatively
With respect to [72][88], the application to be offloaded does 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 9a). The and the execution delay for the partial offloading decision is
UEs are assumed to be able to determine whether to partition delivered in [94]. Similarly as in [90], the application to be
the application and how many parts should be offloaded to offloaded contains only offloadable parts and in extreme case,
the MEC. The problem is formulated as a nonlinear constraint the full offloading may occur (as explained in Section V-B).
problem of a high complexity. As a consequence, it is sim- The offloading decision considers the following parameters:
plified to the problem solvable by linear programming and 1) total number of bits to be processed, 2) computational
resulting in the complexity O(N ) (N is the number of UEs capabilities 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 computation
when compared to the scenario with no offloading. In case of offloading decision is formulated as a joint optimization of
the heuristic low complex algorithm, 30% savings are observed 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 the savings is inconsequential. Moreover, the authors show the
authors assume TDMA based system where time is divided offloading is not profitable if the communication channel is
into slots with duration of T seconds. During each slot, the of a low quality since a high amount of energy is spent to
UEs may offload a part of their data to the MEC according to offload the application. In such situation, the whole application
their channel quality, local computing energy consumption, is preferred to be processed locally at the UE. With an
and fairness among the UEs. In this regard, an optimal intermediate channel quality, a part of the computation is
resource allocation policy is defined giving higher priority to offloaded to the MEC as this results in energy savings. Finally,
those UEs that are not able to meet the application latency if the channel is of a high quality, the full offloading is
constraints if the computation would be done locally. After preferred since the energy consumption for data transmission
that, the optimal resource allocation policy with threshold is low while the savings accomplished by the computation
based structure is proposed. In other words, the optimal policy offloading are high.
makes a binary offloading decision for each UE. If the UE The study in [95] provides more in-depth theoretical anal-
has a priority higher than a given threshold, the UE performs ysis on trade-off between the energy consumption and the
full computation offloading to the MEC. Contrary, if the latency of the offloaded applications preliminarily handled
UE has a lower priority than the threshold, it offloads only in [94]. Moreover, the authors further demonstrate that a
minimum amount of computation to satisfy the application probability of the computation offloading is higher for good
latency constraints. Since the optimal joint allocation of com- channel quality. With higher number of antennas (4x2 MIMO
munication and computation resources is of a high complexity, and 4x4 MIMO is assumed), the offloading is done more often
the authors also propose a sub-optimal allocation algorithm, and the energy savings at the UE are more significant when
which decouples communication and computation resource compared to SISO or MISO (up to 97% reduction of energy
 TABLE III: The comparison of individual papers addressing computation offloading decisions.
Offloading Objective Proposed solution No. of UE Evaluation Reduction of Complexity
type offloading method D/EU E wrt local of proposed
computing algorithm
[74] Full 1) Minimize D One-dimensional search algorithm Single UE Simulations Up to 80% reduc- N/A
finding the optimal offloading pol- tion of D
icy
[75] Full 1) Minimize D, 2) Lyapunov optimization-based dy- Single UE Theoretical Up to 64% reduction N/A
Minimize application namic computation offloading verifications, of D
failure simulations
[78] Full 1) Minimize EU E , 2) Online learning allocation strategy, Single UE Simulations Up to 78% reduction N/A
Satisfy D constraint offline pre-calculated strategy of EU E
[79] Full 1) Minimize EU E , 2) Deterministic and random offline Single UE Simulations Up to 78% reduction N/A
Satisfy D constraint strategies of EU E
[80] Full 1) Minimize EU E , 2) Deterministic offline strategy, de- Multi UEs Simulations N/A N/A
Satisfy D constraint terministic online strategy based on
post-decision learning framework
[81] Full 1) Minimize EU E , 2) Joint allocation of communication Multi UEs Simulations N/A N/A
Satisfy D constraint and computation resources
[82] Full 1) Minimize EU E , 2) Distributed iterative algorithm ex- Multi UEs Simulations N/A N/A
Satisfy D constraint ploiting Successive Convex Ap-
proximation (SCA)
[83] Full 1) Minimize EU E , 2) Distributed iterative algorithm ex- Multi UEs Simulations N/A N/A
Satisfy D constraint ploiting Successive Convex Ap-
proximation (SCA)
[84] Full 1) Minimize EU E , 2) Energy-efficient computation of- Multi UEs Simulations Up to 15% reduction O(max(I2+
Satisfy D constraint floading (EECO) algorithm of EU E N, IK +N ))
[85] Full 1) Trade-off between Computation offloading game Multi UEs Analytical Up to 40% reduction N/A
EU E and D evaluations, of EU E
simulations
[86] Full 1) Trade-off between Heuristic algorithm based on Single UE Simulations Up to 70% reduction N/A
EU E and D semidefinite relaxation and of total cost
randomization mapping method
[87] Full 1) Trade-off between Heuristic algorithm based on Multi UEs Simulations Up to 45% reduction O(N 6 ) per
EU E and D semidefinite relaxation and of total cost iteration
randomization mapping method
[88] Partial 1) Minimize EU E , 2) Adaptive algorithm based on com- Single UE Simulations Up to 47% reduction O(N )
Satisfy D constraint binatorial optimization method of EU E
[72] Partial 1) Minimize EU E , 2) Algorithm exploiting binary parti- Single UE Simulations Up to 25% reduction O(G.K.N 2 )
Satisfy D constraint cle swarm optimizer of EU E
[90] Partial 1) Minimize EU E , 2) Application and delay based re- Multi UEs Simulations Up to 40% reduction O(N )
Satisfy D constraint source allocation scheme of EU E
[91] Partial 1) Minimize EU E , 2) Optimal resource allocation policy Multi UEs Simulations N/A N/A
Satisfy D constraint with threshold based structure for
TDMA system
[92] Partial 1) Minimize EU E , 2) Optimal resource allocation policy Multi UEs Simulations N/A O(K + N )
Satisfy D constraint with threshold based structure for
TDMA and OFDMA system
[93] Partial 1) Minimize EU E , 2) Adapting computing power of the Single UE Simulations N/A N/A
Satisfy D constraint UE by means of DVS to achieve
maximum allowed latency
[94] Partial 1) Trade-off between Joint allocation of communication Single UE Simulations N/A N/A
EU E and D and computational resources
[95] Partial 1) Trade-off between Iterative algorithm finding the op- Single UE Analytical Up to 97% reduction N/A
EU E and D timal value of the number of bits evaluations, of EU E (SINR 45
sent in uplink simulations dB, 4x4 MIMO)
[96] Partial 1) Trade-off between Joint allocation of communication Multi UEs Simulations Up to 90% reduction N/A
EU E and D and computational resources of EU E
[97] Partial 1) Trade-off between Lyapunov optimization-based dy- Multi UEs Simulations Up to 90% reduction N/A
EU E and D namic computation offloading of EU E , up to 98%
reduction of D

consumption for 4x4 MIMO antenna configuration). Note that among multiple UEs. In the paper, it is proven that with more
the same conclusion is also reached, e.g., in [81][82]. UEs in the system, it takes more time to offload the application
The main drawback in [94][95] is that these papers consider and it also lasts longer to process the application in the MEC.
only the single-UE scenario. A trade-of analysis between the The reason for this phenomenon is quite obvious since less
energy consumption at the UE and the execution delay for the radio and computational resources remains for each UE. Still,
multi-UEs scenario is delivered in [96]. In case of the multi- up to 90% of energy savings may be accomplished in multi-
UEs scenario, the whole joint optimization process proposed UE scenario.
in [95] has to be further modified since both communication A trade-off between the power consumption and the execu-
and computation resources provided by the MEC are shared 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 re-
gard, the online algorithm based on Lyapunov optimization 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 application
to the MEC. The proposed algorithm is able to control 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%.
Fig. 14: An example of allocation of computing resources
within the MEC.
C. Summary of works focusing on computation offloading
decision
A comparison of individual computation offloading strate- decision where the individual applications should be placed
gies is illustrated in Table III. The majority of computation depends on the applications priorities (derived from the ap-
offloading decision algorithms aims to minimize the energy plication’s delay requirements, i.e., the application with a low
consumption at the UE (EUE ) while satisfying the execu- delay requirements has higher priority) and availability of the
tion delay (D) acceptable by the offloaded application or computing resources at the MEC. The basic principle for the
to find/analyse a trade-off between these two metrics. The allocation of computation resources is depicted in Fig. 15.
papers indicate up to 90% energy savings achievable by the The offloaded applications are firstly delivered to the local
computation offloading to the MEC and execution delay may scheduler within the MEC. The scheduler checks if there is
be reduced even up to 98%. Besides, all the papers evaluate a computing node with sufficient computation resources. If
the proposed solutions mostly by means of simulation (only there is a computing node with enough available resources,
several studies perform analytical evaluations). the VM is allocated at the node. Then the application is
processed at this MEC node, and finally sent back to the UE
VI. A LLOCATION OF COMPUTING RESOURCES (see Fig. 15). However, if the computation power provided
If a decision on the full or partial offloading of an applica- by the MEC server is not sufficient, the scheduler delegates
tion to the MEC is taken (as discussed in previous section), a the application to the distant CC. In order to maximize the
proper allocation of the computation resources has to be done. amount of applications processed in the MEC while satisfying
Similarly as in case of the computation offloading decision, the their delay requirements, the authors propose a priority based
selection of computation placement is influenced by the ability cooperation policy, which defines several buffer thresholds for
of the offloaded application to be parallelized/partitioned. each priority level. Hence, if the buffer is full, the applications
If the parallelization/partitioning of the application is not are sent to the CC. The optimal size of the buffer thresholds is
possible, only one physical node may be allocated for the found by means of low-complexity recursive algorithm. The
computing since the application cannot be split into several proposed cooperation policy is able to increase the probability
parts (in Fig. 14, the UE1 offloads whole application to the of the application completion within the tolerated delay by
eNB as this application cannot be partitioned). In the opposite 25%.
case, the offloaded application may be processed by resources When compared to the previous paper, the general objective
distributed over several computing nodes (in Fig. 14, the of [99] is to minimize not only the execution delay but also
application offloaded by the UE2 is partitioned and processed the power consumption at the MEC. The paper considers a
by all three eNBs). hot spot area densely populated by the UEs, which are able
This section surveys the papers addressing the problem to access several MEC servers through nearby eNBs. To that
of a proper allocation of the computing resources for the
applications that are going to be offloaded to the MEC (or
in some cases to the CC, if the MEC computing resources
are not sufficient). We categorize the research in this area into
papers focusing on allocation of the computation resources at
1) a single computing node (Section VI-A) and 2) multiple
computing nodes (Section VI-B).

