A Survey on Mobile Edge Computing: The Communication Perspective
Yuyi Mao, Changsheng You, Jun Zhang, Kaibin Huang, Khaled B. Letaief
11 A Survey on Mobile Edge Computing: TheCommunication Perspective
Yuyi Mao, Changsheng You, Jun Zhang, Kaibin Huang, and Khaled B. Letaief
Abstract —Driven by the visions of Internet of Things and5G communications, recent years have seen a paradigm shift inmobile computing, from the centralized Mobile Cloud Computingtowards
Mobile Edge Computing (MEC). The main feature ofMEC is to push mobile computing, network control and storageto the network edges (e.g., base stations and access points) so asto enable computation-intensive and latency-critical applicationsat the resource-limited mobile devices. MEC promises dramaticreduction in latency and mobile energy consumption, tacklingthe key challenges for materializing 5G vision. The promisedgains of MEC have motivated extensive efforts in both academiaand industry on developing the technology. A main thrust ofMEC research is to seamlessly merge the two disciplines ofwireless communications and mobile computing, resulting in awide-range of new designs ranging from techniques for compu-tation offloading to network architectures. This paper provides acomprehensive survey of the state-of-the-art MEC research witha focus on joint radio-and-computational resource management.We also discusse a set of issues, challenges and future researchdirections for MEC research, including MEC system deployment,cache-enabled MEC, mobility management for MEC, greenMEC, as well as privacy-aware MEC. Advancements in thesedirections will facilitate the transformation of MEC from theoryto practice. Finally, we introduce recent standardization effortson MEC as well as some typical MEC application scenarios.
Index Terms —Mobile edge computing, fog computing, mobilecloud computing, computation offloading, resource management,green computing.
I. I
NTRODUCTION
The last decade has seen Cloud Computing emerging as anew paradigm of computing. Its vision is the centralization ofcomputing, storage and network management in the Clouds,referring to data centers, backbone IP networks and cellularcore networks [1], [2]. The vast resources available in theClouds can then be leveraged to deliver elastic computingpower and storage to support resource-constrained end-userdevices. Cloud Computing has been driving the rapid growthof many Internet companies. For example, the Cloud businesshas risen to be the most profitable sector for Amazon [3], andDropbox’s success depended highly on the Cloud service ofAmazon.However, in recent years, a new trend in computing is hap-pening with the function of Clouds being increasingly movingtowards the network edges [4]. It is estimated that tens of bil-lions of Edge devices will be deployed in the near future, and
Y. Mao, J. Zhang and K. B. Letaief are with the Dept. of Elec-tronic and Computer Engineering, The Hong Kong University of Scienceand Technology, Hong Kong (Email: [email protected], [email protected],[email protected]). K. B. Letaief is also affiliated with Hamad bin KhalifaUniversity, Doha, Qatar.C. You and K. Huang are with the Dept. of Electrical and Elec-tronic Engineering, The University of Hong Kong, Hong Kong (Email:[email protected], [email protected]). their processor speeds are growing exponentially, followingMoore’s Law. Harvesting the vast amount of the idle compu-tation power and storage space distributed at the network edgescan yield sufficient capacities for performing computation-intensive and latency-critical tasks at mobile devices. Thisparadigm is called
Mobile Edge Computing (MEC) [5]. Whilelong propagation delays remain a key drawback for CloudComputing, MEC, with the proximate access, is widely agreedto be a key technology for realizing various visions for next-generation Internet, such as Tactile Internet (with millisecond-scale reaction time) [6],
Internet of Things (IoT) [7], andInternet of Me [8]. Presently, researchers from both academiaand industry have been actively promoting MEC technologyby pursuing the fusion of techniques and theories from bothdisciplines of mobile computing and wireless communications .This paper aims at providing a survey of key research progressin this young field from the communication perspective. Weshall also present a research outlook containing an ensembleof promising research directions for MEC.
A. Mobile Computing for 5G: From Clouds to Edges
In the past decade, the popularity of mobile devices and theexponential growth of mobile Internet traffic have been drivingthe tremendous advancements in wireless communications andnetworking. In particular, the breakthroughs in small-cell net-works, multi-antenna, and millimeter-wave communicationspromise to provide users gigabit wireless access in next-generation systems [9]. The high-rate and highly-reliable airinterface allows to run computing services of mobile devicesat the remote cloud data center, resulting in the research areacalled
Mobile Cloud Computing (MCC). However, there isan inherent limitation of MCC, namely, the long propagationdistance from the end user to the remote cloud center, whichwill result in excessively long latency for mobile applications.MCC is thus not adequate for a wide-range of emergingmobile applications that are latency-critical. Presently, newnetwork architectures are being designed to better integrate theconcept of Cloud Computing into mobile networks, as will bediscussed in the latter part of this article.In 5G wireless systems, ultra-dense edge devices, includingsmall-cell base stations (BSs), wireless access points (APs),laptops, tablets, and smartphones, will be deployed, eachhaving a computation capacity comparable with that of acomputer server a decade ago. As such, a large population ofdevices will be idle at every time instant. It will, in particular,be harvesting enormous computation and storage resourcesavailable at the network edges, which will be sufficient toenable ubiquitous mobile computing. In a nutshell, the main a r X i v : . [ c s . I T ] J un target of wireless systems, from 1G to 4G, is the pursuit ofincreasingly higher wireless speeds to support the transitionfrom voice-centric to multimedia-centric traffic. As wirelessspeeds approach the wireline counterparts, the mission of 5Gis different and much more complex, namely to support theexplosive evolution of ICT and Internet. In terms of func-tions, 5G systems will support communications, computing,control and content delivery (4C). In terms of applications,a wide-range of new applications and services for 5G areemerging, such as real-time online gaming, virtual reality (VR) and ultra-high-definition (UHD) video streaming, whichrequire unprecedented high access speed and low latency.The past decade also saw the take-off of different visions ofnext-generation Internet including IoT, Tactile Internet (withmillisecond latency), Internet-of-Me, and social networks. Inparticular, it was predicted by Cisco that about 50 billionIoT devices (e.g., sensors and wearable devices) will beadded to the Internet by 2020, most of which have limitedresources for computing, communication and storage, andhave to rely on Clouds or edge devices for enhancing theircapabilities [10]. It is now widely agreed that relying onlyon Cloud Computing is inadequate to realize the ambitiousmillisecond-scale latency for computing and communicationin 5G. Furthermore, the data exchange between end users andremote Clouds will allow the data tsunami to saturate andbring down the backhaul networks. This makes it essential tosupplement Cloud Computing with MEC that pushes traffic,computing and network functions towards the network edges.This is also aligned with a key characteristic of next-generationnetworks that information is increasingly generated locallyand consumed locally , which arises from the booming ofapplications in IoT, social networks and content delivery [4].The concept of MEC was firstly proposed by the Euro-pean Telecommunications Standard Institute (ETSI) in 2014,and was defined as a new platform that “ provides IT andcloud-computing capabilities within the Radio Access Network(RAN) in close proximity to mobile subscribers ” [5]. Theoriginal definition of MEC refers to the use of BSs foroffloading computation tasks from mobile devices. Recently,the concept of
Fog Computing has been proposed by Cisco as ageneralized form of MEC where the definition of edge devicesgets broader, ranging from smartphones to set-top boxes [11].This led to the emergence of a new research area calledFog Computing and Networking [4], [12], [13]. However,the areas of Fog Computing and MEC are overlapping andthe terminologies are frequently used interchangeably. In thispaper, we focus on MEC but many technologies discussed arealso applicable to Fog Computing.MEC is implemented based on a virtualized platform thatleverages recent advancements in network functions virtualiza-tion (NFV), information-centric networks (ICN) and software-defined networks (SDN). Specifically, NFV enables a singleedge device to provide computing services to multiple mobiledevices by creating multiple virtual machines (VMs) for si-multaneously performing different tasks or operating different The VM is a virtual computer mapped to the physical machine’s hard-wares, providing virtual CPU, memory, hard drive, network interface, andother devices [14].
Image (cid:3)
AquisitionImage (cid:3)
Aquisition Face (cid:3)
DetectionFace (cid:3)
Detection Pre (cid:882) processingClassificationClassification Feature (cid:3)
ExtractionFeature (cid:3)
Extraction
Database (cid:3) (Training (cid:3)
Samples)
Digital (cid:3)
Image Face (cid:3)
Image Normalized (cid:3)
Face (cid:3)
Image Feature (cid:3)
VectorClassification (cid:3)
Results
MEC (cid:3)
Server
Fig. 1. Main computation components in a face recognition application [17].
Video (cid:3)
SourceVideo (cid:3)
Source
Number (cid:3) of (cid:3) Features
RendererRendererTrackerTracker MapperMapper
Resolution Image (cid:3)
Format
MEC (cid:3)
Server
Object (cid:3)(cid:3)
RecognizerObject (cid:3)(cid:3)
Recognizer
Fig. 2. Main computation components in an AR application [18]. network functions [15]. On the other hand, ICN provides analternative end-to-end service recognition paradigm for MEC,shifting from a host-centric to an information-centric onefor implementing context-aware computing. Last, SDN allowsMEC network administrators to manage services via functionabstraction, achieving scalable and dynamic computing [16].A main focus of MEC research is to develop these generalnetwork technologies so that they can be implemented at thenetwork edges.There is an increasing number of emerging mobile ap-plications that will benefit from MEC, by offloading theircomputation-intensive tasks to the MEC servers for cloudexecution. In the following, we will provide two examples toillustrate the basic principles of MEC. One is the face recog-nition application as shown in Fig. 1, which typically con-sists of five main computation components, including imageacquisition, face detection, pre-processing, feature extraction,and classification [17]. While the image acquisition componentneeds to be executed at the mobile device for supporting theuser interface, the other components could be offloaded forcloud processing, which contain complex computation suchas signal processing and machine learning (ML) algorithms.Another popular stream of applications that can leverage therich resources at the network edges are augmented reality (AR) applications, which are able to combine the computer-generated data with physical reality. AR applications as shownin Fig. 2 have five critical components [18]–[20], namely,the video source (which obtains raw video frames from themobile camera), a tracker (which tracks the position of theuser), a mapper (which builds a model of the environment),an object recognizer (which identifies known objects in the environment), and a renderer (which prepares the processedframe for display). Among these components, the video sourceand renderer should be executed locally, while the mostcomputation-intensive components, i.e., the tracker, mapperand object recognizer, can be offloaded for cloud execution.In this way, mobile users can enjoy various benefits fromMEC such as latency reduction and energy savings, as willbe elaborated in the next subsection.
B. Mobile Edge Computing Versus Mobile Cloud Computing
As shown in Table I, there exist significant disparitiesbetween MEC and MCC systems in terms of computing server,distance to end users and typical latency, etc. Compared withMCC, MEC has the advantages of achieving lower latency,saving energy for mobile devices, supporting context-awarecomputing, and enhancing privacy and security for mobileapplications. These advantages are briefly described throughsome examples and applications in the following.
Low Latency:
The latency for a mobile service is theaggregation of three components: propagation , computation ,and communication latency, depending on the propagation dis-tance, computation capacity, and data rate, respectively. First,the information-propagation distances for MEC are typicallytens-of-meters for the cases of dense small-cell networks or device-to-device (D2D) transmissions, and typically no longerthan 1km for general cases. In contrast, Cloud Computingrequires transmissions from end users to nodes in core net-works or data centers with distances ranging from tens ofkilometers to that across continents. This results in muchshorter propagation delay for MEC than that for MCC. Second,MCC requires the information to pass through several net-works including the radio-access network, backhaul networkand Internet, where traffic control, routing and other network-management operations can contribute to excessive delay. Withthe communication constrained at the network edges, MECis free from these issues. Last, for the computation latency,a Cloud has a massive computation power that is severalorders of magnitude higher than that of an edge device (e.g.,a BS). However, the Cloud has to be shared by a much largernumber of users than an edge device, reducing their gap in thecomputation latency. Furthermore, a modern BS is powerfulenough for running highly sophisticated computing programs.For instance, the edge cloud at a BS has 10 -10 times highercomputation capability than the minimum requirement (e.g.,a CPU over 3.3GHz, 8GB RAM, 70GB storage space) forrunning the Call-of-Duty 13, a popular shooter game . Ingeneral, experiments have shown that the total latency forMCC is in the range of 30-100ms [31]. This is unacceptablefor many latency-critical mobile applications such as real-time online gaming, virtual sports and autonomous driving,which may require tactile speed with latency approaching1ms [37]. In contrast, with short propagation distances andsimple protocols, MEC has the potential of realizing tactile-level latency for latency-critical 5G applications. Mobile Energy Savings:
Due to their compact forms,IoT devices have limited energy storage but are expected to cooperate and perform sophisticated tasks such as surveillance,crowd-sensing and health monitoring [38]. Powering the tensof billions of IoT devices remains a key challenge for de-signing IoT given that frequent battery recharging/replacementis impractical if not impossible. By effectively supporting computation offloading , MEC stands out as a promising solu-tion for prolonging battery lives of IoT devices. Specifically,computation-intensive tasks can be offloaded from IoT devicesto edge devices so as to reduce their energy consumption.Significant energy savings by computation offloading havebeen demonstrated in experiments, e.g., the completion of upto 44-time more computation load for a multimedia application eyeDentify [39] or the increase of battery life by 30-50% fordifferent AR applications [40]. Context-Awareness:
Another key feature that differentiatesMEC from MCC is the ability of an MEC server for lever-aging the proximity of edge devices to end users to tracktheir real-time information such as behaviors, locations, andenvironments. Inference based on such information allows thedelivery of context-aware services to end users [41]–[43]. Forinstance, the museum video guide, an AR application, canpredict users’ interests based on their locations in the museumto automatically deliver contents related to e.g., artworks andantiques [44]. Another example is the CTrack system that usesthe BS fingerprints to track and predict the trajectories of alarge number of users for the purposes of traffic monitoring,navigation and routing, and personalized trip management[45].
Privacy/Security Enhancement:
The capability of enhanc-ing the privacy and security of mobile applications is alsoan attractive benefit brought by MEC compared to MCC.In MCC systems, the Cloud Computing platforms are theremote public large data centers, such as the Amazon EC2and Microsoft Azure, which are susceptible to attacks due totheir high concentration of information resources of users. Inaddition, the ownership and management of users’ data areseparated in MCC, which shall cause the issues of privatedata leakage and loss [46]. The use of proximate edge serversprovides a promising solution to circumvent these problems.On one hand, due to the distributed deployment, small-scalenature, and the less concentration of valuable information,MEC servers are much less likely to become the target of asecurity attack. Second, many MEC servers could be private-owned cloudlets, which shall ease the concern of informa-tion leakage. Applications that require sensitive informationexchange between end users and servers would benefit fromMEC. For instance, the enterprise deployment of MEC couldhelp avoid uploading restricted data and material to remotedata centers, as the enterprise administrator itself manages theauthorization, access control, and classifies different levels ofservice requests without the need of an external unit [47].