A. Allocation of computation resources at a single node

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 Fig. 15: Allocation of computation resources according to [98].
end, an optimal policy is proposed using equivalent discrete
MDP framework. However, this method results in a high
communication overhead and high computational complexity
with increasing number of the MEC servers. Hence, this
problem is overcome by developing an application assignment
index policy. In this respect, each eNB calculates its own index
policy according to the state of its computing resources. Then,
this index policy is broadcasted by all eNBs and the UE is able
to select the most suitable MEC server in order to minimize
both execution delay and power consumption. According to
the results, the index policy is in the worst case by 7% more
costly than optimal policy in terms of system cost (note that
system cost represents weighted execution delay and power
consumptions of the MEC). Fig. 17: An example of allocation of computation resources
The minimization of the execution delay of the offloaded for individual UEs according to [101].
application is also the main goal in [100]. Nonetheless, with
respect to [98][99], the other objectives are to minimize
both communication and computing resource overloading and to the main objective: 1) minimize execution delay and/or
the VM migration cost (note that in [100], the computing power consumption of computing nodes (Section VI-B1) and
nodes are represented by the SCeNBs and the VM migration 2) balance both communication and computing loads (Sec-
may be initiated due to the SCeNBs shutdown). The whole tion VI-B1).
problem is formulated as the VM allocation at the SCeNB and 1) Minimization of execution delay and/or power consump-
solved by means of MDP. An example of the VM allocation tion of computing nodes: The minimization of the execution
according to [100] is shown in Fig. 16, where the VM for the delay by allocation of computing resources provided by the
UE1 is allocated at the serving SCeNB1 while the UE2 has cluster of SCeNBs while avoiding to use the CC is proposed in
allocated the VM at the neighbouring SCeNB3. The SCeNB3 [101]. The cluster formation is done by means of a cooperative
is preferred because of a high quality backhaul resulting in a game approach, where monetary incentives are given to the
low transmission delay of the offloaded data. The simulations SCeNBs if they perform the computation for the UEs attached
show that with higher VM migration cost, the VM is preferred to other SCeNBs. The coalition among the SCeNBs is formed
to be allocated at the serving SCeNB (i.e., the SCeNB closest for several time slots and then new coalitions may be created.
to the UE) if this SCeNB has enough computation power. The allocation of computation resources is done as shown in
The main disadvantage of all above-mentioned approaches Fig. 17. Firstly, the serving SCeNB tries to serve their UEs on
is that these do not consider more computing nodes within its own since this results in the shortest communication delay
the MEC for single application in order to further decrease its (e.g., in Fig. 17 SCeNB1 allocates the computation resources
execution delay. to the UE1 and the UE2, etc.). Only if the SCeNB is not
able to process the application on its own, it is forwarded to
all SCeNBs in the same cluster (in Fig. 17, the computation
B. Allocation of computation resources at multiple nodes for the UE3 is done at the SCeNB2 and the SCeNB3). The
(federated clouds) numerical results show that the proposed scheme is able to
When compared to the previous section, the allocation of reduce the execution delay by up to 50% when compared
computation resources at multiple computing nodes is con- to the computation only at the serving SCeNB and by up
sidered here. The papers are split into subsections according to 25% comparing to the scenario when all SCeNBs in
the system participate in the computation. Unfortunately, the
proposed approach does not address a problem of forming new
coalitions and its impact on currently processed applications.
The selection of computing nodes can significantly influence
not only the execution delay, as considered in [101], but also
the power consumption of the computing nodes. Hence, the
main objective of [102] is to analyze an impact of the cluster
size (i.e., the amount of the SCeNBs performing computing)
on both execution latency of the offloaded application and
the power consumption of the SCeNBs. The analysis is done
for different backhaul topologies (ring, tree, full mesh) and
technologies (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
Fig. 16: An example of the VM allocation at single computing (up to 90% execution delay reduction). Contrary, a fiber back-
SCeNB according to [100]. haul in ring topology results in the lowest power consumption.
Moreover, the paper shows that an increasing number of the each SCeNB allocates its computational resources to their
computing SCeNBs does not always shorten execution delay. own UEs according to specific scheduling rules, such as
Quite the opposite, if a lot of SCeNBs process the offloading application latency constraint, computation load or minimum
applications and the transmission delay becomes longer than required computational capacity. In the second step, labelled as
the computing delay at the SCeNBs, the execution delay may establishment of computing clusters, the computation clusters
be increased instead. Besides, with an increasing number of are created for each UE that cannot be served by its serving
the computing SCeNBs, power consumption increases as well. SCeNB. The authors propose three algorithm realizations
Consequently, a proper cluster formation and the SCeNBs differing in applications prioritization (e.g., earliest deadline
selection play a crucial part in system performance. first or according to computation size of application) and the
The problem to find an optimal formation of the clusters of objective (minimization of power consumption or execution
SCeNBs for computation taking into account both execution latency similarly as, e.g., in [104]). The simulations illustrate
delay and power consumption of the computing nodes is ad- that there could be found the algorithm realization resulting
dressed in [103]. The paper proposes three different clustering in the users satisfaction ratio above 95% while keeping a
strategies. The first clustering strategy selects the SCeNBs in moderate power consumption of all computing nodes.
order to minimize execution delay. Since all SCeNBs in the 2) Balancing of communication and computation load: In
system model are assumed to be one hop away (i.e., full mesh the previous section, the allocation of computing resources
topology is considered), basically all SCeNBs are included in is done solely with purpose to minimize the execution delay
the computation resulting in up to 22% reduction of execution and/or the power consumption of the computing nodes. This
delay. This is due to the fact that the computation gain could, however, result in unequal load distribution among
(and, thus, increase in the offloaded application processing) individual computing nodes and backhaul overloading. The
is far greater than the transmission delay. The objective of balancing of communication and computation load of the
the second clustering strategy is to minimize overall power SCeNBs while satisfying the delay requirement of the of-
consumption of the cluster. In this case, only the serving floaded application is addressed in [106]. To this end, an
SCeNB is preferred to compute, thus, any computation at Application Considering Algorithm (ACA) selecting suitable
the neighbouring SCeNBs is suppressed to minimize power SCeNBs according to the current computation and commu-
consumption of the SCeNBs (up to 61% reduction of power nication load of the SCeNBs is proposed. The ACA exploits
consumption is observed). This, however, increases overall knowledge of the offloaded application’s requirements (i.e.,
latency and high variations of the computation load. The last the number of bytes to be transferred and the maximum
clustering strategy aims to minimize the power consumption latency acceptable by the application/user). The selection of
of each SCeNB in the cluster, since the power consumptions the SCeNBs for the computation is done in a static way
of the individual SCeNBs is highly imbalanced in the second prior to the offloading to avoid expensive VMs migration.
strategy. The performance evaluation is done for two backhauls, low
While in [103] the optimal clustering of the SCeNBs is throughput ADSL and high quality gigabit passive optical
done only for single UE, the multi UEs scenario is assumed network (GPON). The proposed ACA algorithm is able to
in [104]. When compared to the previous paper, whenever the satisfy 100% of the UEs as long as number of offloaded tasks
UE is about to offload data for the computation, the computing per second is up to 6. Moreover, the paper shows that tasks
cluster is assigned to it. Consequently, each UE has assigned parallelization helps to better balance computation load.
different cluster size depending on the application and the
UE’s requirements. The core idea of the proposal is to jointly
compute clusters for all active users’ requests simultaneously
to being able efficiently distribute computation and communi-
cation resources among the UEs and to achieve higher QoE.
The main objective is to minimize the power consumption of
the clusters while guaranteeing required execution delay for
each UE. The joint clusters optimization is able to significantly
outperform the successive cluster optimization (allocation of
the clusters are done subsequently for each UE), the static
clustering (equal load distribution among SCeNBs) and no
clustering (computation is done only by the serving SCeNB)
in terms of the users’ satisfaction ratio (up to 95% of UEs is
satisfied). On the other hand, the average power consumption
is significantly higher when compared to ”no clustering” and
”successive clusters optimization” scenarios.
Similar as in [104], the multi-UE cluster allocation is
assumed in [105], but the cluster formation is done jointly with
the UEs scheduling. The proposed resource allocation process Fig. 18: An example of application and physical graph accord-
is split into two steps similarly as proposed in [101]. In the ing to [107] (FD - Face detection, IPFE - Image processing
first step, labeled as local computational resource allocation, and feature extraction, FR - Face recognition, D - Database).
TABLE IV: The comparison of individual papers addressing allocation of computation resources for application/data already
decided to be offloaded.