C. Paper Motivation and Outline
MEC has emerged as a key enabling technology for realiz-ing the IoT and 5G visions [15], [48], [49]. MEC research liesat the intersection of mobile computing and wireless commu-nications, where the existence of many research opportunities
TABLE IC
OMPARISON OF
MEC
AND
MCC S
YSTEMS .MEC MCCServer hardware Small-scale data centers Large-scale data centers (each containswith moderate resources [5], [21] a large number of highly-capable servers) [22], [23]Server location Co-locate with wireless gateways, Installed at dedicated buildings,WiFi routers, and LTE BSs [5] with size of several football fields [24], [25]Deployment Densely deployed by telecom operators, Deployed by IT companies, e.g., GoogleMEC vendors, enterprises, and and Amazon, at a few locationshome users. Require lightweight over the world. Require sophisticatedconfiguration and planning [5] configuration and planning [22]Distance to end users Small Large(tens to hundreds of meters) [15] (may across the country border) [26]Backhaul usage Infrequent use Frequent useAlleviate congestion [27] Likely to cause congestion [27]System management Hierarchical control Centralized control [28](centralized/distributed) [28]Supportable latency Less than tens of milliseconds [15], [29] Larger than 100 milliseconds [30], [31]Applications Latency-critical and computation-intensive Latency-tolerant and computation-intensiveapplications, e.g., AR, automatic driving, applications, e.g., online social networking,and interactive online gaming [5], [32]. and mobile commerce/health/learning [33]–[36]. has resulted in a highly active area. In recent years, researchersfrom both academia and industry have investigated a wide-range of issues related to MEC, including system and networkmodeling, optimal control, multiuser resource allocation, im-plementation and standardization. Subsequently, several surveyarticles have been published to provide overviews of theMEC area with different focuses, including system models,architectures, enabling techniques, applications, edge caching,edge computation offloading, and connections with IoT and 5G[27], [28], [50]–[56]. Their themes are summarized as follows.An overview of MEC platforms is presented in [50] wheredifferent existing MEC frameworks, architectures, and theirapplication scenarios, including FemtoClouds, REPLISM, andME-VOLTE, are discussed. The survey of [51] focuses onthe enabling techniques in MEC such as cloud computing,VM, NFV, SDN that allow the flexible control and multi-tenancy support. In [52], the authors categorize diverse MECapplications, service models, deployment scenarios, as wellas network architectures. The survey in [53] presents a tax-onomy for MEC applications and identifies potential direc-tions for research and development, such as content scaling,local connectivity, augmentation, and data aggregation andanalytics. In [28], emerging techniques of edge computing,caching, and communications (3C) in MEC are surveyed,showing the convergence of 3C. Besides, key enablers of MECsuch as cloud technology, SDN/NFV, and smart devices arealso discussed. The survey in [54] focuses on three criticaldesign problems in computation offloading for MEC, namely,the offloading decision, computation resource allocation, andmobility management. In addition, the role of MEC in IoT,i.e., creating new IoT services, is highlighted in [55] throughMEC deployment examples with reference to IoT use cases.Several attractive use scenarios of MEC in 5G networksare also introduced in [27], ranging from mobile-edge or-chestration, collaborative caching and processing, and multi-layer interference cancellation. Furthermore, potential businessopportunities related to MEC are discussed in [56] from theperspectives of application developers, service providers, andnetwork equipment vendors. In view of prior work, there still lacks a systematic survey article providing comprehensiveand concrete discussions on specific MEC research resultswith a deep integration of mobile computing and wirelesscommunications, which motivates the current work. This paperdiffers from existing surveys on MEC in the following aspects.First, the current survey summarizes existing models of com-puting and communications in MEC to facilitate theoreticalanalysis and provide a quick reference for both researchersand practitioners. Next, we present a comprehensive literaturereview on joint radio-and-computational resource allocationfor MEC, which is the central theme of the current paper. Theliterature review in our paper shall be a valuable addition to theexisting survey literature on MEC, which can benefit readersfrom the research community in building up a systematicunderstanding of the state-of-the-art resource managementtechniques for MEC systems. Furthermore, we identify anddiscuss several research challenges and opportunities in MECfrom the communication perspective, for which potential so-lutions are elaborated. In addition, to bridge the gap betweentheoretical research and real implementation of MEC, recentstandardization efforts and use scenarios of MEC will then beintroduced.This paper is organized as follows. In Section II, we summa-rize the basic MEC models, comprising models of computationtasks, communications, mobile devices and MEC servers,based on which the models of MEC latency and energy con-sumption are developed. Next, a comprehensive review is pre-sented in Section III, focusing on the research of joint radio-and-computational resource management for different typesof MEC systems, including single-user, multiuser systems aswell as multi-server MEC. Subsequently, a set of key researchissues and future directions are discussed in Section III-D3including 1) deployment of MEC systems, 2) cache-enabledMEC, 3) mobility management for MEC, 4) green MEC,and 5) security-and-privacy issues in MEC. Specifically, weanalyze the design challenges for each research problemand provide several potential research approaches. Last, theMEC standardization efforts and applications are reviewed anddiscussed in Section V, followed by concluding remarks in
TABLE IIS
UMMARY OF I MPORTANT A CRONYMS . Acronym Definition Acronym Definition
AF application function MEC mobile edge computingAR augmented reality ML machine learningAP access point mMTC massive machine type communicationBS base station NEF network exposure functionCAPEX capital expenditure NFC near-filed communicationsC-RAN cloud radio access network NFV network functions virtualizationCSI channel-state information OFDMA orthogonal frequency division multiple accessDAG directed acyclic graph PCF policy control functionDCN data-center network PMR peak-to-mean ratioDNS domain name system PoC proof of conceptDP dynamic programming QoS quality of serviceDPP determinantal point process RAM random access memoryDVFS dynamic frequency and voltage scaling RAN radio access networkD2D device-to-device RFID radio frequency identificationEH energy harvesting RNIS radio network information serviceseMBB enhanced mobile broadband SDN software-defined networksESI energy side information SINR signal-to-interference-plus-noise ratioETSI European Telecommunications Standard Institute TOF traffic offloading functionGLB geographical load balancing UE user equipmentHet-MEC heterogeneous MEC UHD ultra-high-definitionHetNets heterogeneous networks UPF user plane functionHPPP homogeneous Poisson point process UPS uninterrupted power supplyIaaS Infrastructure as a Service URLLC ultra-reliable and low latency communicationICN information-centric networks VM virtual machineISG industry specification group VR virtual realityISI inter-symbol interference V2X vehicular-to-everythingIoT Internet of Things WPT wireless power transferKKT Karush-Kuhn-Tucker 3C computing, caching, and communicationsLP linear programming 3GPP 3rd Generation Partnership ProjectLTE long-term evolution 4C communications, computing, control and content deliveryMCC mobile cloud computing 5GPPP European 5G infrastructure Public Private PartnershipMDP Markov decision process 5QI 5G QoS Indicator
Section VI. We summarze the definitions of the acronyms thatwill be frequently use in this paper in TABLE II for ease ofreference.II. MEC C
OMPUTATION AND C OMMUNICATION M ODELS
In this section, system models are introduced for the keycomputation/communication components of the typical MECsystem. The models provide mechanisms for abstracting var-ious functions and operations into optimization problems andfacilitating theoretical analysis as discussed in the followingsections.For the MEC system shown in Fig. 3, the key componentsinclude mobile devices (a.k.a. end users, clients, service sub-scribers) and MEC servers. The MEC servers are typicallysmall-scale data centers deployed by the cloud and telecomoperators in close proximity with end users and can be co-located with wireless APs. Through a gateway, the serversare connected to the data centers via Internet. Mobile devicesand servers are separated by the air interface where reliablewireless links can be established using advanced wirelesscommunication and networking technologies. In the followingsubsections, we will introduce the models for different compo-nents of MEC systems, including models for the computationtasks, wireless communication channels and networks, as wellas the computation latency and energy consumption models ofmobile devices and MEC servers.
A. Computation Task Models
There are various parameters that play critical roles inmodeling the computation tasks, including latency, bandwidthutilization, context awareness, generality, and scalability [57].Though it is highly sophisticated to develop accurate modelsfor tasks, there exist simple ones that are reasonable and allowmathematical tractability. In this subsection, we introduce twocomputation-task models popularly used in existing literatureon MCC and MEC, corresponding to binary and partialcomputation offloading, respectively. Task Model for Binary Offloading : A highly integratedor relatively simple task cannot be partitioned and has to beexecuted as a whole either locally at the mobile device oroffloaded to the MEC server, called binary offloading . Such atask can be represented by a three-field notation A ( L, τ d , X ) .This commonly-used notation contains the information ofthe task input-data size L (in bits), the completion deadline τ d (in second), and the computation workload/intensity X (in CPU cycles per bit). These parameters are related tothe nature of the applications and can be estimated throughtask profilers [58], [59]. The use of these three parametersnot only captures essential properties of mobile applicationssuch as the computation and communication demands, butalso facilitates simple evaluation of the execution latency andenergy consumption performance (which will be analyzed inSection II-C).The task A ( L, τ d , X ) is required to be completed before (cid:19)(cid:28)(cid:29)(cid:19)(cid:19)(cid:68)(cid:80)(cid:20)(cid:19)(cid:29)(cid:19)(cid:19)(cid:68)(cid:80) (cid:19)(cid:28)(cid:29)(cid:22)(cid:19)(cid:68)(cid:80)(cid:20)(cid:19)(cid:29)(cid:22)(cid:19)(cid:68)(cid:80) Connected (cid:3) vehiclesSurveillance (cid:3) networksSmart (cid:3) devices (cid:3) applicationsHealth (cid:3) monitoring
GamingAR (cid:3)
Apps.3D (cid:3)
ModelingSocial (cid:3) networking
MEC (cid:3)
Server (cid:48)(cid:82)(cid:69)(cid:76)(cid:79)(cid:72)(cid:3)(cid:70)(cid:82)(cid:85)(cid:72)(cid:3)(cid:81)(cid:72)(cid:87)(cid:90)(cid:82)(cid:85)(cid:78)(cid:42)(cid:68)(cid:87)(cid:72)(cid:90)(cid:68)(cid:92) (cid:44)(cid:81)(cid:87)(cid:72)(cid:85)(cid:81)(cid:72)(cid:87)(cid:69)(cid:68)(cid:70)(cid:78)(cid:69)(cid:82)(cid:81)(cid:72) (cid:39)(cid:68)(cid:87)(cid:68)(cid:3)(cid:70)(cid:72)(cid:81)(cid:87)(cid:72)(cid:85) (cid:36)(cid:54)(cid:51)(cid:86)(cid:38)(cid:39)(cid:49)(cid:86)
Fig. 3. Architecture of the MEC systems. a hard deadline τ d . This model can also be generalized tohandle the soft deadline requirement which allows a smallportion of tasks to be completed after τ d [60]. In this case,the number of CPU cycles needed to execute 1-bit of taskinput data is modeled as a random variable X . Specifically,define x as a positive integer such that Pr(
X > x ) ≤ ρ where ρ is a small real number: < ρ (cid:28) . It follows that Pr(
LX > W ρ ) ≤ ρ where W ρ = Lx . Then given the L -bittask-input data, W ρ upper bounds the number of required CPUcycles almost surely. Task Models for Partial Offloading : In practice,many mobile applications are composed of multiple proce-dures/components (e.g., the computation components in anAR application as shown in Fig. 2), making it possibleto implement fine-grained (partial) computation offloading.Specifically, the program can be partitioned into two parts withone executed at the mobile device and the other offloaded foredge execution.The simplest task model for partial offloading is the data-partition model , where the task-input bits are bit-wise indepen-dent and can be arbitrarily divided into different groups andexecuted by different entities in MEC systems, e.g., parallelexecution at the mobiles and MEC server.Nevertheless, the dependency among different proce-dures/components in many applications cannot be ignoredas it significantly affects the procedure of execution andcomputation offloading due to the following reasons: • First, the execution order of functions or routines cannotbe arbitrarily chosen because the outputs of some com-ponents are the inputs of others. • Second, due to either software or hardware constraints,some functions or routines can be offloaded to the serverfor remote execution, while the ones can only be executedlocally such as the image display function.This calls for task models that are more sophisticatedthan the mentioned data-partition model that can capturethe inter-dependency among different computation functionsand routines in an application. One such model is called the task-call graph . The graph is typically a directed acyclicgraph (DAG), which is a finite directed graph with nodirected cycles. We shall denote it as G ( V , E ) , where the setof vertices V represents different procedures in the applicationand the set of edges E specifies their call dependencies. Thereare three typical dependency models of sub-tasks (i.e., taskcomponents such as functions or routines), namely sequential , parallel , and general dependency [61], [62], as illustratedin Fig. 4. For the mobile initiated applications, the first andthe last steps, e.g., collecting the I/O data and displayingthe computation results on the screen, are normally requiredto be executed locally. Thus, node 1 and node N in Fig.4(a)-4(c) are components that must be executed locally.Besides, the required computation workloads and resourcesof each procedure, e.g., the number of required CPU cyclesand the amount of needed memory, can also be specifiedin the vertices of the task-call graph, while the amount ofinput/output data of each procedure can be characterized byimposing weights on the edges. (cid:20) (cid:21) (cid:22) (cid:892) NN (cid:16)(cid:20) (cid:171) w (cid:20)(cid:21) w (cid:21)(cid:22) w N (cid:16)(cid:20)(cid:15)(cid:3) N (cid:70) (cid:20) (cid:70) (cid:21) (cid:70) (cid:22) (cid:70) N (cid:16)(cid:20) (cid:70) N (a) Sequential dependency (cid:20) (cid:21)(cid:22) (cid:892) NN (cid:16)(cid:20) (cid:171) (cid:23) (cid:70) (cid:20) (cid:70) (cid:22) (cid:70) N (cid:70) (cid:21) (cid:70) (cid:23) (cid:70) N (cid:16)(cid:20) w (cid:20)(cid:15)(cid:3)(cid:23) w (cid:23)(cid:15)(cid:3) N (b) Parallel dependency (cid:20) (cid:21) (cid:23) (cid:892) NN (cid:16)(cid:20) (cid:171) (cid:24) (cid:26)(cid:22) (cid:892) N (cid:16)(cid:21)(cid:25) (cid:171) (cid:70) (cid:20) (cid:70) (cid:22) (cid:70) (cid:25) (cid:70) (cid:21) (cid:70) (cid:23) (cid:70) (cid:24) (cid:70) N (cid:16)(cid:21) (cid:70) (cid:26) (cid:70) N (cid:16)(cid:20) (cid:70) N w (cid:20)(cid:15)(cid:3)(cid:22) w (cid:20)(cid:15)(cid:3)(cid:25) w N (cid:16)(cid:20)(cid:15)(cid:3) N w N (cid:16)(cid:21)(cid:15)(cid:3) N-1 w (cid:21)(cid:15)(cid:3)(cid:23) w (cid:23)(cid:15)(cid:3)(cid:26) w (cid:21)(cid:15)(cid:3)(cid:24) w (cid:22)(cid:15)(cid:3)(cid:24) (c) General dependencyFig. 4. Typical topologies of the task-call graphs. B. Communication Models