No. of com- Objective Proposed solution Computing Evaluation Results
puting nodes nodes method
for each ap-
plication
[98] Single node 1) Maximize the amount of Priority based cooperation policy MEC servers Simulations 25% reduction of D
served applications, 2) Sat- (e.g., at the eNB wrt offloading only
isfy D constraint or agg. point), CC to the MEC
[99] Single node 1) Minimize D, 2) Mini- Optimal online application assign- MEC servers Simulations N/A
mize EC ment policy using equivalent dis- (e.g., at the eNB
crete MDP framework or agg. point), CC
[100] Single node 1) Minimize D, 2) Mini- Optimal allocation policy obtained SCeNBs Simulations N/A
mize overloading of com- by solving MDP using linear pro-
munication and computing graming reformulation
resources, 3) Minimize VM
migration cost
[101] Multiple nodes 1) Minimize D, 2) Avoid Formation of collaborative coali- SCeNBs Simulations Up to 50% reduction
to use the CC due to high tions by giving monetary incentives of D wrt single com-
delay to the SCeNBs puting SCeNB
[102] Multiple nodes 1) Analyze the impact of N/A SCeNBs Simulations Up to 90% reduction
different network topologies of D wrt single com-
and technologies on execu- puting SCeNB
tion delay and power con-
sumption
[103] Multiple nodes 1) Minimize D, 2) Mini- Three clustering strategies mini- SCeNBs Simulations 22% reduction of
mize EC mizing delay, power consumption D, 61% reduction of
of the cluster and power consump- EC
tion of the SCs
[104] Multiple nodes 1) Minimize D, 2) Mini- Joint cluster formation for all active SCeNBs Simulations Up to 95% of UE
mize EC users requests simultaneously are satisfied (for
max. 5 UEs)
[105] Multiple nodes 1) Minimize D, 2) Mini- Joint cluster formation for all ac- SCeNBs Simulations Up to 95% of UE
mize EC tive users requests simultaneously are satisfied (for
together with users scheduling max. 5 UEs)
[106] Multiple nodes 1) Balance communication ACA algorithm assuming jointly SCeNBs Simulations 100% satisfaction
and computation load of computation and communication ratio for up to 6
computing nodes, 2) Satisfy loads offloaded tasks/s
execution delay requirement
[107] Multiple nodes 1) Balance communication Online approximation algorithms UE, eNB, CC Simulations Reduction of
and computation load of with polynomial-logarithmic (poly- resource utilization
computing nodes, 2) Mini- log) competitive ratio for tree ap- up to 10%
mize resource utilization plication graph placement

The main objective to balance the load (both communi- C. Summary of works dealing with allocation of computing
cation and computation) among physical computing nodes resources
and, at the same time, to minimize the resource utilization
of each physical computing node (i.e., reducing sum resource The comparison of individual methods addressing allocation
utilization) is also considered in [107]. The overall problem is of the computation resources within the MEC is shown in
formulated as a placement of application graph onto a physical Table IV. The main objective of the studies dealing with
graph. The former represents the application where nodes in the allocation of computation resources is to minimize the
graph correspond to individual components of the application execution delay of the offloaded application (D). In other
and edges to the communication requirements between them. words the aim is to ensure QoS to the UEs in order to fully
The latter represents physical computing system, where the exploit proximity of the MEC with respect to the computing
nodes in graph are individual computing devices and edges in faraway CC. Moreover, several studies also focus on mini-
stands for the capacity of the communication links between mization of the energy consumption of computing nodes (EC ).
them (see the example of application and physical graphs In addition, some limited effort has been focused on balancing
in Fig. 18 for the face recognition application). The authors of computing and communication load to more easily satisfy
firstly propose the algorithm finding the optimal solution for the requirements on execution delay and/or to minimize overall
the linear application graph and, then, more general online resources utilization.
approximation algorithms. The numerical results demonstrate A common drawback of all proposed solutions is that only
that the proposed algorithm is able to outperform two heuristic simulations are provided to demonstrate proposed solutions
approaches in terms of resource utilization by roughly 10%. for allocation of MEC computing resources. Moreover, all
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 MANAGEMENT FOR MEC

In the conventional mobile cellular networks, a mobility of
users is enabled by handover procedure when the UE changes
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 ensure the service continuity. In fact, there are several Fig. 19: Principle of CaPC according to [108][109].
options how to cope with the mobility of UEs. The first option,
applicable only for the UEs with a low mobility (e.g. within
a room), is to adapt transmission power of the eNB/SCeNB power is triggered (∆t ) is the same for all SCeNBs and UEs
during the time when the offloaded application is processed independently on the channel quality (i.e., SINR). As a con-
by the MEC (Section VII-A). If the UE performs handover sequence, the CaPC may be triggered too late when sufficient
to the new serving eNB/SCeNB despite of the power control, SINR cannot be guaranteed in due time to successfully deliver
the service continuity may be guarantee either by the VM the offloaded application back to the UE. This problem is
migration (i.e., the process during which the VM run at addressed in [109], where the ∆t is set individually for each
the current computing node(s) is migrated to another, more UEs depending on its current channel quality. The proposed
suitable, computing node(s) as discussed in Section VII-B) or algorithm finds ∆t by iterative process when ∆t is adapted
by selection of a new communication path between the UE after each application is successfully delivered back to the UE.
and the computing 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., B. VM migration
when the UEs are slowly moving inside a building, a proper If the UEs mobility is not limited, as considered in Sec-
setting of the transmission power of the serving and/or neigh- tion VII-A, and power control is no longer sufficient to keep
boring SCeNBs can help to guarantee QoS. This is considered the UE at the same serving eNB/SCeNB, a possibility to
in [108], where the authors propose a cloud-aware power initiate the VM migration should be contemplated in order to
control (CaPC) algorithm helping to manage the offloading guarantee the service continuity and QoS requirements. On one
of real-time applications with strict delay requirements. The hand, the VM migration has its cost (CostM ) representing the
main objective of the CaPC is to maximize the amount of time required for the VM migration and backhaul resources
the offloaded applications processed by the MEC with a given spent by transmission of the VM(s) between the computing
latency constrain. This is achieved by an adaptation of the nodes. On the other hand, there is a gain if the VM migration is
transmission power of the SCeNBs so that the handover to a initiated (GainM ) since the UE can experience lower latency
new SCeNB is avoided if possible (see the basic principle (data is processed in UE’s vicinity) and backhaul resources do
in Fig.19 where the moving UE remains connected to the not have to be allocated for transmission of the computation
same SCeNB as its transmission power is increased). The results back to the UE.
CaPC is composed of coarse and fine settings of the SCeNBs A preliminary analysis how the VM migration influences
transmission power. The purpose of the coarse setting is to performance of the UE is tackled in [110]. The authors
find an optimal default transmission power Pt,def , which is describe analytical model based on Markov chains. Without
applied if all of the UEs attached to the SCeNB are idle. the VM migration, a probability that the UE is connected to
Setting of the Pt,def depends on the power level received the optimal MEC decreases with increasing number of hops
by the serving SCeNB from the most interfering neighboring between the UE and the eNB, where the service is initially
SCeNB and the interference generated by the eNBs. The fine placed. This also results in increasing delay. Contrary, the
setting consists in a short-term adaptation of the SCeNB’s connection of the UE to the optimal MEC server results in
transmission power when the UE would not be able to receive the lowest delay but at the high cost of the migration. The
the offloaded application from the cloud due to low SINR. reason for this phenomenon is that the VM migration should
If the CaPC is utilized, up to 95% applications computed at be ideally initiated after each handover performed by the UE
the SCeNBSs are successfully delivered back to the UE with to keep minimum delay.
satisfying delay. Contrary, a conventional, non-cloud-aware, While the previous paper is more general and focused on
power control is able to successfully deliver only roughly 80% preliminary analysis regarding the VM migration, the main
of offloaded applications. objective of [111] is to design a proper decision policy
The main disadvantage of the CaPC presented in [109] is determining whether to initiate the VM migration or not. As
that the time when the fine adjustment of the transmission discussed above, there is a trade-off between the migration
 optimal 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][113] is that these assume
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
sequential decision making problem for the VM migration
using MDP and define algorithm for finding optimal policy
Fig. 20: VM migration principle according to [111]. with the complexity O(N 3 ), where N is the number of states
(note that the state is defined as the number of hops between
the eNB to which the UE is connected and the location of
cost (CostM ) and migration gain (GainM ). The authors the MEC server analogously to [113]). Since the proposed
formulate the VM migration policy as a Continuous Time optimal VM migration strategy is too complex, the authors
MDP (CTMDP) and they try to find an optimal threshold propose an approximation of the underlying state space by
policy when the VM migration is initiated. Consequently, after defining the space as a distance between the UE and the
each handover to the new eNB, the optimal threshold policy MEC server where the service is running. In this case, the
decides whether the VM migration should be initiated or not. time complexity is reduced to O(N 2 ). As demonstrated by
An example of this principle is shown in Fig. 20, where the numerical evaluations, the proposed migration strategy is
the UE exploiting the MEC1 moves from the eNB1 to the able to decrease sum cost by roughly 35% compared to both
eNBn. While the conventional radio handover is performed never and always 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 scheme is able to: 1) estimate in advance a throughput that
optimal policy always achieves the maximum expected gain user can receive from individual MEC servers as it roams
if 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 ) the proposed scheme is able to decrease latency by 35%
and VM migration gain (GainM ) is also studied in [112]. with respect to scheme proposed in [111]. Nonetheless, a
The paper proposes a Profit Maximization Avatar Placement disadvantage of the proposal is that it requires huge amount
(PRIMAL) strategy deciding whether the VM should be of information in order to predict the throughput. Moreover,
migrated or not. Since the PRIMAL problem is NP-hard, the paper does not consider the cost of migration itself.
the authors use Mixed-Integer Quadratic Programming tool In [116], the VM migration decision process is further
to find the heuristic solution. The proposed solution is able enhanced by the mechanism predicting future migration cost
to significantly reduce execution delay when compared to the with specified upper bound on a prediction error. The main
situation with no migration (roughly by 90%) while reducing objective of the paper is, similarly as in [113][114], to
the migration cost approximately by 40%. When compared minimize the sum cost over a given time. First, the authors
to [111], the authors also show the influence of α parameter propose an offline algorithm for finding the optimal placement
weighing CostM and GainM . Basically, with increasing α, sequence for a specific look-ahead window size T , which
the migration cost is decreasing (i.e., migration is not done so represents the time to which the cost prediction is done.
frequently), but at the cost of higher execution delay. For the offline algorithm, an arrival and a departure of the
An optimal threshold policy for the VM migration is also applications offloaded to the MEC are assumed to be exactly
considered in [113]. The problem is again formulated as the known. The time complexity of the algorithm is O(M 2 T ),
MDP and the VM migration is initiated always if the state of where M stands for the number of MEC serves in the system.
the UE is bounded by a particular set of thresholds. The state The VM migration is strongly dependent on the size of T . If
of the UE is defined as the number of hops between the eNB to T is too large, the future predicted values may be far away
which the UE is connected and the location of the MEC server from the actual values and, thus, the VM migration far from
where the computing service is running (in the paper labelled the optimal. Contrary if T is too short, a long term effect of
as the offset). The main objective of the paper is to minimize the service placement is not considered. As a result, also a
the overall sum cost by the optimal VM migration decision binary search algorithm finding the optimal window size is
(i.e., the VM migration is performed if CostM < GainM proposed in the paper. The proposed offline algorithm is able
as explained earlier). The authors proof the existence of the to reduce cost by 25% (compared to never migrate strategy)
and by 32% (compared to always migrate strategy). Although based on Locator/Identifier Separation Protocol (LISP) [121].
the simulation results are demonstrated for the multi-UEs If the user is experiencing latency or jitter above maximum
scenario, the problem is formulated only for the single-UE. tolerated threshold, the VM migration to a new MEC server is
Hence, the paper is further extended in [117] for the multi- initiated. The selection of the new MEC server, which is about
UEs offloading K applications to the MEC. Similarly as in to host VM of the user is based on the required computing
[116], the problem is solved by the offline algorithm with power and availability of resources at the MEC servers. The
complexity of O(M K2 T ). Since the offline algorithm is of proposal is evaluated by means of both experiments on real
high complexity and impractical for real systems, the paper testbed and simulations. The results show that the system
also propose an online approximation algorithm reducing the throughput is increased by up to 40% when compared to the
complexity to O(M 2 KT ). The proposed online algorithm case without VM migration.
outperforms never migrate and always migrate strategies by
approximately 32% and 50%, respectively.
C. Path selection and/or VM migration
So far, all the studies focusing on the VM migration do
not consider an impact on a workload scheduling, i.e., how The VM migration is not a convenient option when a huge
the VM migration would be affected by a load of individual amount of data needs to be migrated among the computing
MEC servers. As suggested in [118], the problem of the VM nodes and the whole process may take minutes or even hours
migration and scheduling of the MEC workloads should be [119]. Even if the migration process lasts few seconds, real-
done jointly. Although the problem could be formulated as time applications cannot be offloaded to the MEC. Moreover,
a sequential decision making problem in the framework of the load imposed on backhaul links may be too significant.
MDPs (like in above studies) it would suffer from several In such cases, finding and optimizing new paths for delivery
drawbacks, such as, 1) extensive knowledge of the statistics of of the computed data from the MEC are a more viable
the users mobility and request arrival process is impractical, option. This eventuality is considered in [122], where the path
2) problem can is computationally challenging, and 3) any selection algorithm for a delivery of the offloaded data from
change in the mobility and arrival statistics would require re- the cluster of computing SCeNBs to the UE is proposed. The
computing the optimal solution. Hence, the main contribution main objective of the path selection algorithm is to minimize
of [118] is a development of a new methodology overcoming transmission delay taking into account quality of both radio
these drawbacks inspired by Lyapunov optimization frame- and backhaul links. Moreover, the authors enable to enforce
work. The authors propose online control algorithm making handover to new serving SCeNB to minimize the transmission
decision on where the application should be migrated so that delay. An example of the data delivery is shown in Fig. 21,
the overall transmission and reconfiguration costs are mini- where three SCeNBs are computing the application offloaded
mized. The complexity of the algorithm is O(M !/(M − K)!), by the UE. The data processed by the serving SCeNB2 are
where M is the number of MEC servers and K is the amount received by the UE directly while data from the SCeNB3 are
of applications host by the MEC. By means of proposed delivered to the UE through the CN and the serving SCeNB2.
optimization framework, the reconfiguration cost is reduced Finally, the UE performs handover to the SCeNB1 and, then, it
when compared to always migrate strategy (by 7%) and never receives results from the SCeNB3 directly via radio link. The
migrate strategy (by 26%). complexity of the proposed algorithm is O(mn ), where m is
While the main objective of the previous papers focusing the number of UEs and n the amount of the SCeNBs in cluster.
on the VM migration is to make a proper decision on whether The proposed algorithm is able to reduce transmission delay
to migrate or not, the main aim of the authors in [119] is by up to 9% with respect to a case when the UE receives all
to minimize the VM migration time when the migration is data from the same serving SCeNB. In [123], the algorithm’s
about to be performed. This is accomplished by a compression complexity is further decreased to O(I n ), where I is the set
algorithm reducing the amount of transmission data during the of SCeNBs with sufficient radio/backhaul link quality. It is
migration itself. On one hand, if the compression rate of the
algorithm is low, more data has to be transmitted, but the com-
pression itself is shorter in terms of time. On the other hand, a
higher compression rate results in a significant reduction of the
transmitted data during the VM migration, but the compression
takes significant amount of time. Hence, the paper proposes
a dynamic adaptation of the compression rate depending on
the current backhaul available bandwidth and the processing
load of the MEC. The paper presents extensive experiments
on real system showing that the dynamic adaptation during
the VM migration is able to cope with changing of available
bandwidth capacity.
A proper VM migration may not result only in an execution
delay reduction, but it can also increase throughput of the sys- Fig. 21: An example of path selection algorithm proposed
tem as demonstrated in [120]. The paper proposes a protocol in [122] (cR B
x and Cx stands for capacity of radio links and
architecture for cloud access optimization (PACAO), which is backhaul links, respectively).
 TABLE V: The comparison of individual papers focusing on mobility management in MEC.
Mobility Objective Proposed method Mobility Evaluation Results Algorithm
man. model method complex-
method ity
[108] Power 1) Maximize the amount Adaptation of transmission 2D limited Simulations Up to 95% offloaded -
control of delivered requests power of SCeNBs mobility (e.g., applications successfully
from the MEC, 2) apartment) delivered
Guaranteeing latency
constraints
[109] Power 1) Maximize the amount Adaptation of transmission 2D limited Simulations Up to 98% offloaded -
control of delivered requests power of SCeNBs, optimiza- mobility (e.g., applications successfully
from the MEC, 2) tion of power control trigger apartment) delivered
Guaranteeing latency time
constraints
[110] VM mi- 1) Define analytical - 2D random Analytical N/A -
gration model for VM walk model model
migration, 2) Analyze
how VM migration
influences e2e delay
[111] VM mi- 1) Maximize total ex- Formulation of an optimal 1D random Simulations Always maximize total -
gration pected reward threshold decision policy ex- walk expected reward wrt al-
ploiting MDP ways/never migrate
[112] VM mi- 1) Find a trade-off Profit Maximization Avatar Random way Simulations Reducing execution de- -
gration between CostM and Placement (PRIMAL) strat- point model lay by 90% wrt no mi-
GainM egy deciding whether VM gration, reducing migra-
should be migrated or not tion cost by 40% wrt to
always migrate
[113] VM mi- 1) Minimize system sum Formulation of an optimal 1D asymmetric Simulations Always minimize over- O(|M |N )
gration cost over a given time threshold decision policy us- random walk all cost wrt always/never
ing MDP mobility model migrate
[114] VM mi- 1) Minimize system sum Formulation of an optimal 2D mobility, Simulations 30% reduction of O(N 2 )
gration cost over a given time sequential decision policy real mobility average cost wrt to
using MDP traces never/always migrate
[115] VM mi- 1) Minimize execution Estimation of throughput of- Cars moving Simulations Reducing latency by -
gration delay fered by MEC servers based by a predefined 35% wrt [111]
on mobility prediction paths
[116] VM mi- 1) Minimize system sum Offline algorithm for find- Real world user Simulations 25%(32%) red. of O(M 2 T )
gration cost over a given time ing optimal placement se- mobility traces average cost wrt to
quence for a specific look- never(always) migrate
ahead window size
[117] VM mi- 1) Minimize system sum Offline and online algorithms Real world user Analytical, 32%(50%) red. of O(M K2 T )
gration cost over a given time for finding optimal place- mobility traces simulations average cost wrt to
ment sequence for a specific never(always) migrate
look-ahead window size
[118] VM mi- 1) Minimize overall Online control algorithm 1) Random Analytical 7%(26%) red. of re- O(M !/(M −
gration transmission and making decision where walk, 2) Real evalu- configuration cost wrt to K)!) for
reconfiguration costs application should be placed world user ations, never(always) migrate M ≥K
and migrated mobility traces simulations
[119] VM mi- 1) Minimize VM migra- Adaptation of compression - Experiments N/A -
gration tion time rate during VM migration on real
depending on available band- system
width and processing load
[120] VM mi- 1) Maximize throughput Protocol architecture for Real world user Experiments Increase of throughput -
gration cloud access optimization mobility traces on testbed, up to 40%
exploiting LISP simulations
[122] Path se- 1) Minimize transmis- Path selection exploiting Manhattan mo- Simulations Reduction of transmis- O(mn )
lection sion delay handover mechanism bility model sion delay by up to 9%
[123] Path se- 1) Minimize transmis- Path selection exploiting Manhattan mo- Simulations Reduction of transmis- O(Z n )
lection sion delay handover mechanism bility model sion delay by up to 54%
[124] Path se- 1) Minimize transmis- Cooperative service migra- Smooth random Simulations Reduction of transmis- O(|Z||I|τ ),
lection + sion delay ton and path selection algo- mobility model sion delay by up to 10% O(|I|τ )
VM mi- rithm with movement predic- wrt [123]
gration tion

shown that the proposed path selection algorithm is able to the dynamic VM migration and the path selection algorithm
reduce transmission delay by 54%. proposed in [123] further enhanced by consideration of a
mobility prediction. The first algorithm decides whether the
The path selection algorithm contemplated in [122][123] VM migration should be initiated or not based on the mobility
may not be sufficient if the UE is too far away from the prediction and the computation/communication load of the
computing location since increased transmission delay may eNB(s). The second algorithm, then, finds the most suitable
result in QoS reduction notwithstanding. Hence, the authors route for downloading the offloaded data with the mobility
in [124] suggest a cooperation between an algorithm for
prediction outcomes taken into account. The complexity of the sending only small amount of data [82]. The reason is that
first algorithm is O(|Z||I|τ ) and the complexity of the second the energy spent by transmission/reception of the offloaded
algorithm equals to O(|I|τ ), where Z is the number of eNBs computing is small while the energy savings achieved by
with sufficient channel quality and computing capacity, and τ the computation offloading are significant. Contrary, the
stands for the size of the prediction window. The proposed applications that need to offload a lot of data should
algorithm is able reducing the average offloading time by be computed locally as the offloading simply does not pay
27% comparing to the situation when the VM migration is off due to huge amount of energy spent by the offloading
performed after each conventional handover and by roughly and high offloading delays.
10% with respect to [123]. • If the computing capacities at the MEC are fairly
limited, the probability to offload data for processing
D. Summary of works focused on mobility management is lowered. This is due to the fact that the probabilities of
A comparison of the studies addressing the mobility issues the offloading and local processing are closely related to
for the MEC is shown in Table V. As it can be observed 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 offloading
VM migration. Basically, the related papers try to find an as well as its processing at the MEC last longer [96].
optimal decision policy whether the VM migration should Consequently, if there is high amount of UEs in the system,
be initiated or not to minimize overall system cost (up to the local processing may be more profitable, especially if
32% and up to 50% reduction of average cost is achieved the minimization of execution delay is the priority (such is
compared to never and always migrate options, respectively the case of real-time applications).