In the literature of MCC, communication channels betweenthe mobile devices and cloud servers are typically abstracted asbit pipes with either constant rates or random rates with givendistributions. Such coarse models are adopted for tractabilityand may be reasonable for the design of MCC systemswhere the focuses are to tackle the latency in the corenetworks and management of large-scale cloud but not thewireless-communication latency. The scenario is different forMEC systems. Given small-scale edge clouds and targetinglatency-critical applications, reducing communication latencyby designing a highly efficient air interface is the maindesign focus. Consequently, the mentioned bit-pipe modelsare insufficient as they overlook some fundamental propertiesof wireless propagation and are too simplified to allow theimplementation of advanced communication techniques. To bespecific, wireless channels differ from the wired counterpartsin the following key aspects [63]:1) Due to atmospheric ducting, reflection and refractionfrom scattering objects in the environment (e.g., build-ings, walls and trees), there exists the well-known multi-path fading in wireless channels, making the channelshighly time-varying and can cause severe inter-symbolinference (ISI). Thus, effective ISI suppression tech-niques, such as equalization and spread spectrum, areneeded for reliable transmissions.2) The broadcast nature of wireless transmissions resultsin a signal being interfered by other signals occupy-ing the same spectrum, which reduces their respectivereceive signal-to-interference-plus-noise ratios (SINRs)and thereby results in the probabilities of error in detec-tion. To cope with the performance degradation, inter-ference management becomes one of the most importantdesign issues for wireless communication systems andhas attracted extensive research efforts [64]–[66].3) Spectrum shortage has been the main foe for veryhigh-rate radio access, motivating extensive research onexploiting new spectrum resources [67], [68], designingnovel transceiver architectures [69]–[71] and networkparadigms [72], [73] to improve the spectrum efficiency,as well as developing spectrum sharing and aggregationtechniques to facilitate efficient use of fragmented andunderutilized spectrum resources [74]–[76].The random variations of wireless channels in time, fre- quency and space make it important for designing efficientMEC systems to seamlessly integrate control of computationoffloading and radio resource management. For instance, whenthe wireless channel is in deep fade, the reduction on executionlatency by remote execution may not be sufficient to compen-sate for the increase of transmission latency due to the steepdrop in transmission-data rates. For such cases, it is desirableto defer offloading till the channel gain is favorable or switchto an alternative frequency/spatial channel with a better qualityfor offloading. Furthermore, increasing transmission power canincrease the data rate, but also lead to a larger transmissionenergy consumption. The above considerations necessitate thejoint design of offloading and wireless transmissions, whichshould be adaptive to the time-varying channels based on theaccurate channel-state information (CSI).In MEC systems, communications are typically betweenAPs and mobile devices with the possibility of direct D2Dcommunications. The MEC servers are small-scale data cen-ters deployed by the Cloud Computing/telecom operators,which can be co-located with the wireless APs, e.g., thepublic WiFi routers and BSs, as so to reduce the capitalexpenditure (CAPEX) (e.g., site rental). As shown in Fig. 3,the wireless APs not only provide the wireless interface forthe MEC servers, but also enable the access to the remotedata center through backhaul links, which could help the MECserver to further offload some computation tasks to other MECservers or to large-scale cloud data centers. For the mobiledevices that cannot communicate with MEC servers directlydue to insufficient wireless interfaces, D2D communicationswith neighboring devices provide the opportunity to forwardthe computation tasks to MEC servers. Furthermore, D2Dcommunications also enable the peer-to-peer cooperation onresource sharing and computation-load balancing within acluster of mobile devices.Presently, there exist different types of commercializedtechnologies for mobile communications, including the near-filed communications (NFC), radio frequency identification (RFID), Bluetooth, WiFi, and cellular technologies such as the long-term evolution (LTE). Besides, the 5G network, whichwill be realized by the development of LTE in combinationwith new radio-access technologies, is currently being stan-dardized and will be put into commercial use as early as2020 [77]. These technologies can support wireless offloadingfrom mobiles to APs or peer-to-peer mobile cooperation for
TABLE IIIC
HARACTERISTICS OF T YPICAL W IRELESS C OMMUNICATION T ECHNOLOGIES .NFC RFID Bluetooth WiFi LTE 5GMax. Coverage 10cm 3m 100m 100m up to 5km Excellent coverageOperation Freq. LF: 120-134kHz13.56MHz HF: 13.56MHz 2.4GHz 2.4GHz, 5GHz TDD: 1.85-3.8GHz 6-100GHzUHF: 850-960MHz FDD: 0.7-2.6GHzData Rate Indoor/dense outdoor:106, 212, Low (LF) to 135Mbps DL: 300Mbps up to 10Gbps414kbps high (UHF) 22Mbps (IEEE 802.11n) UL: 75Mbps Urban/suburban: > hundreds of Mbps varying data rates and transmission ranges. We list the keycharacteristics of typical wireless communication technologiesin Table III, which differ significantly in terms of the operationfrequency, maximum coverage range, and data rate. For NFC,the coverage range and data rate are very low and thusthe technology is suitable for applications that require littleinformation exchange, e.g., e -payment and physical accessauthentication. RFID is similar to NFC, but only allows one-way communications. Bluetooth is a more powerful tech-nique to enable short-range D2D communications in MECsystems. For long-range communications between mobiles andMEC servers, WiFi and LTE (or 5G in the future) are twoprimary technologies enabling the access to MEC systems,which can be adaptively switched depending on their linkreliability. For the deployment of wireless technologies inMEC systems, the communication and networking protocolsneed to be redesigned to integrate both the computing andcommunication infrastructures, and effectively improve thecomputation efficiency that is more sophisticated than the datatransmission. C. Computation Models of Mobile Devices
In this subsection, we introduce the computation models ofmobile devices and discuss methodologies of evaluating thecomputation performance.The CPU of a mobile device is the primary engine forlocal computation. The CPU performance is controlled bythe CPU-cycle frequency f m (also known as the CPU clockspeed). The state-of-the-art mobile CPU architecture adoptsthe advanced dynamic frequency and voltage scaling (DVFS)technique, which allows stepping-up or -down of the CPU-cycle frequency (or voltage), resulting in growing and reducingenergy consumption, respectively. In practice, the value of f m is bounded by a maximum value, f maxCPU , which reflectsthe limitation of the mobile’s computation capability. Basedon the computation task model introduced in Section II-A,the execution latency for task A ( L, τ, X ) can be calculatedaccordingly to t m = LXf m , (1)which indicates that a high CPU clock speed is desirable inorder to reduce the execution latency, at the cost of higherCPU energy consumption.As the mobile devices are energy-constrained, the energyconsumption for local computation is another critical mea-surement for the mobile computing efficiency. According to the circuit theory [78]–[81], the CPU power consumption canbe divided into several factors including the dynamic , short-circuit , and leakage power consumption , where the dynamicpower consumption dominates the others. In particular, itis shown in [80] that the dynamic power consumption isproportional to the product of V f m where V cir is the circuitsupplied voltage. It is further noticed in [78], [81] that, theclock frequency of the CPU chip is approximately linearproportional to the voltage supply when operating at the lowvoltage limits. Thus, the energy consumption of a CPU cycleis given by κf m , where κ is a constant related to the hardwarearchitecture. For the computation task A ( L, τ, X ) with CPUclock speed f m , the energy consumption can be derived: E m = κLXf m . (2)One can observe from (1) and (2) that the mobile device maynot be able to complete a computation-intensive task withinthe required deadline, or else the energy consumption incurredby mobile execution is so high that the onboard battery will bedepleted quickly. In such cases, offloading the task executionprocess to an MEC server is desirable.Besides CPUs, other hardware components in the mobiledevices, e.g., the random access memory (RAM) and flashmemory, also contribute to the computation latency andenergy consumption [82], while detailed discussions arebeyond the scope of this survey. D. Computation Models of MEC Servers
In this subsection, we introduce the computation modelsof the MEC servers. Similar as the mobile devices, thecomputation latency and energy consumption are of particularinterests.The server-computation latency is negligible compared withcommunication or local-computation latency in MEC systemswhere the computation loads for servers are much lower thantheir computation capacities [81], [83]. This model can be alsorelevant for multiuser MEC systems with resource-constrained The dynamic power consumption comes from the toggling activities of thelogic gates inside a CPU, which shall charge/discharge the capacitors insidethe logic gates. When a logic gate toggles, some of its transistors may changestates, and thus, there might be a short period of time when some transistorsare conducting simultaneously. In this case, the direct path between the sourceand ground will result in some short-circuit power loss. The leakage powerdissipation is due to the flowing current between doped parts of the transistors[80], available on https://en.wikipedia.org/wiki/CPU power dissipation. servers if the servers’ computation loads are regulated by mul-tiuser resource management under latency and computation-capacity constraints [84].On the other hand, as edge servers have relatively limitedcomputation resources, it is necessary to consider the non-negligible server execution time in the general design ofMEC systems, yielding the computation model for the seversdiscussed in the remainder of this subsection. Two possiblemodels are considered in the literature, corresponding to the deterministic and stochastic server-computation latency. Thedeterministic model is proposed to consider the exact server-computation latency for latency-sensitive applications, whichis implemented using techniques such as VMs and DVFS.Specifically, assume the MEC server allocates different VMsfor different mobile devices, allowing independent computa-tion [85]. Let f s,k denote the allocated servers’ CPU-cyclenumber for mobile device k . Similar to Section II-C, itfollows that the server execution time denoted by t s,k canbe calculated as t s,k = w k f s,k , where w k is the number ofrequired CPU cycles for processing the offloaded computationworkload. This model has been widely used for designingcomputation-resource allocation policies [86]–[88]. A similarmodel was proposed in [84], where the MEC server is assumedto perform load balancing for the total offloaded computationworkloads. In other words, the CPU cycles at the MEC serverare proportionally allocated to each mobile device such thatthey experience the same execution latency. Furthermore, inaddition to the CPU processing time, the server schedulingqueuing delay should be accounted for MEC servers withrelatively small computation capacities, where parallel com-puting via virtualization techniques is not feasible and thusit needs to process the computation workloads sequentially.Without loss of generality, denote k as the processing order fora mobile device and name it as mobile k . Thus, the total server-computation latency including the queuing delay for device k denoted by T s,k can be given as T s,k = (cid:88) i ≤ k t s,i . (3)For latency-tolerant applications, the average server-computation time can be derived based on stochastic models.For example, in [89], the task arrivals and service timeare modeled by the Poisson and exponential processes,respectively. Thus, the average server-computation time canbe derived using techniques from queuing theory. Last,for all above models, as investigated in [1], multiple VMssharing the same physical machine will introduce the I/Ointerference among different VMs. It results in the longercomputation latency for each VM denoted by T (cid:48) s,k , whichcan be modeled by T (cid:48) s,k = T s,k (1 + (cid:15) ) n where (cid:15) is theperformance degradation factor as the percentage increasingof the latency [90].The energy consumption of an MEC server is jointlydetermined by the usage of the CPU, storage, memory, andnetwork interfaces. Since the CPU contribution is dominantamong these factors, it is the main focus in the literature. Twotractable models are widely used for the energy consumption of MEC servers. One model is based on the DVFS techniquedescribed as follows. Consider an MEC server that handles K computation tasks and the k -th task is allocated with w k CPUcycles with CPU-cycle frequency f s,k . Hence, the total energyconsumed by the CPU at the MEC server, denoted by E s , canbe expressed as E s = K (cid:88) k =1 κw k f s,k , (4)which is similar to that for the mobile devices. The other modelis based on an observation in recent works [91]–[93] that theserver-energy consumption is linear to the CPU utilizationratio which depends on the computation load. Moreover, evenfor an idle server, it still, on average, consumes up to 70%of the energy consumption for the case with the full CPUspeed. Thus, the energy consumption at the MEC server canbe calculated according to E s = αE max + (1 − α ) E max u, (5)where E max is the energy consumption for a fully-utilizedserver, α is the fraction of the idle energy consumption (e.g.,70%) and u denotes the CPU utilization ratio. This modelsuggests that energy-efficient MEC should allow servers to beswitched into the sleep mode in the case of light load andconsolidation of computation loads into fewer active servers. E. Summary and Insights
The MEC computation and communication models aresummarized in Fig. 5, laying the foundation for the analysis ofMEC resource management in the next section. These modelsshed several useful insights on the offloading design, listed asfollows. • The effective design of MEC should leverage and inte-grate advanced techniques from both areas of wirelesscommunications and mobile computing. • It is vital to choose suitable computation task modelsfor different MEC applications. For example, the soft-deadline task model can be applied for social network-ing applications but is not suitable for AR applicationsdue to the stringent computation latency requirements.Moreover, for a specific application, the task model alsodepends on the offloading scenario, e.g., the data-partitionmodel can be used when the input-data is offloaded, andthe task-call graph should be considered when each taskcomponent can be offloaded as a whole. • The wireless channel condition significantly affects theamount of energy consumption for computation offload-ing. MEC has the potential to reduce the transmissionenergy consumption due to short distances between usersand MEC servers. Advanced wireless communicationtechniques, such as interference cancelation and adaptivepower control, can further reduce the offloading energyconsumption. • Dynamic CPU-cycle frequency control is the key tech-nique for controlling the computation latency and en-ergy consumption for both mobile devices and MECservers. Specifically, increasing the CPU-cycle frequency Fig. 5. Summary of MEC models. can reduce the computing time but contributes to higherenergy consumption. The effective CPU-cycle frequencycontrol should approach the optimal tradeoff betweencomputation latency and energy consumption. • Apart from the task-execution latency, the computationscheduling delay is non-negligible if the MEC server hasa relatively small computation capacity or heavy compu-tation loads are offloaded to the server. Load-balancingand intelligent scheduling policies can be designed toreduce the total computation latency.III. R
ESOURCE M ANAGEMENT IN
MEC S
YSTEMS
The joint radio-and-computational resource managementplays a pivotal role in realizing energy-efficient and low-latency MEC. The implementation of relevant techniques isfacilitated by the network architecture where MEC serversand wireless APs (e.g., BSs and WiFi routers) are co-located.In this section, we provide a comprehensive overview ofthe literature on resource management for MEC systemssummarized in Fig. 6. Our discussion starts from the simplesingle-user systems comprising a single mobile device and asingle MEC server, allowing the exposition of the key design considerations and basic design methodologies. Subsequently,more complex multiuser MEC systems are considered wheremultiple offloading users compete for the use of both the radioand server-computation resources and have been coordinated.Last, we extend the discussion to MEC systems with het-erogeneous servers which not only provide the freedom ofserver selection but also allow the cooperation among servers.Such network-level operations can significantly enhance theperformance of MEC systems.
A. Single-User MEC Systems
This subsection focuses on the simple single-user MEC sys-tems and reviews a set of recent research efforts for this case.The discussion is divided according to three popularly-usedtask models, namely, deterministic task model with binaryoffloading, deterministic task model with partial offloading,and stochastic task model. Deterministic Task Model with Binary Offloading : Consider the mentioned single-user MEC system where thebinary offloading decision is on whether a particular taskshould be offloaded for edge execution or local computation.The investigations for the optimal offloading policies can be Resource Management in MEC Systems Multiuser MEC SystemsSingle-User MEC Systems
1. Deterministic Task Model with Binary Offloading 2. Deterministic Task Model with Partial Offloading 3. Stochastic Task Model 1. Joint Radio-and-Computational Resource Allocation 2. MEC Server Scheduling 3. Multiuser Cooperative Edge Computing
MEC Systems with Heterogeneous Servers
1. Server Selection 2. Server Cooperation 3. Computation Migration
Fig. 6. Classification of resource management techniques for MEC. dated back to those for conventional Cloud Computing sys-tems, where the communication links were typically assumedto have a fixed rate B . In [94] and [95], general guidelinesare developed for determining the offloading decision for thepurposes of minimizing the mobile-energy consumption andcomputation latency. Denote w as the amount of computation(in CPU cycles) for a task, f m as the CPU speed of the mobiledevice, d as the input data size, and f s as the CPU speed at thecloud server. Offloading the computation to the cloud servercan improve the latency performance only when wf m > dB + wf s , (6)which holds for applications that require heavy computationand have small amount of data input, or when the cloud serveris fast, and the transmission rate is sufficiently high. Moreover,let p m represent the CPU power consumption at the mobiledevice, and p t as the transmission power, p i as the powerconsumption at the device when the task is running at theserver. Offloading the task could help save mobile energy when p m × wf m > p t × dB + p i × wf s (7)holds, i.e., applications with heavy computation and lightcommunication should be offloaded.Nevertheless, the data rates for wireless communications arenot constant and change with the time-varying channel gainsas well as depend on the transmission power. This calls forthe design of control policies for power adaptation and datascheduling to streamline the offloading process. In addition, asthe CPU power consumption increases super-linearly with theCPU-cycle frequency, the computation energy consumption formobile execution can be minimized using DVFS techniques.These issues led to the active field of adaptive MEC assummarized below.In [96], the problem of transmission-energy minimizationunder a computation-deadline constraint was formulated withthe optimization variable being the input-data transmissiontime, where the famous Shannon-Hartley formula gives the power-rate function. The optimization problem is convex andcan be solved in closed form. In particular, task offloadingis desirable when the channel power gain is greater than athreshold and the server CPU is fast enough, which revealsthe effects of wireless channels on the offloading decision.A further study was conducted by Zhang et al. in [81] tominimize the energy consumption for executing a task witha soft real-time requirement, targeting e.g., multimedia appli-cations, which requires the task to be completed within thedeadline with a given probability ρ . The offloading decisionwas determined by the computation mode (either offloadingor local computing) that incurs less energy consumption. Onone hand, the energy consumption for local execution was op-timized using the DVFS technique, which was formulated as aconvex optimization problem with the objective function beingthe expected energy consumption of the W ρ CPU cycles anda time duration constraint for these CPU cycles. The optimalCPU-cycle frequencies over the computation duration werederived in closed form by solving the
Karush-Kuhn-Tucker (KKT) conditions, suggesting that the processor should speedup as the number of completed CPU cycles increases. On theother hand, the expected energy consumption for task offload-ing was minimized via data transmission scheduling. Under theGilbert-Elliott channel model, the optimal data transmissionscheduling was obtained through dynamic programming (DP)techniques, and the scaling law of the minimum expectedenergy consumption with respect to the execution deadline wasalso derived. This framework was further developed in [83]where both the local computing and offloading are poweredby wireless energy transfer. Specifically, the optimal CPU-cycle frequencies for local computing and time division foroffloading should be adaptive to the transferred power. Deterministic Task Model with Partial Offloading : Therunning of a relatively sophisticated mobile application can bedecomposed into a set of smaller sub-tasks. Inspired by recentadvancements of parallel computing, partial offloading (alsoknown as program partitioning) schemes were proposed tofurther optimize MEC performance in [61], [62], [97]–[102].In [97], full granularity in program partitioning was con- sidered where the task-input data can be arbitrarily dividedfor local and remote executions. Joint optimization of theoffloading ratio, transmission power and CPU-cycle frequencywas performed to minimize the mobile-energy consumption(or latency) subject to a latency (or energy consumption)constraint. Both the energy and latency minimization problemsare non-convex in contrast to the ones for binary-offloading.The former problem can be solved optimally with a variable-substitution technique while a sub-optimal algorithm wasproposed for the latter one in [97].In [61], [62], [98]–[102], applications were modeled bytask-call graphs discussed earlier that specify the dependencyamong different sub-tasks, and the code partitioning schemesdesigned to dynamically generate the optimal set of tasks foroffloading. In [61], by leveraging the concept of load balancingbetween the mobile device and the server, a heuristic program-partitioning algorithm was developed to minimize the execu-tion latency. Kao et al. investigated the latency minimizationproblem with a prescribed resource utilization constraint in[98], and proposed a polynomial-time approximate solutionwith guaranteed performance. To maximize the energy savingsachieved by computation offloading, the scheduling and cloudoffloading decisions were jointly optimized using an integerprogramming approach in [62]. In [99], considering the wire-less channel models including the block fading channel, in-dependent and identical distributed (i.i.d.) stochastic channel,and the Markovian stochastic channel, the expected energyconsumption minimization problem with a completion timeconstraint was found to be a stochastic shortest-path problem,and the one-climb policies (i.e., the execution only migratesonce from the mobile device to the server) were shown to beoptimal. In addition, the program-partitioning schemes werealso optimized together with the physical layer parameters,such as the transmission and reception power, constellationsize, as well as the data allocation for different radio interfaces[100]–[102]. Stochastic Task Model : Resource management policieshave been also developed for MEC systems with stochastictask models characterized by random task arrivals, where thearrived but not yet executed tasks join the queues in buffers[103]–[108]. For such systems, the long-term performance,e.g., the long-term average energy consumption and executionlatency, are more relevant compared with those of determin-istic task arrivals, and the temporal correlation of the optimalsystem operations makes the design more challenging. As aresult, the design of MEC systems with random task arrivalsis an area less explored compared with the simpler caseswith deterministic task models. In [103], in order to minimizethe mobile-energy consumption while keeping the proportionof executions violating the deadline requirement below athreshold, a dynamic offloading algorithm was proposed todetermine the offloaded software components from an applica-tion running at a mobile user based on Lyapunov optimizationtechniques, where 3G and WiFi networks are accessible to thedevice but their rates vary at different locations. Assuming thatconcurrent local and edge executions are feasible, the latency-optimal task scheduling policies were designed in [104] basedon the theory of
Markov decision process (MDP), which controls the states of the local processing and transmissionunits and the task buffer queue length based on the channelstate. It was shown that the optimal task-scheduling policysignificantly outperforms the greedy scheduling policy (i.e.,tasks are scheduled to the local CPU/transmission unit when-ever they are idle). To jointly optimize the computation latencyand energy consumption, the problem of minimizing the long-term average execution cost was considered in [102] and [106],where the former only optimized the offloading data sizebased on the MDP theory while the latter jointly controlledthe local CPU frequency, modulation scheme as well as datarates under a semi-MDP framework. In [107], the energy-latency tradeoff in MEC systems with heterogeneous types ofapplications was investigated, including the non-offloadableworkload, cloud-offloadable workload and network traffic.A Lyapunov optimization-based algorithm was proposed tojointly decide the offloading policy, task allocation, CPU clockspeed, and selected network interface. It was also shownthat the energy consumption decreases inversely proportionalto V while the latency increases linearly with V , where V is a control parameter in the proposed algorithm. Similarinvestigation was conducted for MEC systems with a multi-core mobile device in [108].
4) Summary and Insight:
The comparison of resourcemanagement schemes for single-user MEC systems is shownin Table IV. This series of work yields a number of usefulinsights on controlling computation offloading as summarizedbelow. • Consider binary offloading. For energy savings, com-putation offloading is preferred to local computationwhen the user has desirable channel condition or smalllocal computation capability. Moreover, beamforming andMIMO techniques can be exploited to reduce the energyconsumption for offloading. For latency reduction, com-putation offloading is advantageous over local computa-tion when the user has a large bandwidth and the MECserver is provisioned with huge computation capacity. • Partial offloading allows flexible components/data par-titioning. By offloading time-consuming or energy-consuming sub-tasks to MEC servers, partial offloadingcan achieve larger energy savings and smaller compu-tation latency compared with binary offloading. Graphtheory is a powerful tool for designing the offloadingscheduling according to the task dependency graph. • For stochastic task models, the temporal correlation oftask arrivals and channels can be exploited to designadaptive dynamic computation offloading policies. More-over, it is critical to maintain the task buffer stability atthe user and MEC server via offloading rate control.