[117]). Moreover, some papers aim to find a proper trade-off • The energy savings achieved by the computation offload-
between VM migration cost and VM migration gain [112], ing is strongly related to the radio access technology
minimizing execution delay [115], minimizing VM migration used at radio link. To be more specific, OFDMA enables
time [119], or maximizing overall throughput [120]. significantly higher energy savings of the UEs than TDMA
From Table V can be further observed that all papers dealing due to higher granularity of radio resources [92].
with the VM migration assume the computation is done by a • The partial offloading can save significantly more energy
single computing node. Although this option is less complex, at the UE when compared to the full offloading [72].
the parallel computation by more nodes should not be entirely Nevertheless, in order to perform the partial offloading,
neglected as most of the papers focusing on the allocation of the application has to enable parallelization/partitioning.
computing resources assume multiple computing nodes (see Hence, the energy savings accomplished by computation
Section VI-B). offloading is also strongly related to the application type
and the way how the code of the application is written.
VIII. L ESSONS LEARNED
From the surveyed papers focused on allocation of comput-
This section summarizes lessons learned from the state of
ing resources, the following key facts are learned:
the art focusing on computation offloading into the MEC. We
again address all three key items: decision on computation • The allocation of computation resources is strongly
offloading, allocation of computing resources, and mobility related to the type of the application being offloaded
management. in a sense that only applications allowing paralleliza-
From the surveyed papers dealing with the decision on com- tion/partitioning may be distributed to multiple computing
putation offloading, following key observations are derived: nodes. Obviously, a proper parallelization and code
• If the channel quality between the UE and its serving partitioning of the offloaded application can result in
station is low, it is profitable to compute rather locally shorter execution delays as multiple nodes may pool their
[95]. The main reason is that the energy spent by the trans- computing resources (up to 90% reduction of execution
mission/reception of the offloaded data is too expensive delay when compared to single computing node). On the
in terms of the energy consumption at the UE. Contrary, other hand, the allocation of computation resources for
with increasing quality of the channel, it is better to parallelized applications is significantly more complex.
delegate the computation to the MEC since the energy • An increase in the number of computing nodes does
cost required for transmission/reception of the offloaded not have to result always in a reduction in the execu-
data is reduced and it is easily outweighed by the energy tion delay [102]. On the contrary, if the communication
saving due to the remote computation. Consequently, the delay becomes predominant over the computation delay,
computation can be offloaded more frequently if MIMO the overall execution delay may be even increased. Hence,
is exploited as it improves channel quality. Moreover, it a proper trade-off between the number of computing nodes
is efficient to exploit connection through SCeNBs for and execution delay needs to be carefully considered when
the offloading as the SCeNBs are supposed to serve fewer allocating computing resources to offloaded data.
users in proximity providing high channel quality and more • If the backhaul is of a low quality, it is mostly preferred
available radio resources. to perform the computation locally by the serving
• The most suitable applications for offloading are those node (e.g., SCeNB/eNB) since the distribution of data for
requiring high computational power (i.e., high compu- computing is too costly in terms of the transmission latency.
tational demanding applications) and, at the same time, Contrary, a high quality backhaul is a prerequisite for
 an efficient offloading to multiple computing nodes. selection should be adopted [124].
• The execution delay of the offloaded application de-
pends not only on the backhaul quality, but also on a IX. O PEN RESEARCH CHALLENGES AND FUTURE WORK
backhaul topology (e.g., mesh, ring, tree, etc.) [102]. The
As shown in the previous sections, the MEC has attracted a
mesh topology is the most advantageous in terms of the
lot of attention in recent years due to its ability to significantly
execution delay since all computing nodes are connected
reduce energy consumption of the UEs while, at the same time,
directly and distribution of the offloaded data for computing
enabling real-time application offloading because of proximity
is more convenient. On the other hand, mesh topology
of computing resources to the users. Despite this fact the
would require huge investment in the backhaul.
MEC is still rather immature technology and there are many
Finally, after surveying the papers addressing mobility is- challenges that need to be addressed before its implementation
sues in the MEC, we list following key findings: into mobile network to be beneficial. This section discusses
• There are several options of the UE’s mobility management several open research challenges not addressed by the current
if the data/application is offloaded to the MEC. In cases of researcher.
the low mobility, the power control at the SCeNBs/eNBs
side can be sufficient to handle mobility (up to 98% of A. Distribution and management of MEC resources
offloaded applications can be successfully delivered back
to the UE [109]). This is true as long as the adaption of In Section III, we have discussed several possible options for
transmission power enables keeping the UE at the same placement of the computing nodes enabling the MEC within
serving station during the computation offloading. However, the mobile network architecture. To guarantee ubiquitous
if the UE performs handover, the power control alone is not MEC services for all users wanting to utilize the MEC, the
sufficient and the VM migration or new communication MEC servers and the computation/storage resource should
path selection may be necessary to comply with require- be distributed throughout whole network. Consequently, the
ments of offloaded applications in terms of latency. individual options where to physically place the MEC servers
• A decision on VM migration depends strongly on three should complement each other in a hierarchical way. This
metrics: will allow efficient usage of the computing resources while
respecting QoS and QoE requirements of the users. In this
1) The VM migration cost (CostM ) representing the time
context, an important challenge is to find an optimal way
required for the service migration and the backhaul
where to physically place the computation depending on
resources spent by the transmission of VM(s) between
expected users demands while, at the same time, consider
the computing nodes.
related CAPEX and OPEX (as initially tackled in [67][68]).
2) The VM migration gain (GainM ) is the gain consti-
Another missing topic in the literature is a design of
tuting delay reduction (data are computed in proximity
efficient control procedures for proper management of the
of the UE) and saving of the backhaul resources (data
MEC resources. This includes design of signalling messages,
does not have to be sent through several nodes).
their exchange and optimization in terms of signalling over-
3) The computing load of the node(s) to which the VM is
head. The control messages should be able to deliver status
reallocated since, in some situations, the optimal com-
information, such as load of individual computing nodes
puting node for the VM migration may be unavailable
and quality of wireless/backhaul links in order to efficiently
due to its high computation load.
orchestrate computing resources within the MEC. There is a
• The VM migration is impractical if huge amount of data trade-off between high signalling overhead related to frequent
needs to be transmitted between the computing nodes exchange of the status information and an impact on the MEC
and/or if the backhaul resources between VMs are inad- performance due to aging of the status information if these are
equate since it may take minutes or even hours to migrate exchanged rarely. This trade-off have to be carefully analysed
whole VM. This is obviously too long for real-time services and efficient signalling mechanisms need to be proposed to
and it also implies significant load on backhaul, especially ensure that the control entities in the MEC have up to date
if the VM migration would need to be performed frequently. information at their disposal while the cost to obtain them is
Note that time consuming migration goes against the major minimized.
benefit of the MEC, i.e., low latency resulting in suitability
of the offloading for real-time services.
• The minimization of the VM migration time can be B. Offloading decision
done by reduction of the amount of migrated data [119]. The offloading decision plays a crucial part as it basically
Nonetheless, even this option is not enough for real-time determines whether the computation would be performed
services. Thus, various path selection algorithms should locally, remotely or jointly in both locations as discussed in
be employed with purpose to find the optimal path for Section V. All papers focusing on the offloading decision
delivery of the offloaded data back to the UEs while consider only the energy consumption at the side of the UE.
computing is done by the same node(s) (i.e., without VM However, to be in line with future green networking, also
migration) [123]. However, if the UE moves too far away the energy consumption at the MEC (including computation
from the computation placement, more robust mobility as well as related communication) should be further taken
management based on joint VM migration and path into account during the decision. Moreover, all papers dealing
with the offloading decision assume strictly static scenarios, efficiently handle the VM migration procedure when appli-
i.e., the UEs are not moving before and during the offload- cation is offloaded to several computing nodes. Moreover,
ing. Nevertheless, the energy necessary for transmission of the VM migration impose high load on the backhaul and
the offloaded data can be significantly changed even during leads to high delay, which makes it unsuitable for real-
offloading if channel quality drops due to low movement or time applications. Hence, new advanced techniques enabling
fading. This can result in the situation when the offloading may very fast VM migration in order of milliseconds should be
actually increase the energy consumption and/or execution developed. However, this alternative is very challenging due
delay comparing to local computation. Hence, it is necessary to communication limits between computing nodes. Therefore,
to propose new advanced methods for the offloading decision, more realistic challenge is how to pre-migrate the computation
for instance, exploiting various prediction techniques on the in advance (e.g., based on some prediction techniques) so that
UEs mobility and channel quality during the offloading to there would be no service disruption observed by the users.
better estimate how much the offloading will cost for varying Despite of above-mentioned suggestions potentially reduc-
conditions. ing VM migration time, stand-alone VM migration may be
Besides, current papers focusing on the partial offloading unsuitable for real-time applications notwithstanding. Conse-
decision disregard the option to offload individual parts to quently, it is important to aim majority of research effort
multiple computing nodes. Multiple computing nodes enables towards a cooperation of the individual techniques for mobility
higher flexibility and increases a probability that the offloading management. In this regard, dynamic optimization and joint
to the MEC will be efficient for the UE (in terms of both consideration of all techniques (such as power control, VM
energy consumption and execution delay). Of course, a sig- migration, compression of migrated data, and/or path selec-
nificant challenge in this scenario belongs to consideration of tion) should be studied more closely in order to enhance QoE
backhaul between the MEC servers and ability to reflect their for the UEs and to optimize overall system performance for
varying load and parameters during the offloading decision. moving users.

C. Allocation of computing resources E. Traffic paradigm imposed by coexistence of offloaded data

The studies addressing the problem of an efficient allocation and conventional data
of the computing resources for the application offloaded to Current research dealing with the decision on computation
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 nication resources both for the offloaded data to the MEC
processed 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 during the fact that the offloaded data represents additional load
processing of the offloaded applications in the MEC is an on already resource starving mobile cellular networks. The
interesting research challenge to be addressed in the future. efficient scheduling of the communication resources may also
So far all the studies focusing on the allocation of computing increase the amount of data offloaded to the MEC because of
resources assume a ”flat” MEC architecture in a sense that more efficient utilization radio and backhaul communication
the MEC computing nodes are equally distributed and of the links.
same computing power. In this respect, it would be interesting Besides, the offloading reshapes conventional perception of
to consider more hierarchical placement of the computing uplink/downlink utilization as the offloading is often more
nodes within the MEC. More specifically, computing resources demanding in terms of the uplink transmission (offloading
should be distributed within the network as described in Sec- from the UE to the MEC). The reason for this is that many ap-
tion III-B3 (e.g., cluster of SCeNBs, eNBs, aggregation points plications require delivering of large files/data to the MEC for
or even at the edge of CN). A hierarchical MEC placement processing (e.g., image/video/voice recognition, file scanning,
should result in a better distribution of the computing load and etc.) while the results delivered to the UE are of significantly
a lower execution delay experienced by the users since the use lower volume. This paradigm motivates for rethinking and
of distant CC can be more easily avoided. reshaping research effort from sole downlink to the mixed
downlink and uplink in the future.
D. Mobility management
So far, the works focusing on mobility management and F. Concept validation
particularly on the VM migration consider mostly a scenario As shown in Section V, VI, and VII, the MEC concept
when only a single computing node (SCeNB or eNB) makes is analyzed and novel algorithms and proposed solutions are
computation for each UE. Hence, the challenge is how to validated typically by numerical analysis or by simulations. In
addition, majority of work assume rather simple, and some- [9] T. Nishio, R. Shinkuma, T. Takahashi, and N. B. Mandayam, “Service-
times unrealistic, scenarios for simplification of the problem. oriented heterogeneous resource sharing for optimizing service latency in
mobile cloud”, Proceedings of the first international workshop on Mobile
Although these are a good starting point in uncovering MEC cloud computing & networking, 19-26, 2013.
potentials, it is important to validate key principles and find- [10] K. Habak, M. Ammar, K. A. Harras, and E. Zegura, “Femto Clouds:
ings by means of simulations under more complex and realistic Leveraging Mobile Devices to Provide Cloud Service at the Edge”, IEEE
International Conference on Cloud Computing, 9-16, 2015.
situations and scenarios such as, e.g., in [114]-[118] where [11] F Liu, P. Shu, H. Jin, L. Ding, J. Yu, D. Niu, and B. Li, “Gearing
at least real world user mobility traces are considered for resource-poor mobile devices with powerful clouds: architectures, chal-
evaluation and proposals on VM migration. At the same time, lenges, and applications”, IEEE Wireless Communications, 20(3), 14-22,
2013.
massive trials and further experiments in emulated networks [12] I. Yaqoob, E. Ahmed, A. Gani, S. Mokhtar, M. Imran, and S. Guizani,
(like initially provided in [42]) or real networks (similar to “Mobile ad hoc cloud: A survey”, Wireless Communications and Mobile
those just recently performed by Nokia [46]) are mandatory Computing, 2016.
[13] C. Ragona, F. Granelli, C. Fiandrino, D. Kliazovich, and P. Bouvry,
to move the MEC concept closer to the reality. “Energy-Efficient Computation Offloading for Wearable Devices and
Smartphones in Mobile Cloud Computing”, IEEE Global Communica-
tions Conference (GLOBECOM), 1-6, 2015.
X. C ONCLUSION [14] W. Zhang, Y. Wen, J. Wu, and H. Li, “Toward a unified elastic computing
platform for smartphones with cloud support”, IEEE Network, 27(5), 34-
The MEC concept brings computation resources close to 40, 2013.
the UEs, i.e., to the edge of mobile network. This enables [15] F. Bonomi, R. Milito, J. Zhu, and S. Addepalli, “Fog Computing and
Its Role in the Internet of Things,” MCC workshop on Mobile cloud
to offload highly demanding computations to the MEC in computing, 13-16, 2012.
order to cope with stringent requirements of applications on [16] J. Zhu, D. S. Chan, M. S. Prabhu, P. Natarajan, H. Hu, and F.
latency (e.g., real time applications) and to reduce energy Bonomi, “Improving Web Sites Performance Using Edge Servers in Fog
consumption at the UE. Although the research on the MEC Computing Architecture,” IEEE International Symposium on Service-
Oriented System Engineering, 320-323, 2013.
gains its momentum, as reflected in this survey after all, the [17] I. Stojmenovic, S. Wen, “The Fog Computing Paradigm: Scenarios
MEC itself is still immature and highly unproved technology. and Security Issues,” Federated Conference on Computer Science and
In this regard, the MEC paradigm introduces several critical Information Systems, 1-8, 2014.
[18] I. Stojmenovic, “Fog computing: A cloud to the ground support for
challenges waiting to be addressed to the full satisfaction of all smart things and machine-to-machine networks,” Australasian Telecom-
involved parties such as mobile operators, service providers, munication Networks and Applications Conference (ATNAC), 117-122,
and users. The alpha and the omega of current research 2014.
[19] T. H. Luan, L. Gao, Y. Xiang, Z. Li, L. Sun, “Fog Com-
regarding the MEC is how to guarantee service continuity in puting: Focusing on Mobile Users at the Edge. Available at
highly dynamic scenarios. This part is lacking in terms of https://arxiv.org/abs/1502.01815, 2016.
research and is one of the blocking point to enroll the MEC [20] M. Yannuzzi, R. Milito, R. Serral-Gracia, and D. Montero, “Key
ingredients in an IoT recipe: Fog Computing, Cloud computing, and more
concept. Moreover, recent research validates solution mostly Fog Computing”, International Workshop on Computer Aided Modeling
under very simplistic scenarios and by means of simulations and Design of Communication Links and Networks (CAMAD), 325-329,
or analytical evaluations. Nevertheless, to demonstrate the 2014.
[21] A. Checko, H. L. Christiansen, Y. Yan, L. Scolari, G. Kardaras, M. S.
expected values introduced by the MEC, real tests and trials Berger, and L. Dittmann“Cloud RAN for Mobile Networks - A Tech-
under more realistic assumptions are further required. nology Overview”, IEEE Communications Surveys & Tutorials, 17(1),
405-426, First Quarter 2015.
[22] D. Kliazovich and F. Granelli, “Distributed Protocol Stacks: A Frame-
R EFERENCES work for Balancing Interoperability and Optimization”, IEEE Interna-
tional Conference on Communications (ICC) Workshop, 241-245, 2008.
[1] Hoang T. Dinh, Chonho Lee, Dusit Niyato and Ping Wang, “A survey [23] Ji. Cheng, Y. Shi, B. Bai, and W. Chen, “Computation Offloading in
of mobile cloud computing: architecture, applications, and approaches”, Cloud-RAN Based Mobile Cloud Computing System”, IEEE Interna-
Wireless Communications and Mobile Computing, 1587-1611, 13, 2013. tional Conference on Communications (ICC), 1-6, 2016.
[2] S. Barbarossa, S. Sardellitti, and P. Di Lorenzo, “Communicating while [24] Y. Ch. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young, “Mobile
Computing: Distributed mobile cloud computing over 5G heterogeneous Edge Computing A key technology towards 5G-First edition”, 2015.
networks”, IEEE Signal Processing Magazine, 31(6), 45-55, 2014. [25] Z. Sanaei, S. Abolfazli, A. Gani, and R. Buyya, “Heterogeneity in
[3] A. R. Khan, M. Othman, S. A. Madani, and S. U. Khan, “A Survey of Mobile Cloud Computing: Taxonomy and Open Challenges”, IEEE
Mobile Cloud Computing Application Models”, IEEE Communications Communications Surveys & Tutorials, 16(1), 369-392, First Quarter 2014.
Surveys & Tutorials, 16(1), 393-413, First Quarter 2014. [26] Y. Wen, X. Zhu, J. J. P. C. Rodrigues, and C. W. Chen, “Cloud Mobile
[4] M. Satyanarayanan, P. Bahl, R. Caceres, and N. Davies, “The Case for Media: Reflections and Outlook”, IEEE Transactions on Multimedia,
VM-Based Cloudlets in Mobile Computing”, IEEE Pervasive Computing, 16(4), 885-902, June 2014.
8(4), 14-23, 2009. [27] A. Ahmed, E. Ahmed, “A Survey on Mobile Edge Computing”, IEEE
[5] C. Shi, V. Lakafosis, M. H. Ammar, and E. W. Zegura, “Serendipity: International Conference on Intelligent Systems and Control (ISCO 2016),
Enabling remote computing among intermittently connected mobile de- 1-8, 2016.
vices”, ACM international symposium on Mobile Ad Hoc Networking and [28] R. Roman, J. Lopez, M. Mambo, “Mobile Edge Computing, Fog et
Computing, 145-154, 2012. al.: A Survey and Analysis of Security Threats and Challenges”, Future
[6] U. Drolia, et. al., “The Case For Mobile Edge-Clouds,” IEEE 10th Generation Computer Systems, Nov 2016.
International Conference on Ubiquitous Intelligence & Computing and [29] N. C. Luong, P. Wang, D. Niyato, W. Yonggang, and Z. Han, “Resource
IEEE 10th International Conference on Autonomic & Trusted Computing, Management in Cloud Networking Using Economic Analysis and Pricing
209-215, 2013. Models: A Survey”, IEEE Communications Surveys & Tutorials, PP(99),
[7] A. Mtibaa, A. Fahim, K. A. Harras, and M. H. Ammar, “Towards resource Jan. 2017.
sharing in mobile device clouds: Power balancing across mobile devices”, [30] E. Cuervo et al., “MAUI: Making smartphones last longer with code
ACM SIGCOMM workshop on Mobile cloud computing, 51-56, 2013. offload,” Int. Conf. Mobile Syst., Appl. Serv. (Mobysis), 4962, 2010.
[8] A. Mtibaa, K. Harras, and A. Fahim, “Towards computational offloading [31] B. G. Chun, S. Ihm, P. Maniatis, M. Naik, and A. Patti, “CloneCloud:
in mobile device clouds”, IEEE International Conference on Cloud Elastic execution between mobile device and cloud,” Eur. Conf. Comput.
Computing Technology and Science, 331-338, 2013. Syst. (Eurosys), 301314, 2011.
[32] S. Kosta, A. Aucinas, P. Hui, R. Mortier, and X Zhang, “ThinkAir: [56] S. Wang, et. al., “Mobile Micro-Cloud: Application Classification,
Dynamic resource allocation and parallel execution in the cloud for Mapping, and Deployment”, Annual Fall Meeting of ITA (AMITA), 2013.
mobile code offloading”, IEEE INFOCOM, 945-953, 2012. [57] K. Wang, M. Shen, and J. Cho, “MobiScud: A Fast Moving Personal
[33] W. Zhang, Y. Wen, and D. O. Wu, “Energy-efficient Scheduling Policy Cloud in the Mobile Network”, Workshop on All Things Cellular:
for Collaborative Execution in Mobile Cloud Computing”, IEEE INFO- Operations, Applications and Challenge, 19-24, 2015.