B. Multiuser MEC Systems
While the preceding subsection aims at resource manage-ment policies for single-user MEC systems with a dedicatedMEC server, this subsection considers the multiuser MECsystems comprising multiple mobile devices that share oneedge server. Several new challenges are investigated in thesequel, including the multiuser joint radio-and-computational TABLE IVT
HE COMPARISON OF PAPERS FOCUSING ON SINGLE - USER
MEC
SYSTEMS . Task model Design Objective Reference Proposed Solution
Binary Offloading Energy [81] Optimize local computing and offloading by controlling the CPUfrequency and transmission rate[83] Propose a novel framework of wirelessly powered MEC and optimizeboth local computing and offloading[94] Propose general guidelines to make offloading decision for energyconsumption minimization[96] Propose the optimal binary computation offloading decision usingconvex optimizationEnergy and latency [95] Propose general guidelines to make offloading decision for energy-consumption and computation-latency minimizationPartial Offloading Energy [62] Propose a joint scheduling and computation offloading algorithm byparallel processing appropriate components in the mobile and cloud[99] Formulate a stochastic shortest-path problem and derive the one-climboptimal policy[101] Jointly optimize the program partitioning with the selection of transmitpower and constellation size[102] Propose an iterative algorithm for the optimal offloading scheduling aswell as the percentage of the data to be carried on each radio interfaceLatency [61] Propose a heuristic load-balancing program-partitioning algorithm[98] Propose a polynomial-time approximate solution with guaranteedperformanceEnergy and latency [97] Jointly optimize the offloading ratio, transmission power and CPU-cycle frequency using variable-substitution technique[100] Propose an algorithmic to leverage the structure of the call graphs bymeans of message passing under both serial and parallel implementa-tions of processing and communicationStochastic Model Energy [103] Propose a Lyapunov optimization-based dynamic computation offload-ing policyLatency [104] Dynamically control the local processing and transmission using MDP[105] Optimize local computing and transmission using semi-MDP andpropose a one-dimensional heuristic search algorithmEnergy and Latency [106] Jointly control the local CPU frequency, modulation scheme as wellas the data rates under a semi-MDP framework[107] Propose a Lyapunov optimization-based algorithm to decide the of-floading policy, task allocation, CPU clock speed, and selected networkinterface[108] Propose a Lyapunov optimization-based scheme for cloud offloadingscheduling, as well as download scheduling for cloud execution output resource allocation, MEC server scheduling, and multiusercooperative edge computing. Joint Radio-and-Computational Resource Allocation : Compared with the central cloud, the MEC servers have muchless computational resources. Therefore, one key issue indesigning a multiuser MEC system is how to allocate thefinite radio-and-computational resources to multiple mobilesfor achieving a system-level objective, e.g., the minimumsum mobile-energy consumption. Both the centralized anddistributed resource allocation schemes have been studied fordifferent MEC systems as reviewed in the following..For centralized resource allocation [84], [86], [101], [109]–[114], the MEC server obtains all the mobile information, in-cluding the CSI and computation requests, makes the resource-allocation decisions, and informs the mobile devices aboutthe decisions. In [84], mobile users time-share a single edgeserver and have different computation workloads and local-computation capacities. A convex optimization problem wasformulated to minimize the sum mobile-energy consumption.The key finding is that the optimal policy for controllingoffloading data size and time allocation has a simple threshold-based structure. Specifically, an offloading priority functionwas firstly derived according to mobile users’ channel condi-tions and local computing energy consumption. Then, the users with priorities above and below a given threshold will performfull and minimum offloading (so as to meet a given compu-tation deadline), respectively. This result was also extendedto the OFDMA-based MEC systems for designing a close-to-optimal computation offloading policy. In [86], insteadof controlling the offloading data size and time, the MECserver determined the mobile-transmission power and assignedserver CPU cycles to different users in order to reduce thesum mobile-energy consumption. The optimal solution showsthat, there exists an optimal one-to-one mapping between thetransmission power and the number of allocated CPU cyclesfor each mobile device. This work was further extended in[101] to account for the optimal binary offloading based onthe model of task-call graphs. In [112], the authors con-sidered the multiuser video compression offloading in MECand minimized the latency in local compression, edge cloudcompression and partial compression offloading scenarios.Besides, in order to minimize the energy and delay cost formulti-user MEC systems where each user has multiple tasks,Chen et al. jointly optimized the offloading decisions and theallocation of communication resource via a separable semidef-inite relaxation approach in [113], which was later extendedin [114] by taking the computational resource allocation andprocessing cost into account. Different from [84], [86], [101], [112]–[114], the revenue of service providers was maximizedin [109] under constraints of quality of service (QoS) require-ments for all mobile devices. The assumed fixed resourceusage of each user results in a semi-MDP problem, whichwas transformed into a linear programming (LP) model andefficiently solved. In [110], assuming a stochastic task arrivalmodel, the energy-latency tradeoff in multiuser MEC systemswas investigated via a Lyapunov optimization-based onlinealgorithm, which jointly manages the available radio-and-computational resources. Centralized resource management formultiuser MEC system based on cloud radio access network (C-RAN) has also been investigated in [111].Another thrust of research targets distributed resource allo-cation for multiuser MEC systems which were designed usinggame theory and decomposition techniques [87], [88], [115]–[119]. In [115] and [87], the computation tasks were assumedto be either locally executed or fully offloaded via singleand multiple interference channels, respectively. With fixedmobile-transmission power, an integer optimization problemwas formulated to minimize the total energy consumption andoffloading latency, which was proved to be NP-hard. Instead ofdesigning a centralized solution, the game-theoretic techniqueswere applied to develop a distributed algorithm that is ableto achieve a Nash equilibrium. Moreover, it was shown thatfor each user, offloading is beneficial only when the receivedinterference power is lower than a threshold. Furthermore, thiswork was extended in [116] and [117], where each mobile hasmultiple tasks and can offload computation to multiple APsconnected by a common edge-server, respectively. For the of-floading process, in addition to transmission energy, this workhas also accounted for the scanning energy of the APs and thefixed circuit power. The proposed distributed offloading policyshowed that a mobile device should handover the computationto a different AP only when a new user choosing the sameAP achieves a larger benefit. Building on the system modelin [87], the joint optimization for the mobile-transmissionpower and the CPU-cycle allocation of the edge server wasinvestigated in [88]. To solve the formulated mixed-integerproblem, the decomposition technique was utilized to optimizethe resource allocation and offloading decision sequentially.Specifically, the offloading decision problem was reduced to asub-modular maximization problem and solved by designinga heuristic greedy algorithm. Similar decomposition techniqueand successive convex approximation technique were utilizedin [118] and [119] respectively to design distributed resourceallocation algorithm for MEC systems. MEC Server Scheduling : The works discussed earlier[84], [86]–[88], [109], [117] are based on the assumptions ofuser synchronization and the feasibility of parallel local-and-edge computation. However, studying practical MEC serverscheduling requires relaxation of these assumptions as dis-cussed below together with the resultant designs. First, thearrival times of different users are in general asynchronous sothat it is desirable for the edge server with finite computationalresource to buffer and compute the tasks sequentially, whichincurs the queuing delay. In [120], to cope with the burstytask arrivals, the server scheduling was integrated with uplink-downlink transmission scheduling to minimize the average latency using queuing theory. Second, even for synchronizedtask arrivals, the latency requirements can differ significantlyover users running different types of applications ranging fromlatency-sensitive to latency-tolerant applications. This factcalls for the server scheduling assigning users different levelsof priorities based on their latency requirements. In [121], afterthe pre-resource allocation, the MEC server will check thedeadline of different tasks during the server computing processand adaptively adjust the task execution order to satisfy theheterogeneous latency requirements. Last, some computationtasks each consists of several dependent sub-tasks such that thescheduling of these modules must satisfy the task-dependencyrequirements. The task model with a sequential sub-taskarrangement was considered in [122] that jointly optimizesthe program partitioning for multiple users and the server-computation scheduling to minimize the average completiontime. As a result, a heuristic algorithm was proposed tosolve the formulated mixed-integer problem. Specifically, itfirst optimizes the computation partition for each user. Underthese partitions, it will search the time intervals violating theresource constraint and adjust them accordingly. Furthermore,the general dependency-task model as shown in Fig. 4(c) wasconsidered for multiple users in [118]. This model drasticallycomplicates the computing time characterization. To addressthis challenge, a measure of ready time was defined for eachsub-task as the earliest time when all the predecessors havebeen computed. Then, the offloading decision, mobile CPU-cycle frequency and mobile-transmission power were jointlyoptimized to reduce the sum mobile-energy consumption andcomputation latencies with a proposed distributed algorithm. Multiuser Cooperative Edge Computing : Multiusercooperative computing is envisioned as a promising techniqueto improve the MEC performance by providing two advantages[123]–[129]. First, MEC servers with limited computationalresources may be overloaded when they have to serve alarge number of offloading mobile users. In such cases, theburdens on the servers can be lightened via peer-to-peer mobilecooperative computing. Second, sharing the computationalresources among the users can balance the uneven distributionof the computation workloads and computation capabilitiesover users. In [123], D2D communication was proposed toenable multiuser cooperative computing. In particular, thiswork studied how to detect and utilize computational resourceson other users. This idea was adopted in [124] to proposea D2D-based heterogeneous MCC networks. Such a novelframework was shown to enhance the network capacity andoffloading probability. Moreover, for wireless sensor networks,cooperative computing was proposed in [125] to enhanceits computation capability. First, the optimal computationpartition for minimizing the total energy consumption oftwo cooperative nodes was investigated. This result was thenutilized to design the fairness-aware energy-efficient cooper-ative node selection. Furthermore, Song et al. showed thatsharing computation results among the peer users can sig-nificantly reduce the communication traffic for a multiuserMEC system [126]. Assuming the task can either be offloadedor computed locally, a mixed-integer optimization problemwas formulated to minimize the total energy consumption TABLE VT
HE COMPARISON OF PAPERS FOCUSING ON MULTIUSER
MEC
SYSTEMS . Theme DesignType/Motivation Design Objective Reference Proposed Solution
Centralized Energy [84] Design the optimal threshold-based resource allocation policybased on defined offloading priority function for TDMA andOFDMA systems[86] Jointly optimize the allocation of communication and com-putation resources[101] Design the optimal resource allocation and code partitioningby call-graph selection approach[111] Solve the non-convex resource allocation problem for C-RANusing iterative algorithmsLatency [112] Minimize the latency in multiuser video compression viaresource allocationEnergy and latency [110] Propose a Lyapunov optimization-based dynamic computa-tion offloading policy[113], [114] Jointly optimize the offloading decisions and the allocationof resource via semidefinite relaxationJoint radio-and-computational Revenue [109] Design the optimal resource allocation based on semi-MDPresourceallocation Distributed Energy [119] Propose a distributed iterative algorithm using successiveconvex approximation techniqueEnergy and latency [87], [115] Develop a distributed algorithm that is able to achieve a Nashequilibrium[116] Propose a distributed algorithm for multi-user MEC systemswhere each user has multiple tasks[117] Consider multiple servers and develop a distributed algorithmadmitting the Nash equilibrium[118] Propose a decomposition algorithm to control the computationoffloading selection, clock frequency control and transmissionpower allocation iterativelyUtility [88] Propose a decomposition algorithm to optimize the resourceallocation and offloading decisionsMEC server Bursty dataarrivals Latency [120] Optimize the uplink and downlink scheduling using queuingtheoryHeterogeneousdeadlines Energy [121] Propose a pre-resource allocation and joint scheduling schemescheduling Task dependency Latency [122] Propose heuristic algorithm with searching and adjustingphases based on constraint relaxationEnergy and latency [118] Propose a decomposition algorithm to control the computationoffloading selection, clock frequency control and transmissionpower allocation iterativelyCooperative D2D Task success rate [123] Propose the optimal and periodic mobile cloud access schemecommunication Network capacity andoffloading probability [124] Propose D2D communication techniques in heterogeneousMEC systemsCooperation Energy [125] Propose a fairness-aware energy-efficient cooperative nodeselection scheme[127] Propose a four-slot protocol to enable joint computation andcommunication cooperationcomputing Sharecomputationresults Energy [126] Propose a Lyapunov optimization-based cooperative comput-ing policySharecomputationalresource Energy [128] Propose a “string-pulling” offloading policy based on con-structed offloading feasibility tunnelSmall BSs cooper-ation Delay cost [129] Propose a peer offloading framework that allows both cen-tralized and autonomous decision making under the constraint of the system communication traffic. Totackle this challenging problem, two online task schedulingalgorithms were proposed based on pricing and Lyapunovoptimization theories. In addition, by employing a helper,a four-slot joint computation-and-communication cooperationprotocol was proposed in [127], where the helper not onlycomputes part of the tasks offloaded from the user, but alsoacts as a relay node to forward the tasks to the MEC server.Another recent work [128] investigated the optimal offloadingpolicies in a peer-to-peer cooperative computing system wherethe computing helper has time-varying computation resources. Specifically, an offloading feasibility tunnel was constructedbased on the helper’s CPU profile and buffer size. Given thetunnel, the optimal offloading was shown to be achieved bythe well-known “string-pulling” strategy, graphically referringto pulling a string across the tunnel. Last, Chen et al. proposedan online peer offloading framework based on Lyapunovoptimization and game theoretic approaches in [129], whichenables small BSs cooperation to handle the spatially unevencomputation workloads in the network.
4) Summary and Insight:
The comparison of resource man-agement schemes for multiuser MEC systems is provided in Table V. We draw several conclusions on resource allocation,MEC server scheduling and mobile cooperative computing asfollows. • Consider multiuser MEC systems with finite radio-and-computational resources. For system-lever objectives,e.g., to minimize the sum mobile energy consumption, theusers with large channel gains and low local-computationenergy consumption have higher priorities for offloadingcomputation since they can contribute to larger energysavings. Too many offloading users, however, will causesevere inter-user interference of communication and com-putation, which will, in turn, reduce the system revenue. • To effectively reduce the sum computation latency ofmultiple users, the scheduling design for a MEC servershould assign higher priorities to the users with morestringent latency requirements and heavy computationloads. Moreover, parallel computing can further boost thecomputation speed at the server. • Scavenging the enormous amount of distributed computa-tion resources can not only alleviate the network conges-tion, but also improves resource utilization and enablesubiquitous computing. This vision can be materializedby peer-to-peer mobile cooperative edge computing. Thekey advantages include short-range transmission via D2Dtechniques and computation resource and result sharing.
C. MEC Systems with Heterogeneous Servers
To enable ubiquitous edge computing, heterogeneous MEC (Het-MEC) systems were proposed in [130] comprising onecentral cloud and multiple edge servers. The coordinationand interaction of multi-level central/edge clouds introducemany interesting new research challenges and recently haveattracted extensive relevant investigations on server selection,cooperation and computation migration, as discussed in thesequel. Server Selection : For users served by a Het-MECsystem, a key design issue is to determine the destination ofcomputation offloading, i.e., either the edge or central cloudserver. In [131], the server selection problem was studiedfor a multiuser system comprising a single edge server anda single central cloud. To maximize the total successfuloffloading probability, a heuristic scheduling algorithm wasproposed to leverage both the low communication latency dueto the proximity of the MEC server and the low computationlatency arising from abundant computational resources at thecentral-cloud server. Specifically, when the computation loadof the MEC server exceeds a given threshold, latency-toleranttasks are offloaded to the central cloud to spare enoughcomputational resources at the edge server for processinglatency-sensitive tasks. In addition, [132] explored the problemof server selection over multiple MEC servers. The majorchallenge arises from the correlation between the amountsof the offloaded computation and selected edge servers formultiple users. To cope with this issue, a congestion gamewas formulated and solved to minimize the sum energyconsumption of mobile users and edge servers. Most recently,a computation offloading framework that allows a mobile device to offload tasks to multiple MEC servers was proposedin [133], and semidefinite relaxation-based algorithms wereproposed to determine the task allocation decisions and CPUfrequency scaling. Server Cooperation : Resource sharing via server co-operation can not only improve the resource utilization andincrease the revenue of computing service providers, but alsoprovide more resources for mobile users to enhance theiruser experience. This framework was originally proposed in[134], which includes components such as resource allocation,revenue management and service provider cooperation. First,resource allocation was optimized for cases with deterministicand random user information to maximize the total revenues.Second, considering self-interested cloud service providers, adistributed algorithm based on game theory was proposed tomaximize service providers’ own profits, which was shownto achieve the Nash equilibrium. This study was furtherextended in [135], which considered both the local and remoteresource sharing. The former refers to resource sharing amongdifferent service providers within the same data center, whilethe latter one means the cooperation across different datacenters. To realize the resource sharing and cooperation amongdifferent servers, a coalition game was formulated and solvedby a game-theoretic algorithm with stability and convergenceguarantees. Moreover, the recent work [136] proposed a newserver cooperation scheme where edge servers exploit both thecomputational and storage resources by proactively cachingcomputation results to minimize the computation latency. Thecorresponding task distributing problem was formulated as amatching game and solved by an efficient algorithm based ona proposed deferred-acceptance algorithm. Computation Migration : In [137]–[139], apart fromoptimizing the offloading decisions, the authors also inves-tigated the computation migration among different remoteservers. Specifically, the computation migration over MECservers was motivated by the mobility of offloading users.When a user moves closer to a new MEC server, the networkcontroller can choose to migrate the computation to thisserver, or compute the task in the original server and thenforward the results back to the user via the new server. Thecomputation migration problem was formulated as an MDPproblem based on a random-walk mobility model in [137].It was shown that the optimal policy has a threshold-basedstructure, i.e., the migration should be selected only when thedistance of two servers is bounded by two given thresholds.This work was further extended in [138] where the workloadscheduling in edge servers was integrated with the servicemigration to minimize the average overall transmission andreconfiguration costs using Lyapunov optimization techniques.Another computation migration framework was proposed in[139], where the MEC server can either process offloadedcomputation tasks locally or migrate them to the central cloudserver. An optimization problem was formulated to minimizethe sum mobile-energy consumption and computation latency.This problem was solved by a heuristic two-stage algorithm,which first determines the offloading decision for each userby the semi-definite relaxation and randomization techniques,and then performs the resource allocation optimization for all TABLE VIT
HE COMPARISON OF PAPERS FOCUSING ON
MEC
SYSTEMS WITH HETEROGENEOUS SERVERS . Theme Design Type Design Objective Reference Proposed Solution
Server selection Edge/centralserver selection Successful offload-ing probability [131] Propose a heuristic server selection algorithm according tothe deadline requirementsEdge server selec-tion Energy [132] Formulate a congestion game and propose a distributed algo-rithm admitting the Nash equilibriumMultiple edgeservers Energy and latency [133] Propose semidefinite relaxation-based algorithms for taskallocation decisions and frequency scalingServer cooperation Edge server coop-eration Revenue [134] Propose a distributed resource allocation algorithm admittingthe Nash equilibriumEdge/remoteserver cooperation Utility [135] Formulate a coalition game and propose a game-theoreticalgorithmEdge serverproactive caching Latency [136] Study the distribution and proactive caching of computingtasks in MECComputation Edge server Cost [137] Propose a threshold-based computation migration schemeaccording to the distancemigration migration [138] Propose online workload scheduling and migration algorithmsusing Lyapunov optimization techniquesRemote server mi-gration Energy and latency [139] Propose a heuristic two-stage algorithm including migrationdecision and resource allocation the users.