COM, 190-194, 2013. [58] A. Manzalini, et. al., “Towards 5G Software-Defined Ecosystems: Tech-
[34] W. Zhang, Y. Wen, and D. O. Wu, “Collaborative task execution in nical Challenges, Business Sustainability and Policy Issues”, white paper,
mobile cloud computing under a stochastic wireless channel, IEEE Trans. 2014.
Wireless Commun., 14(1), 8193, Jan. 2015. [59] T. Taleb, A. Ksentini, “Follow Me Cloud: Interworking Federated Clouds
[35] Y. Wen, W. Zhang, and H. Luo, “Energy-Optimal Mobile Application and Distributed Mobile Networks”, IEEE Network, 27(5), 12-19, 2013.
Execution: Taming Resource-Poor Mobile Devices with Cloud Clones”, [60] T. Taleb, A. Ksentini, and P. A. Frangoudis, “Follow-Me Cloud: When
IEEE INFOCOM, 2716-20, 2012. Cloud Services Follow Mobile Users”, IEEE Transactions on Cloud
[36] H. Flores, P. Hui, S. Tarkoma, Y. Li, S. Srirama, and R. Buyya, Computing, PP(99), 1-1, 2016.
“Mobile Code Offloading: From Concept to Practice and Beyond”, IEEE [61] A. Aissioui, A. Ksentini, and A. Gueroui, “An Efficient Elastic Dis-
Communication Magazine, 53(3), 80-88, 2015. tributed SDN Controller for Follow-Me Cloud”, IEEE International
[37] L. Jiao, R. Friedman, X. Fu, S. Secci, Z. Smoreda, and H. Tschofenig, Conference on Wireless and Mobile Computing, Networking and Com-
“Cloud-based Computation Offloading for Mobile Devices: State of the munications (WiMob), 876-881, 2015.
Art, Challenges and Opportunities”, Future Network & Mobile Summit, [62] J. Liu, T. Zhao, S. Zhou, Y. Cheng, and Z. Niu, “CONCERT: a cloud-
1-11, 2013. based architecture for next-generation cellular systems”, IEEE Wireless
[38] ETSI GS MEC 002: Mobile Edge Computing (MEC); Technical Re- Communications, 21(6), 14-22, Dec. 2014.
quirements V1.1.1, March 2016. [63] ETSI GS MEC 001: Mobile Edge Computing (MEC); Terminology
[39] M. T. Beck, M. Werner, S. Feld, and T. Schimper, “Mobile Edge V1.1.1, March 2016.
Computing: A Taxonomy” International Conference on Advances in [64] ETSI GS MEC 005: Mobile Edge Computing (MEC); Proof of Concept
Future Internet (AFIN), 2014. Framework V1.1.1, March 2016.
[40] N. Takahashi, H. Tanaka, and R. Kawamura, “Analysis of process [65] ETSI GS MEC 004: Mobile Edge Computing (MEC); Service Scenarios
assignment in multi-tier mobile cloud computing and application to Edge V1.1.1, March 2016.
Accelerated Web Browsing”, IEEE International Conference on Mobile [66] ETSI GS MEC 003: Mobile Edge Computing (MEC); Framework and
Cloud Computing, Services, and Engineering, 233-234, 2015. Reference Architecture V1.1.1, March 2016.
[41] Y. Zhang, H. Liu, L. Jiao, and X. Fu, “To offload or not to offload: [67] A. Ceselli, M. Premoli, and S. Secci, “Cloudlet Network Design Opti-
an efficient code partition algorithm for mobile cloud computing”, 1st mization”, IFIP Networking, 1-9, 2015.
International Conference on Cloud Networking (CLOUDNET), 80-86, [68] A. Ceselli, M. Premoli, and S. Secci, “Mobile Edge Cloud Network
2012. Design Optimization”, IEEE/ACM Transactions on Networking, PP(99),
[42] J. Dolezal, Z. Becvar, and T. Zeman, “Performance Evaluation of 2017.
Computation Offloading from Mobile Device to the Edge of Mobile [69] D. Kreutz, F. M. V. Ramos, P. E. Verissimo, C. E. Rothenberg, S.
Network”, IEEE Conference on Standards for Communications and Azodolmolky, and S. Uhlig, “Software-Defined Networking: A Compre-
Networking (CSCN), 1-7, 2016. hensive Survey”, Proceedings of the IEEE, 103(1), 14-76, Jan. 2015.
[43] O. Salman, I. Elhajj, A. Kayssi, and A. Chehab, “Edge Computing [70] N. A. Jagadeesan and B. Krishnamachari, “Software-Defined Network-
Enabling the Internet of Things”, IEEE 2nd World Forum on Internet ing Paradigms in Wireless Networks: A Survey”, ACM Computing
of Things (WF-IoT), 603-608, 2015. Surveys, 47(2), Jan. 2015.
[44] S. Abdelwahab, B. Hamdaoui, M. Guizani, and T. Znati, “REPLISOM: [71] X. Jin, L. E. Li, L. Vanbever, and J. Rexford, “SoftCell: scalable and
Disciplined Tiny Memory Replication for Massive IoT Devices in LTE flexible cellular core network architecture”, ACM conference on Emerging
Edge Cloud”, IEEE Internet of Things Journal, 3(3), 327-338, 2016. networking experiments and technologies (CoNEXT), 163-174, 2013.
[45] X. Sun and N. Ansari, “EdgeIoT: Mobile Edge Computing for the [72] M. Deng, H. Tian, and B. Fan, “Fine-granularity Based Application
Internet of Things”, IEEE Communications Magazine, 54(12), 22-29, Offloading Policy in Small Cell Cloud-enhanced Networks”, IEEE In-
2016. ternational Conference on Communications Workshops (ICC), 638-643,
[46] NOKIA: Multi-access Edge Computing, [online] https://networks. 2016.
nokia.com/solutions/multi-access-edge-computing. [73] S. E. Mahmoodi, R. N. Uma, and K. P. Subbalakshmi, “Optimal joint
[47] NOKIA: Mobile Edge Computing, [online] scheduling and cloud offloading for mobile applications,” IEEE Trans.
http://resources.alcatel-lucent.com/asset/200546. Cloud Comput., PP(99), 2016.
[48] FP7 European Project, Distributed computing, storage and radio resource [74] J. Liu, Y. Mao, J. Zhang, and K. B. Letaief, “Delay-Optimal Com-
allocation over cooperative femtocells (TROPIC). [Online]. Available: putation Task Scheduling for Mobile-Edge Computing Systems”, IEEE
http://www.ict-tropic.eu/, 2012. International Symposium on Information Theory (ISIT), 1451-55, 2016.
[49] H2020 European Project, Small cEllS coordinAtion for [75] Y. Mao, J. Zhang, K. B. Letaief, “Dynamic Computation Offloading for
Multi-tenancy and Edge services (SESAM), Available: Mobile-Edge Computing with Energy Harvesting Devices”, IEEE Journal
http://www.sesame-h2020-5g-ppp.eu/, 2015. on Selected Areas in Communications, 34(12), 3590-3605, 2016.
[50] I. Giannoulakis, et. al., “The emergence of operator-neutral small cells as [76] W. Zhang, Y. Wen, K. Guan, D. Kilper, H. Luo, and D. O. Wu, “Energy-
a strong case for cloud computing at the mobile edge”, Trans. Emerging Optimal Mobile Cloud Computing under Stochastic Wireless Channel”,
Tel. Tech., 27, 115259, 2016. IEEE Transactions on Wireless Communications, 12(9), 4569-81, 2013.
[51] M. Chiosi, et. al. “Network Functions Virtualisation: An Introduction, [77] S. Ulukus, A. Yener, E. Erkip, O. Simeone, M. Zorzi, P. Grover, and
Benefits, Enablers, Challenges & Call for Action”, Introductory white K. Huang, “Energy Harvesting Wireless Communications: A Review of
paper, 2012. Recent Advances”, IEEE Journal on Selected Areas in Communications,
[52] ETSI GS NFV 002: Architectural Framework, V1.1.1, Oct. 2013. 33(3), 360-381, 2015.
[53] F. Lobillo, et. al., “An architecture for mobile computation offloading [78] M. Kamoun, W. Labidi, and M. Sarkiss, “Joint resource allocation
on cloud-enabled LTE small cells”, Workshop on Cloud Technologies and and offloading strategies in cloud enabled cellular networks”, IEEE
Energy Efficiency in Mobile Communication Networks (IEEE WCNCW), International Conference on Communications (ICC), 5529-34, 2015.
1-6, 2014. [79] W. Labidi, M. Sarkiss, and M. Kamoun, “Energy-Optimal Resource
[54] M. A. Puente, Z. Becvar, M. Rohlik, F. Lobillo, and E. C. Strinati, Scheduling and Computation Offloading in Small Cell Networks”, In-
“A Seamless Integration of Computationally-Enhanced Base Stations into ternational Conference on Telecommunications (ICT), 313-318, 2015.
Mobile Networks towards 5G”, IEEE Vehicular Technology Conference [80] W. Labidi, M. Sarkiss, and M. Kamoun, “Joint Multi-user Resource
(IEEE VTC-Spring 2015) workshops, 1-5, 2015. Scheduling and Computation Offloading in Small Cell Networks”, IEEE
[55] Z. Becvar, M. Rohlik, P. Mach, M. Vondra, T. Vanek, M.A. Puente, F. International Conference onWireless and Mobile Computing, Networking
Lobillo, “Distributed Architecture of 5G Mobile Networks for Efficient and Communications (WiMob), 794-801, 2015.
Computation Management in Mobile Edge Computing”, Chapter in [81] S. Barbarossa, S. Sardellitti, and P. Di Lorenzo, “Joint allocation of
5G Radio Access Network (RAN) - Centralized RAN, Cloud-RAN and computation and communication resources in multiuser mobile cloud
Virtualization of Small Cells (H. Venkataraman, R. Trestian, eds.), Taylor computing”, IEEE Workshop on Signal Processing Advances in Wireless
and Francis Group, USA, March 2017. Communications (SPAWC), 26-30, 2013.
[82] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint Optimization of Radio tional Symposium on Personal, Indoor, and Mobile Radio Communication
and Computational Resources for Multicell Mobile Cloud Computing”, (PIMRC), 1474-79, 2014.
IEEE International Workshop on Signal Processing Advances in Wireless [104] J. Oueis, E. Calvanese Strinati, S. Sardellitti, and S. Barbarossa, “Small
Communications (SPAWC), 354-358, 2014. Cell Clustering for Efficient Distributed Fog Computing: A Multi-User
[83] S. Sardellitti, S. Barbarossa, G. Scutari, “Distributed Mobile Cloud Case”, IEEE Vehicular Technology Conference (VTC Fall), 1-5, 2015.
Computing: Joint Optimization of Radio and Computational Resources”, [105] J. Oueis, E. Calvanese Strinati, and S. Barbarossa, “The Fog Balancing:
IEEE Globecom Workshops (GC Wkshps), 1505-10, 2014. Load Distribution for Small Cell Cloud Computing”, IEEE 81st Vehicular
[84] K. Zhang, Y. Mao, S. Leng, Q. Zhao, L. Li, X. Peng, L. Pan, S. Technology Conference (VTC Spring) 1-6, 2015.
Maharjan, and Y. Zhang, “Energy-Efficient Offloading for Mobile Edge [106] M. Vondra and Z. Becvar, “QoS-ensuring Distribution of Computation
Computing in 5G Heterogeneous Networks”, IEEE Access, 4, 5896-5907, Load among Cloud-enabled Small Cells”, IEEE International Conference
2016. on Cloud Networking (CloudNet), 197-203, 2014.
[85] X. Chen, L. Jiao, W. Li, and X. Fu, “Efficient Multi-User Computation [107] S. Wang, M. Zafer, and K. K. Leung, “Online Placement of Multi-
Offloading for Mobile-Edge Cloud Computing”, IEEE/ACM Transactions Component Applications in Edge Computing Environments”, IEEE Ac-
on Networking, 24(5), 2795-2808, 2016. cess, PP(99), 2017.
[86] M.-H. Chen, B. Liang, and M. Dong, “A Semidefinite Relaxation [108] P. Mach and Z. Becvar, “Cloud-aware power control for cloud-enabled
Approach to Mobile Cloud Offloading with Computing Access Point”, small cells”, IEEE Globecom Workshops (GC Wkshps), 1038-43, 2014.
IEEE International Workshop on Signal Processing Advances in Wireless [109] P. Mach and Z. Becvar, “Cloud-aware power control for real-time appli-
Communications (SPAWC), 186-190, 2015. cation offloading in mobile edge computing”, Transactions on Emerging
[87] M.-H. Chen, M. Dong, and B. Liang, “Joint offloading decision and Telecommunications Technologies, 2016.
resource allocation for mobile cloud with computing access point”, IEEE [110] T. Taleb and A. Ksentini, “An Analytical Model for Follow Me
International Conference on Acoustics, Speech and Signal Processing Cloud”, IEEE Global Communications Conference (GLOBECOM), 1291-
(ICASSP), 3516-20, 2016. 96, 2013.
[88] S. Cao, X. Tao,Y. Hou, and Q. Cui, “An Energy-Optimal Offloading [111] A. Ksentini, T. Taleb, and M. Chen, “A Markov Decision Process-based
Algorithm of Mobile Computing Based on HetNets”, International Con- Service Migration Procedure for Follow Me Cloud”, IEEE International
ference on Connected Vehicles and Expo (ICCVE), 254-258, 2015. Conference on Communications (ICC), 1350-54, 2014
[89] J. Kennedy and R. C. Eberhart, “A Discrete Binary Version of the [112] X. Sun and N. Ansari, “PRIMAL: PRofIt Maximization Avatar pLace-
Particle Swarm Algorithm”, IEEE International Conference on Systems, ment for Mobile Edge Computing”, IEEE International Conference on
Man, and Cybernetics, 4104-08, 1997. Communications (ICC), 1-6, 2016.
[90] Y. Zhao, S. Zhou, T. Zhao, and Z. Niu, “Energy-Efficient Task Offload- [113] S. Wang, R. Urgaonkar, T. He, M. Zafer, K. Chan, and K. K. Leung,
ing for Multiuser Mobile Cloud Computing”, IEEE/CIC International “Mobility-Induced Service Migration in Mobile Micro-Clouds”, IEEE
Conference on Communications in China (ICCC), 1-5, 2015. Military Communications Conference, 835-840, 2014.
[91] C. You and K. Huang, “Multiuser Resource Allocation for Mobile- [114] S. Wang, R. Urgaonkar, M. Zafer, T. He, K. Chan, and K. K. Leung,
Edge Computation Offloading”, IEEE Global Communication Conference “Dynamic Service Migration in Mobile Edge-Clouds”, IFIP Networking
(GLOBECOM), 1-6, 2016. Conference (IFIP Networking), 1-9, 2015.
[115] A. Nadembega, A. S. Hafid, and R. Brisebois, “Mobility Prediction
[92] C. You, K. Huang, H. Chae, and B.-H. Kim, “Energy-Efficient Resource
Model-based Service Migration Procedure for Follow Me Cloud to sup-
Allocation for Mobile-Edge Computation Offloading”, IEEE Transactions
port QoS and QoE”, IEEE International Conference on Communications
on Wireless Communications, 16(3), 1397-1411, 2017.
(ICC), 1-6, 2016.
[93] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li, “Mobile-Edge
[116] S. Wang, R. Urgaonkar, K. Chan, T. He, M. Zafer, and K. K. Leung,
Computing: Partial Computation Offloading Using Dynamic Voltage
“Dynamic Service Placement for Mobile Micro-Clouds with Predicted
Scaling”, IEEE Transactions on Communications, 64(10), pp. 4268-82,
Future Costs”, IEEE International Conference on Communications (ICC),
2016.
5504-10, 2015.
[94] O. Munoz, A. Pascual-Iserte, and J. Vidal, “Joint Allocation of Radio and [117] S. Wang, R. Urgaonkar, T. He, K. Chan, M. Zafer, and K. K. Leung,
Computational Resources in Wireless Application Offloading”, Future “Dynamic Service Placement for Mobile Micro-Clouds with Predicted
Network and Mobile Summit, 1-10, 2013. Future Costs”, IEEE Transactions on Parallel and Distributed Systems,
[95] O. Munoz, A. Pascual-Iserte, and J. Vidal, “Optimization of Radio and PP(99), 2016.
Computational Resources for Energy Efficiency in Latency-Constrained [118] R. Urgaonkar, S. Wang, T. He, M. Zafer, K. Chan, and K. K. Leung,
Application Offloading”, IEEE Transactions on Vehicular Technology, “Dynamic service migration and workload scheduling in edge-clouds”,
64(10), 4738-55, 2015. Performance Evaluations, 91(2015), 205-228, 2015.
[96] O. Munoz, A. Pascual-Iserte, J. Vidal, and M. Molina, “Energy-Latency [119] K. Ha, Y. Abe, Z. Chen, W. Hu, B. Amos, P. Pillai, and M. Satya-
Trade-off for Multiuser Wireless Computation Offloading”, IEEE Wireless narayanan, “Adaptive VM Handoff Across Cloudlets”, Technical Report
Communications and Networking Conference Workshops (WCNCW), 29- CMU-CS-15-113, Computer Science Department, June 2015.
33, 2014. [120] S. Secci, P. Raad, and P. Gallard, “Linking Virtual Machine Mobility to
[97] Y. Mao, J. Zhang, S.H. Song, and K. B. Letaief, “Power-Delay Tradeoff User Mobility”, IEEE Transactions on Network and Service Management,
in Multi-User Mobile-Edge Computing Systems”, IEEE Global Commu- 13(4), pp. 927-940, 2016.
nications Conference (GLOBECOM), 1-6, Dec. 2016. [121] D. Farinacci, V. Fuller, D. Meyer, and D. Lewis, “The locator/ID
[98] T. Zhao, S. Zhou, X. Guo, Y. Zhao, and Z. Niu, “A Cooperative separation protocol (LISP),” Internet Eng. Task Force, Fremont, CA, USA,
Scheduling Scheme of Local Cloud and Internet Cloud for Delay-Aware RFC 6830, 2013.
Mobile Cloud Computing”, IEEE Globecom Workshops (GC Wkshps), [122] Z. Becvar, J. Plachy, and P. Mach, “Path Selection Using Handover in
1-6, 2015. Mobile Networks with Cloud-enabled Small Cells”, IEEE International
[99] X. Guo, R. Singh, T. Zhao, and Z. Niu, “An Index Based Task Symposium on Personal, Indoor and Mobile Radio Communications
Assignment Policy for Achieving Optimal Power-Delay Tradeoff in Edge (PIMRC), 1480-85, 2014.
Cloud Systems”, IEEE International Conference on Communications [123] J. Plachy, Z. Becvar, and P. Mach, “Path Selection Enabling User
(ICC), 1-7, 2016. Mobility and Efficient Distribution of Data for Computation at the Edge
[100] V. Di Valerio and F. Lo Presti, “Optimal Virtual Machines Allocation of Mobile Network”, Computer Networks, 108, 357-370, 2016.
in Mobile Femto-cloud Computing: an MDP Approach”, IEEE Wireless [124] J. Plachy, Z. Becvar, and E. Calvanese Strinati, “Dynamic Resource
Communications and Networking Conference Workshops (WCNCW), 7- Allocation Exploiting Mobility Prediction in Mobile Edge Computing”,
11, 2014. IEEE International Symposium on Personal, Indoor and Mobile Radio
[101] S. M. S. Tanzil, O. N. Gharehshiran, and V. Krishnamurthy, “Femto- Communications (PIMRC), 1-6, 2016.
Cloud Formation: A Coalitional Game-Theoretic Approach”, IEEE
Global Communications Conference (GLOBECOM), 1-6, 2015.
[102] J. Oueis, E. Calvanese-Strinati, A. De Domenico, and S. Barbarossa,
“On the Impact of Backhaul Network on Distributed Cloud Computing”,
IEEE Wireless Communications and Networking Conference Workshops
(WCNCW), 12-17, 2014.
[103] J. Oueis, E. Calvanese Strinati, and S. Barbarossa, “Small Cell Clus-
tering for Efficient Distributed Cloud Computing”, IEEE Annual Interna-