4) Summary and Insight:
Table VI provides the summaryof resource management schemes for MEC systems withheterogeneous servers. The literature provides a set of insightson server selection, cooperation, and computation migration,described as follows. • Consider MEC systems with multiple computation tasksand heterogeneous servers. To reduce the sum computa-tion latency, it is desirable to offload latency-insensitivebut computation-intensive tasks to remote central cloudserver and latency-sensitive ones to the edge servers. • Server cooperation can significantly improve the compu-tation efficiency and resource utilization at MEC servers.More importantly, it can balance the computation loaddistribution over the networks so as to reduce sum com-putation latency while the resources are better utilized.Moreover, the server cooperation design should con-sider temporal-and-spatial computation task arrivals andserver’s computation capacities, time-varying channels,and servers’ individual revenue. • Computation migration is an effective approach for mo-bility management in MEC. The decision of migrate-or-not depends on the migration overhead, distances betweenusers and servers, channel conditions, and servers’ com-putation capacities. Specifically, when a user moves faraway from its original MEC server, it is preferred tomigrate the computation to nearby servers.
D. Challenges
In the preceding subsections, we have conducted a compre-hensive survey on the state-of-the-art resource managementtechniques for MEC systems. However, the progress is stillin the infancy stage and many critical factors have beenoverlooked for simplicity, which need to be addressed in futureresearch efforts. In the following, we identify three criticalresearch challenges for resource management in MEC thatremain to be solved. Two-Timescale Resource Management : In most exist-ing works, e.g., [87], [88], [96], [119], [121], [140], wirelesschannels were assumed to remain static during the whole taskexecution process for simplicity. Nevertheless, this assumptionmay be unreasonable when the channel coherence time ismuch shorter than the latency requirement. For instance, at acarrier frequency of 2GHz, the channel coherence time can beas small as 2.5ms when the speed is 100km/h. For some mobileapplications such as the MMORPG game PlaneShift , theacceptable response time is 440ms and the excellent latencyis 120ms [141]. In such scenarios, the task offloading processmay be across multiple channel blocks, necessitating the two-timescale resource management for MEC. This problem isvery challenging even for a single-user MEC system withdeterministic task arrivals [81]. Online Task Partitioning : For ease of optimization,existing literature tackling the task partitioning problems ig-nores the fluctuation of the wireless channels, and obtain thetask partitioning decision before the start of the executionprocess. With such an offline task partitioning decision, thechange of the channel condition may lead to inefficient oreven infeasible offloading, which shall severely degrade thecomputation performance. To develop online task partitioningpolicies, one should incorporate the channel statistics into theformulated task partitioning problem, which may easily belongto an NP-hard problem even under a static channel. In [99] and[142], approximate online task partitioning algorithms werederived for applications with serial and tree-topology task-call graphs, respective, while solutions for general task modelsremain unexploited. Large-Scale Optimization : The collaboration of multi-ple MEC servers allows their resources to be jointly managedfor serving a large number of mobile devices simultaneously.However, the increase of the network size renders the re-source management a large-scale optimization problem withrespect to a large number of offloading decisions as well asradio-and-computational resource allocation variables. Con-ventional centralized joint radio-and-computational resource Mobility Management for MEC
1. Mobility-Aware Online Prefetching 2. Mobility-Aware Offloading Using D2D Communications 3. Mobility-Aware Fault-Tolerant MEC 4. Mobility-Aware Server Scheduling
Deployment of MEC Systems
1. Site Selection for MEC Servers 2. MEC Network Architecture 3. Server Density Planning
Cache-Enabled MEC
1. Service Caching for MEC Resource Allocation 2. Data Caching for MEC Data Analytics
Green MEC
1. Dynamic Right-Sizing for Energy-Proportional MEC 2. Geographical Load Balancing for MEC 3. Renewable Energy-Powered MEC Systems
Security and Privacy Issues in MEC
1. Trust and Authentication Mechanisms 2. Networking Security 3. Secure and Private Computation Future Research Directions for MEC
Fig. 7. Future research directions for MEC. management algorithms require a huge amount of informationand computation when applied to large-scale MEC systems,which will inevitably incur a significant execution delay andmay whittle away the potential performance improvement,e.g., latency reduction, brought by the MEC paradigm. Toachieve efficient resource management, it is required to designdistributed low-complexity large-scale optimization algorithmswith light signaling and computation overhead. Althoughthe recent advancements in large-scale convex optimization[143] provide powerful tools for radio resource management,they cannot be directly applied to optimize the computationoffloading decision due to its combinatorial and non-convexnature, which calls for new algorithmic techniques.IV. I
SSUES , C
HALLENGES , AND F UTURE R ESEARCH D IRECTIONS
Recent years have witnessed substantial research efforts onresource management for MEC as surveyed in the precedingsection. However, there are lots of emerging research direc-tions of MEC that are still largely uncharted. In this section,technical issues, challenges and research opportunities will beidentified and discussed as summarized in Fig. 7, includingthe large-scale MEC system deployment, cache-enabled MEC,mobility management, green MEC and security-and-privacyissues in MEC.
A. Deployment of MEC Systems
The primary motivation of MEC is to shift the CloudComputing capability to the network edges in order to reduce the latency caused by congestion and propagation delays inthe core network. However, there is no formal definition ofwhat an MEC server should be, and the server locationsin the system are not specified. These invoke the site se-lection problems for MEC servers, which are significantlydifferent from the conventional BS site selection problems,as the optimal placement of edge servers is coupled with thecomputational resource provisioning, and both of them areconstrained by the deployment budget. Besides, the efficiencyof an MEC system relies heavily on its architecture, whichshould account for various aspects such as workload intensityand communication rate statistics. In addition, it is criticalfor MEC vendors to determine the required server densityfor catering the service demand, which is closely related tothe infrastructure deployment cost and marketing strategies.Nonetheless, the large-scale nature of MEC systems makestraditional simulation-based methods inapplicable, and thussolutions based on network-scale analysis are preferred. In thissubsection, we will discuss three research problems relatedto MEC deployment, including the site selection for MECservers, the MEC network architecture, and server densityplanning. Site Selection for MEC Servers : Selecting the sitesfor MEC infrastructures, especially MEC servers, is the firststep towards building up the MEC system. To make thecost-effective server-site selection, the system planners andadministrators should account for two important factors: siterentals and computation demands. In general, given the system deployment budget, more MEC servers should be installed atregions with higher computation demands, such as businessdistricts, commercial areas and densely populated areas. This,however, contradicts the cost requirement as such areas arelikely to have high site rentals. Fortunately, thanks to the well-deployed telecom networks, it is a promising idea to install theMEC servers co-located with the existing infrastructures suchas macro BSs, which is even more attractive for the telecomoperators who would like to participate in the MEC market.However, this would not solve all the problems. On onehand, due to the ever-increasing computation-quality require-ment and ubiquitous smart devices, satisfactory user experi-ence cannot be guaranteed due to the poor signal quality andcongestion in the macro cells. For some applications, e.g.,smart home [144], it is desirable to move the computationcapability even closer to the end users. This can be achievedby injecting some computational resources at small-cell BSs[72], [73], which are low-cost and small-size BSs. Despite thepotential benefits, there are still obstacles on the way: • First, due to physical limitations, the computation ca-pabilities of such kind of MEC servers will be muchsmaller than those at macro BSs, making it challenging tohandle computation-intensive tasks. One feasible solutionis to build a hierarchical network architecture for MECsystems comprising MEC servers with heterogeneouscommunication-and-computation capabilities as detailedin the sequel. • Second, some of the small-cell BSs may be self-deployedby the home users, and many femto BS owners may nothave the motivation to collaborate with MEC vendors.To overcome this issue, MEC vendors need to designa proper incentive mechanism in order to stimulate theowners of small-cell BSs for renting the sites. • Moreover, deploying MEC servers at small-cell BSs mayincur security problems as they are easy-to-reach andvulnerable to external attacks, which shall degrade thelevels of reliability.On the other hand, the computation hot spots do not alwayscoincide with the communication hot spots. In other words, forsome of the computation hot spots, there exists no availablecommunication infrastructure (either macro or small-cell BS).For these circumstances, we need to deploy edge servers withwireless transceivers by properly choosing new locations.Besides, the site selection for MEC servers is dependenton the computational resource-allocation strategy, which posesextra challenges compared to the conventional BS site selec-tion. Intuitively, concentrating the computational resources ata few MEC servers can help save the site rentals. However,this comes at the prices of potential degradation of the servicecoverage and communication quality. In addition, the optimalcomputational resource allocation should take into accountboth site rentals and computation demands. For example, foran MEC server at a site with a high site rental, it is preferredto allocate huge computational resource and thus serve a largenumber of users, for achieving the high revenue. Hence, ajoint site selection and computational resource provisioningproblem needs to be solved before deploying MEC systems. (cid:39)(cid:68)(cid:87)(cid:68)(cid:3)(cid:38)(cid:72)(cid:81)(cid:87)(cid:72)(cid:85) (cid:47)(cid:55)(cid:40)(cid:3)(cid:37)(cid:54)(cid:54)(cid:80)(cid:68)(cid:79)(cid:79)(cid:16)(cid:70)(cid:72)(cid:79)(cid:79)(cid:3)(cid:37)(cid:54)(cid:58)(cid:76)(cid:41)(cid:76)(cid:3)(cid:53)(cid:82)(cid:88)(cid:87)(cid:72)(cid:85) (cid:36)(cid:71)(cid:16)(cid:75)(cid:82)(cid:70)(cid:3)(cid:38)(cid:79)(cid:82)(cid:88)(cid:71) (cid:39)(cid:82)(cid:90)(cid:81)(cid:79)(cid:82)(cid:68)(cid:71)(cid:56)(cid:83)(cid:79)(cid:82)(cid:68)(cid:71) (cid:55)(cid:76)(cid:72)(cid:85)(cid:16)(cid:20)(cid:55)(cid:76)(cid:72)(cid:85)(cid:16)(cid:21)(cid:55)(cid:76)(cid:72)(cid:85)(cid:16)(cid:22)(cid:40)(cid:81)(cid:71)(cid:3)(cid:56)(cid:86)(cid:72)(cid:85)(cid:86)
Fig. 8. A 3-tier heterogeneous MEC system. Tier-1 servers are located in closeproximity to the end users, such as at WiFi routers and small-cell BSs, whichare of relatively small computation capabilities. Tier-2 servers are deployedat LTE BSs with moderate computation capabilities. Tier-3 servers are theexisting Cloud Computing infrastructures, such as data centers. MEC Network Architecture : The promotion of MECdoes not mean the extinction of the data-center networks (DCNs). Instead, future mobile computing networks are en-visioned to be consisted of three layers as shown in Fig. 8,i.e., cloud, edge (a.k.a. fog layer), and the service subscriberlayer [130], [145]. While the cloud layer is mature and well-deployed, there is still some flexibility and uncertainty indesigning the edge layer.By analogy to the heterogeneous networks (HetNets) incellular systems, it is intuitive to design the Het-MEC systems,which consist of multiple tiers. Specifically, the MEC serversin different tiers have distinct computation and communicationcapabilities. Such kinds of hierarchical MEC system structurescan not only preserve the advantage of efficient transmissionoffered by HetNets, but also possess strong ability to handlethe peak computation workloads by distributing them acrossdifferent tiers [146]. However, the computation capacity provi-sioning problem is highly challenging and remains unsolved,as it should account for many different factors, such asthe workload intensity, communication cost between differenttiers, workload distribution strategies, etc.Another thrust of research efforts focuses on exploitingthe potential of the service subscriber layer, and utilizingthe undedicated computational resources, e.g., laptops, smartphones, and vehicles, overlaid with dedicated edge nodes.This paradigm is termed as the
Ad-hoc mobile cloud inliterature [147]–[150]. The ad-hoc mobile cloud enjoys thebenefits of amortizing the stress of MEC systems, increasingthe utilization of the computational resources, and reducingthe deployment cost. However, it also brings difficulties inresource management and security issues due to its ad-hocand self-organized nature. Server Density Planning : As mentioned in Sec-tion IV-A2, the MEC infrastructure may be a combination −500 −400 −300 −200 −100 0 100 200 300 400 500−500−400−300−200−1000100200300400500 Meter M e t e r MEC serverMobile device Building A Building BBuilding C Building D
Fig. 9. Illustration of the clustering behavior of the computation demands.The mobile devices requesting for MEC services will be more concentratedaround the MEC servers. of different types of edge servers, which provides variouslevels of computation experience and contributes differentdeployment costs. Hence, it is critical to determine the numberof edge nodes as well as the optimal combination of differenttypes of MEC servers with a given deployment budget andcomputation demand statistics. Conventionally, this problemcan only be addressed by numerical simulations, which istime-consuming and has poor scalability. Fortunately, owingto the recent development of stochastic geometry theory andits successful applications in performance analysis for wirelessnetworks [151]–[154], as well as the similarity between Het-MEC systems and HetNets, it is feasible to conduct per-formance analysis for MEC systems using techniques fromstochastic geometry theory. Such analysis of MEC systemsshould address the following challenges: 1) The timescalesof computation and wireless channel coherence time maybe different [81], [104], which makes existing results forwireless networks not readily applicable for MEC systems.One possible solution is to combine the Markov chain andstochastic geometry theories to capture the steady behaviorof computations. 2) The computation offloading policy willaffect the radio resource management policy, which shouldbe taken into consideration. 3) The computation demands arenormally non-uniformly distributed and clustered (see Fig. 9),prohibiting the use of the homogeneous Poisson point process (HPPP) model for edge servers and service subscribers. It thuscalls for the investigation of more advanced point processes,e.g., the Ginibre α - determinantal point process (DPP), tocapture the clustering behaviors of edge nodes [155]. B. Cache-Enabled MEC
It has been predicted by Cisco that mobile video streamingwill occupy up to 72% of the entire mobile data traffic by 2019[156]. One unique property of such services is that the contentrequests are highly concentrated and some popular contentswill be asynchronously and repeatedly requested. Motivated
Fig. 10. Cache-enabled MEC systems. by this fact, wireless content caching or FemtoCaching wasproposed in [157]–[160] to avoid frequent replication for thesame contents by caching them at BSs. This technology hasattracted extensive attention from both academia and industrydue to its striking advantages on reducing content acquisitionlatency, as well as relieving heavy overhead burden of thenetwork backhaul. While caching is to move popular contentsclose to end users, MEC is to deploy edge servers to handlecomputation-intensive tasks for edge users to enhance userexperience. Note that these two techniques seem to targetfor diverse research directions, i.e., one for popular contentdelivery and the other for individual computation offloading.However, they will be integrated seamlessly in this subsectionand envisioned to create a new research area, namely, the cache-enabled MEC .Consider the novel cache-enabled MEC system shown inFig. 10. In such systems, the MEC server can cache severalapplication services and their related database, called servicecaching (or service placement [161]) and data caching , respec-tively, and handle the offloaded computation from multipleusers. To efficiently reduce the computation latency, severalkey and interesting problems need to be solved, which aredescribed in the following with potential solutions. Service Caching for MEC Resource Allocation : Unlikethe central cloud server that is always assumed with huge anddiverse resources (e.g., computing, memory and storage), thecurrent edge server has much less resources, making it unableto accommodate all users’ computation requests. On the otherhand, different mobile services require different resources,based on which, they can be classified into CPU-hungry (e.g.,cloud chess and VR), memory-hungry (e.g., online Matlab),and storage-hungry (e.g., VR) applications. Such a mismatchbetween resource and demand introduces a key challenge onhow to allocate heterogeneous resources for service caching.Note that similar problems have been investigated in con-ventional Cloud Computing systems [162]–[165], termed as
VM placement , as well as MCC systems [161]. Specifically, theauthors in [162] proposed a novel architecture for VM manage-ment and optimized the VM placement over multiple cloudsto reduce the deployment costs and improve user experience, given constraints on hardware configuration, the number ofVMs as well as load balancing. Similar VM-placement prob-lems were also investigated in [163], [164] for maximizing theenergy savings of cloud servers and in [165] for different cloudscheduling strategies. Recently, the authors in [161] extendedthe VM placement idea to MCC systems and studied thejoint optimization of service caching/placement over multipleclouds and load dispatching for end users’ requests. As aresult, one efficient algorithm was proposed to minimize boththe computation latency and service placement transition cost.These works, however, cannot be directly applied to designefficient service caching policies for MEC systems, since itshould take into account more refined information includingusers’ location, preference, experience as well as edge servers’capacities in terms of the memory, storage and VM instance.To this end, two possible approaches are described as follows.The first one is spatial popularity-driven service caching ,referring to caching different combinations and amounts ofservices in different MEC servers according to their specificlocations and surrounding users’ common interests. This ideais motivated by the fact that users in one small region are likelyto request similar computing services. For example, visitors ina museum tend to use AR for better sensational experience.Thus, it is desirable to cache multiple AR services at the MECserver of this region for providing the real-time service. Toachieve the optimal spatial service caching, it is essential toconstruct a spatial-application popularity distribution model for characterizing the popularity of each application overdifferent locations. Based on this, we can design resource-allocation policies using various optimization algorithms, e.g.,the game theory and convex optimization techniques.An alternative approach is temporal popularity-driven ser-vice caching . The main idea is similar to that of the spatialcounterpart, but it exploits the popularity information in thetemporal domain, since the computation requests also dependon the time period. One example is that users are apt to playmobile cloud gaming after dinner. This kind of informationwill suggest MEC operators to cache several gaming servicesduring this typical period for handling the huge computationloads. One disadvantage of this temporal-based approach isthe additional server cost resulted from frequent cache-and-tear operations since popularity information is time-varyingand MEC servers possess finite resources. Data Caching for MEC Data Analytics : Many modernmobile applications involve intensive computation based ondata analytics, e.g., ranking and classification. Take VR as aninstance. It creates an imaginary environment similar to thereal world by generating realistic images, sounds and othersensations for enhancing users’ experience. Achieving this endis nontrivial as it requires the MEC server to finish multiplecomplicated processes within the ultra-short duration (e.g.,1ms), such as recognizing users’ actions via pattern recog-nition, “understanding” users’ requests via data mining, aswell as rendering virtual settings via video streaming or othersensation techniques [166]. All the above data-analytics basedtechniques should be supported by comprehensive database,which, however, imposes extremely heavy burden on the edgeserver storage. This challenge can be relieved by intelligent data caching that only reserves frequently-used database. Fromanother perspective, caching parts of computation-result datathat is likely to be reused by others can further boost thecomputation performance of the entire MEC system. Onetypical example is mobile cloud gaming, which enables fastand energy-efficient gaming by shifting game computingengines from mobiles to edge servers and supporting real-time gaming by game video streaming. Thus, it emerges asa leading technique for next generation mobile computinginfrastructures [167]. Since certain game rendered videos, e.g.,gaming scenes, can be reused by other players, caching thesecomputation results would not only significantly reduce thecomputation latency of the players with the same computationrequest, but also ease the computation burden for edge servers.Similar idea has been proposed in [168], which investigatedcollaborative multi-bitrate video caching and processing inMEC.For MEC data caching at a single edge server, one keyproblem is how to balance the tradeoff between massivedatabase and finite storage capacity . Unlike FemtoCachingnetworks where content (data) caching mainly introducesa new multiple-access mechanism termed as cache-enabledaccess [169], data caching in MEC systems brings aboutmanifold effects on the computation accuracy, latency andedge server-energy consumption, which, however, have notbeen characterized in existing literature. This calls for modelbuilding research efforts for accurately quantifying the men-tioned effects for various MEC applications. Furthermore, itis also essential to establish a practical database popularitydistribution model that is able to statistically characterize theusage of each database set for different MEC applications.Based on the above models, the said tradeoff can be achievedby solving an optimization problem that maximizes the achiev-able QoS and minimizes the storage cost in MEC systemssimultaneously.The above framework can be further extended to MECsystems with multiple servers where each server can servemultiple users and each user can offload computation to mul-tiple edge servers. The fundamental problem is similar to thatof the cache-enabled HetNets [170], that is, how to spatially distribute the database over heterogeneous edge servers underboth storage and computation-load constraints on each ofthem, for increasing network-wide revenue. Intuitively, foreach MEC server, it is desirable to spare more storage to cachethe database of the most popular applications in its cell, andit also needs to utilize partial storage to accommodate lesspopular ones, whose computation performance will be furtherimproved by cooperative caching in different MEC servers.Moreover, the performance of large-scale cache-enabled MECnetworks can be analyzed using stochastic geometry by mod-eling nearby users as clusters [171].
C. Mobility Management for MEC
Mobility is an intrinsic trait of many MEC applications,such as VR assisted museum tour to enhance experience ofvisitors. In these applications, the movement and trajectory ofusers provide location and personal preference information for Mobile Device’ Trajectory
Fig. 11. Mobility management for MEC. the edge servers to improve the efficiency of handling users’computation requests. On the other hand, mobility also posessignificant challenges for realizing ubiquitous and reliablecomputing (i.e., without interruptions and errors) due to thefollowing reasons. First, MEC will be typically implementedin the HetNet architecture comprising of multiple macro,small-cell BSs and WiFi APs. Thus, users’ movement willcall for frequent handovers among the small-coverage edgeservers as shown in Fig. 11, which is highly complicateddue to the diverse system configurations and user-server as-sociation policies. Next, users moving among different cellswill incur severe interference and pilot contamination, whichshall greatly degrade the communication performance. Last,frequent handovers will increase the computation latency andthus deteriorate users’ experience.Mobility management has been extensively studied for tra-ditional heterogeneous cellular networks [172]–[174]. In theseprior works, users’ mobility is modeled by the connectivityprobability or the link reliability according to such informationas the users’ moving speeds. Based on such models, dynamicmobility management has been proposed to achieve high datarate and low bit-error rate. However, these policies cannotbe directly applied for MEC systems with moving users,since they neglect the effects of the computation resourcesat edge servers on the handover policies. Recent works in[175]–[178] have made initial efforts to design mobility-awareMEC systems. Specifically, the inter-contact time and contactrate were defined in [175] to model users’ mobility. Anopportunistic offloading policy was then designed by solvinga convex optimization problem for maximizing the successfultask offloading probability. Alternatively, to account for themobility, the number of edge servers that users can access wasmodeled by an HPPP in [176]. Then, the offloading decisionwas optimized by addressing the formulated MDP problemto minimize the offloading cost including mobile-energy con-sumption, latency and failure penalty. Other mobility modelswere also proposed in [177], [178], which characterize themobility by a sequence of networks that users can connect toand a two-dimensional location-time workflow, respectively.In addition, mobility management for MEC was integratedwith traffic control in [179] to provide better experience for users with latency-tolerant tasks via designing intelligent cellassociation mechanisms. In [160], edge caching was integratedwith mobility prediction in Follow-Me Cloud for enhancingthe content-caches migration located at the edges. Recentproposals on mobility-aware wireless caching in [180] alsoprovided valuable guidelines on mobility management in MECsystems.Note that most of the existing works focused on optimizingmobility-aware server selection. However, to achieve betteruser experience and higher network-wide profit, the offloadingtechniques at mobile devices and scheduling policies at MECservers should be jointly considered. This introduces a setof interesting research opportunities with some described asfollows. Mobility-Aware Online Prefetching : In practice, thefull information of the user trajectory may be unavailable.Conventional design for mobile computation offloading willfetch a computation task to another server only when it ishandoverred. This mechanism requires excessive fetching ofa large volume of data for handover and thus brings longfetching latency. Moreover, it also causes heavy loads on theMEC network. One promising solution to handle this issue isto leverage the statistical information of the user trajectory andprefetch parts of future computation data to potential serversduring the server-computation time, referred to as onlineprefetching [181]. This technique can not only significantlyreduce the handover latency via mobility prediction, but alsoenable energy-efficient computation offloading by enlargingthe transmission time. However, it also encounters severalchallenges with two most critical ones described as follows.The first challenge arises from the trajectory prediction. Ac-curate prediction can allow seamless handovers among edgeservers and reduce the prefetching redundancy. Achieving it,however, requires precise modeling and high-complexity MLtechniques, e.g., Bayesian, reinforcement and deep learning.For example, the trajectory of a typical visitor in a museumcan be predicted according to his own interest-informationand statistical route-information of some previous visitorswith similar interests that can be obtained by ML algorithms.Therefore, it is important to balance the tradeoff between themodeling accuracy and computation complexity. The second challenge lies in the selection of the prefetched computationdata. To maximize the successful offloading probability ofedge users, the computation-intensive components should beprefetched earlier with adaptive transmission power control indynamic fading channels. Mobility-Aware Offloading Using D2D Communica-tions : D2D communications was first proposed in [182] toimprove the network capacity and alleviate the data trafficburden in cellular systems. This paradigm can also be usedto handle the user mobility problems in MEC systems [123],which creates numerous D2D communication links. Theselinks allow the computation of a user to be offloaded to itsnearby users which have more powerful computation capabil-ities. The short-range communication offered by D2D linksreduces energy consumption of data transmission as well.However, user mobility brings new design issues as follows.The first one is how to exploit the advantages of both D2Dand cellular communications. One possible approach is tooffload the computation-intensive data to the edge serversat BSs that have huge computation capabilities in order toreduce the server-computing time; while the components oflarge data sizes and strict computation requirements shouldbe fetched to nearby users via D2D communications forhigher energy efficiency. Next, the selection of surroundingusers for offloading should be optimized to account for users’mobility information, dynamic channels and heterogeneoususers’ computation capabilities. Last, massive D2D links willintroduce severe interference for reliable communications.This issue is more complicated in the mobility-based MECsystems due to the fast-changing wireless fading environments.Hence, advanced interference cancellation and cognitive radiotechniques can be applied for MEC systems, together withmobility prediction to increase the offloading rate and reducethe service latency. Mobility-Aware Fault-Tolerant MEC : User mobilityposes significant challenges for providing reliable MEC ser-vices due to dynamic environments. Computation offloadingmay fail due to intermittent connections and rapid-changingwireless channels. The induced failure is catastrophic forthe latency-sensitive and resource-demanding applications. Forinstance, AR-based museum video guide aims to provide fluentand fancy virtual sensations for visitors, and the disruptionor failure of video streaming due to intermittent connectionswould upset visitors. Another example is the military operationwhich always requires fast and ultra-reliable computation,even in high-mobility environments. Any computation failurewould bring serious consequences. These facts necessitate thedesign for mobility-aware fault-tolerant MEC systems [183]–[185], with three major and interesting problems illustrated asfollows, including fault prevention, fault detection and faultrecovery. Fault prevention is to avoid or prevent MEC faultby backing up extra stable offloading links. Macro BSs orcentral clouds can be chosen as protection-clouds, since theyhave large network coverage that allows continuous MECservice. The key design challenges lie in how to balance thetradeoff between QoS (i.e., the failure probability) and energyconsumption due to extra offloading links for the single-user case, and how to allocate protection-clouds for multiuser MEC applications. Next, fault detection is to collect faultinformation, which can be realized by setting intelligent timingchecks or receiving feedbacks for MEC services. In addition,channel and mobility estimation techniques can also be appliedto estimate the fault so as to reduce the detection time.Last, for detected MEC faults, recovery approaches should beperformed to continue and accelerate the MEC service. Thesuspended service can be switched to more reliable backupwireless links with adaptive power control for higher-speedoffloading. Alternative recovery approaches include migratingthe workloads to neighboring MEC systems directly or throughad-hoc relay nodes as proposed in [185]. Mobility-Aware Server Scheduling : For multiuser MECsystems, traditional MEC server scheduling servers usersaccording to the offloading priority order that depends onusers’ distinct local computing information, channel gainsand latency requirements [84]. However, this static schedulingdesign cannot be directly applied for the multiuser MECsystems with mobility due to dynamic environments, e.g.,time-varying channels and intermittent connectivities. Suchdynamics motivate the design of adaptive server schedulingthat regenerates the scheduling order from time to time,incorporating the real-time user information. In such adaptivescheduling mechanisms, users with worse conditions willbe allocated with higher offloading priorities to meet theircomputing deadlines. Another potential approach is to designmobility-aware offloading priority function by the followingtwo steps. The first step is to accurately predict users’ mobilityprofiles and channels, where the major challenge is howto reflect the mobility effects and re-define the offloadingpriority function. The second step is resource reservation thatcan enhance the server scheduling performance [186], [187].Specifically, to guarantee the QoS of latency-sensitive andhigh-mobility users, MEC servers can reserve some dedicatedcomputational resources and provide reliable computing ser-vice for such users. While for other latency-tolerant users, theMEC server can perform on-demand provisioning. For such ahybrid MEC server provisioning scheme, the server schedulingcan be optimized for serving the maximum number of userswith QoS guarantees, as well as maximizing MEC servers’revenue.
D. Green MEC
MEC servers are small-scale data centers, each of whichconsumes substantially less energy than the conventional clouddata center. However, their dense deployment pattern raises abig concern on the system-wide energy consumption. There-fore, it is unquestionably important to develop innovative tech-niques for achieving green MEC [188], [189]. Unfortunately,designing green MEC is much more challenging comparedto green communication systems or green DCNs. Comparedto green communication systems, the computational resourceneeds to be managed to guarantee satisfactory computationperformance, making the traditional green radio techniques notreadily applicable. On the other hand, the previous researchefforts on green DCNs have not considered the radio resourcemanagement, which makes them not suitable for green MEC. Besides, the highly unpredictable computation workload pat-tern in MEC servers poses another big challenge for resourcemanagement in MEC systems, calling for advanced estimationand optimization techniques. In this subsection, we will intro-duce different approaches on designing green MEC systems,including dynamic right-sizing for energy-proportional MEC, geographical load balancing (GLB) for MEC, and MECsystems powered by renewable energy. Dynamic Right-Sizing for Energy-Proportional MEC : The energy consumption of an MEC server highly dependsthe utilization radio [see Eq. (5)]. Even when the server isidling, it still consumes around 70% of the energy as itoperates at the full speed. This fact motivates the designof energy-proportional (or power-proportional ) servers, i.e.,the energy consumption of a server should be proportionalto its computation load [190]. One way to realize energy-proportional servers is to switch off/slow down the processingspeeds of some edge servers with light computation loads.Such an operation is termed as dynamic right-sizing in theliterature on green DCNs [191]. However, along with thepotential energy savings, toggling servers between the activeand sleep modes could bring detrimental effects. First of all,it will incur the switching energy cost and application data-migration latency. Also, user experience may be degradeddue to the less amount of allocated computational resources,which may, in turn, reduce the operator’s revenue. Besides,the risk associated with server toggling as well as the wear-and-tear cost of the servers might be increased, which can inturn increase the maintenance costs of MEC vendors. As aresult, switching off the edge servers in a myopic manner isnot always beneficial.In order to make an effective decision on dynamic right-sizing, the profile of computation workload at each edge servershould be accurately forecasted. In conventional DCNs, thiscan be achieved rather easily as the workload at each datacenter is an aggregation of the computation requests acrossa large physical region, e.g., several states in the UnitedStates, which is relatively stable so that it can be estimatedby referring to the readily available historical data at thedata centers. However, for MEC systems, the serving areaof each edge server is much smaller, and hence its workloadpattern is affected by many factors, such as the location of theserver, time, weather, the number of nearby edge servers, anduser mobility. This leads to a fast-changing workload pattern,and requires more advanced prediction techniques. Moreover,online dynamic right-sizing algorithms that require less futureinformation need to be developed. Geographical Load Balancing for MEC : GLB is an-other key technique for green DCNs [192], [193], whichleverages the spatial diversities of the workload patterns,temperatures, and electricity prices, to make workload routingdecision among different data centers. This technique can alsobe applied to MEC systems. For instance, a cluster of MECservers can coordinate together to serve a mobile user, i.e.,the tasks can be routed from the edge server located in ahot spot (such as a restaurant) to a nearby edge server withlight workload (such as the one in a park). On one hand, thishelps to improve the energy efficiency of the lightly-loaded edge servers as well as user experience. On the other hand, itcan prolong the battery lives of mobile devices, as offloadingthe tasks through the nearby server could save transmissionenergy. It is worthwhile to note that the implementation ofGLB requires efficient resource management techniques atedge servers, such as dynamic right-sizing and VM manage-ment [194]–[197].Meanwhile, there are many factors to be incorporated whenapplying GLB in MEC environments. Firstly, since the mi-grated tasks should go through the cellular core network, thenetwork congestion state should be monitored and consideredwhen making the GLB decisions. Secondly, to enable seamlesstask migration, a VM should be migrated/set up in anotheredge server beforehand, which may cause additional energyconsumption. Thirdly, the mutual interests of MEC operatorsand edge computing service subscribers should be carefullyconsidered when performing GLB, due to the tradeoff betweenthe energy savings and latency reduction. Last but not least,the existence of conventional Cloud Computing infrastructuresendows the edge servers with an extra option of offloadingthe latency-critical and computation-intensive tasks to remotecloud data centers, creating a new design dimension andfurther complicating the optimization. Renewable Energy-Powered MEC Systems : Traditionalgrid energy is normally generated by coal-fired power plants.Hence, powering mobile systems with grid energy inevitablycauses a huge amount of carbon emission, which opposes thetarget of green computing. Off-grid renewable energy, suchas solar radiation and wind energy, recently, has emerged asa viable and promising power source for various IT systemsthanks to the recent advancements of energy harvesting (EH)techniques [198], [199]. This fact motivates the design ofinnovative MEC systems, called renewable energy-poweredMEC systems, which are shown in Fig. 12 comprising bothEH-powered MEC servers and mobile devices. On one hand,as the MEC servers are expected to be densely-deployedand have low power consumption similar to that of small-cell BSs [200], it is reasonable and feasible to power theMEC infrastructures with the state-of-the-art EH techniques.On the other hand, the mobile devices can also get benefitsfrom using renewable energy as EH is able to prolong theirbattery lives, which is one of the most favorable features formobile phones [201]. Besides, the use of renewable energysources eliminates the need of human intervention such asreplacing/recharging the batteries, which is difficult if notimpossible for certain types of application scenarios wherethe devices are hard and dangerous to reach. Meanwhile, theseadvantages of using renewable energy are accompanied withnew design challenges.A fundamental problem to be addressed for renewableenergy-powered MEC systems is the green energy-awareresource allocation and computation offloading . Instead ofminimizing the energy consumption subject to satisfactoryuser experience, the design principle for the renewable energy-powered MEC systems should be changed to optimizing theachievable performance given the renewable energy constraint,as the renewable energy almost comes for free. Also, withrenewable energy supplies, the energy side information (ESI), Fig. 12. Renewable energy-powered MEC systems. which indicates the amount of available renewable energy, willplay a key role in the decision making. Initial investigationson renewable energy-powered MEC systems were conductedin [202] and [203], which focused on EH-powered MECservers and EH-powered mobile devices, respectively. For EH-powered MEC servers, the system operator should decidethe amount of workload required to be offloaded from theedge server to the central cloud, as well as the processingspeed of the edge server, according to the information ofthe core network congestion state, computation workload, andESI. This problem was solved by a learning-based onlinealgorithm in [202]. While for EH-powered mobile devices, adynamic computation offloading policy has been proposed in[203] using Lyapunov optimization techniques based on boththe CSI and ESI. However, these two works only consideredsmall-scale MEC systems that consist of either one edgeserver (in [202]) or one mobile device (in [203]). Thus, theycannot provide a comprehensive solution for large-scale MECsystems.For large-scale MEC systems where multiple MEC serversare deployed across a large geographic region, the conceptof GLB could be modified as the green energy-aware GLB to optimize the MEC systems by further utilizing the spatialdiversity of the available renewable energy. This idea wasoriginally proposed for green DCNs, where the “follow therenewables” routing scheme offers a huge opportunity inreducing the grid energy consumption [192], [204]–[207].Moreover, as mentioned before, there exist significant differ-ences between MEC systems and conventional DCNs in termsof the wireless channel fluctuation and resource-managementdesign freedom of system operators. These factors make theoffloading decision making for the green energy-aware GLB inMEC systems much more complicated, as it needs to considerthe CSI and ESI in the whole system.The randomness of renewable energy may introduce theoffloading unreliability and risks of failure, bringing abouta major concern for using renewable energy to power MECsystems. Fortunately, there are several potential solutions tocircumvent this issue as described below. • First, thanks to the low deployment cost, renewable energy-powered edge servers can be densely deployedover the system to provide more offloading opportunitiesfor the users. The resultant overlapping serving areasoffer the offloading diversity in the available energy toavoid performance degradation. A similar idea has beenproposed for EH cooperative communication systems in[208]. • Second, the chance of energy shortage can be reducedby properly selecting the renewable energy sources. Itwas found in [192] that solar energy is more suitablefor workloads with a high peak-to-mean ratio (PMR),while wind energy fits better for workloads with a smallPMR. This provides guidelines for renewable energyprovisioning for edge servers. • Third, MEC servers can be powered by hybrid energysources to improve reliability [209]–[211], i.e., poweredby both the electric grid and the harvested energy. Also,equipping uninterrupted power supply (UPS) units at theedge servers can provide a short period of stable energysupply when green energy is in deficit, and it can berecharged when the surrounding energy condition returnsto a good state. • Moreover, wireless power transfer (WPT), which chargesmobile devices using RF wave [212], [213], is a newly-emerged solution that enables wireless charging andextends the battery life. This technique has been providedin modern mobile phones such as Samsung Galaxy S6.In renewable energy-powered MEC systems, the edgeservers can be powered by WPT when the renewableenergy is insufficient for reliability [214]. This technologyalso applies to the computation offloading for mobiledevices in MEC systems [83] and data offloading forcollaborate mobile clouds [215]. However, novel en-ergy beamforming techniques are needed to increase thecharging efficiency. Moreover, due to the double near-far problem in wireless powered systems, it requires adelicate scheduling to guarantee fairness among multiplemobile devices. E. Security and Privacy Issues in MEC
There are increasing demands for secure and privacy-preserving mobile services. While MEC enables new typesof services, its unique features also bring new security andprivacy issues. First of all, the innate heterogeneity of MECsystems makes the conventional trust and authentication mech-anisms inapplicable. Second, the diversity of communicationtechnologies that support MEC and the software nature ofthe networking management mechanisms bring new securitythreats. Besides, secure and private computation mechanismsbecome highly desirable as the edge servers may be aneavesdropper or an attacker. These motivate us to developeffective mechanisms as described in the following. Trust and Authentication Mechanisms : Trust is animportant security mechanism in almost every mobile system,behind which, the basic idea is to know the identity of the entitythat the system is interacting with . Authentication managementprovides a possible solution to ensure “trust” [216]. However,the inherent heterogeneity of MEC systems, i.e., differenttypes of edge servers may be deployed by multiple vendorsand different kinds of mobile devices coexist, makes theconventional trust and authentication mechanisms designedfor Cloud Computing systems inapplicable. For example, thereputation-based trust model will lead to severe trust threatsin MEC systems, as demonstrated in [217]. This fact callsfor a unified trust and authentication mechanism that is ableto assess the reliability of edge servers and identify thecamouflaged edge servers. Besides, within the mobile network,there will be a large number of edge servers serving mas-sive mobile devices. This makes the trust and authenticationmechanism design much more complicated compared withthat in conventional Cloud Computing systems, since edgeservers are of small computation capabilities and designed toenable latency-sensitive applications. Therefore, it is criticalto minimize the overhead of authentication mechanisms anddesign distributed policies [218], [219]. Networking Security : The communication technologiesto support MEC systems, e.g., WiFi, LTE and 5G, have theirown security protocols to protect the system from attacks andintrusions. However, these protocols inevitably create differenttrust domains. The first challenge of networking security inMEC systems comes from the difficulties in the distributionof credentials, which can be used to negotiate session keysamong different trust domains [216]. In existing solutions,the certification authority can only distribute the credentialsto all the elements located within its own trust domain [216],making it hard to guarantee the privacy and data integrity forcommunications among different trust domains. To addressthis problem, we can use the cryptographic attributes ascredentials in order to exchange session keys [220], [221].Also, the concept of federated content networks, which defineshow multiple trust domains can negotiate and maintain inter-domain credentials [222], can be utilized.Besides, techniques such as SDN and NFV are introducedto MEC systems to simplify the networking management aswell as to provide isolation [5]. However, these techniquesare softwares by nature and thus vulnerable [223], [224]. Moreover, the large number of devices and entities in MECsystems increase the chance of successfully attacking a singledevice, which provides means to launch an attack to thewhole system [225]. Therefore, novel and robust securitymechanisms, such as hypervisor introspection, run-time mem-ory analysis, and centralized security management [226], areneeded to guarantee a secured networking environment forMEC systems. Secure and Private Computation : Migratingcomputation-intensive applications to the edge servers isthe most important function and motivation of building MECsystems. In practice, the task input data commonly containssensitive and private information such as personal clinicaldata and business financial records. Therefore, such datashould be properly pre-processed before being offloaded toedge servers, especially the untrusted ones, in order to avoidinformation leakage. In addition to information leakage,the edge servers may return inaccurate and even incorrectcomputation results due to either software bugs or financialincentives, especially for tasks with huge computationdemands [227]. To achieve secure and private computation,it is highly preferred that the edge platforms can execute thecomputation tasks without the need of knowing the originaluser data and the correctness of the computation results canbe verified, which can be realized by encryption algorithmsand verifiable computing techniques [228]. An interestingexample of secure computation mechanisms for LP problemswas developed in [227], where the LP problem is decomposedinto the public-owned solvers and the private-owned data.By using a privacy-preserving transformation, the customeroffloads the encrypted private data for cloud execution, andthe server returns the results for the transformed LP problem.A set of necessary and sufficient conditions for verifying thecorrectness of the results were developed based on dualitytheory. Upon receiving the correct result, the clients canmap back the desired solution for the original problem usingthe secret transformation. This method of result validationachieves a big improvement in computation efficiency viahigh-level LP computation compared to the generic circuitrepresentation, and it incurs close-to-zero additional overheadon both the client and cloud server, which provides hints todevelop secure and private computation mechanisms for othercloud applications.V. S
TANDARDIZATION E FFORTS AND U SE S CENARIOS OF
MECStandardization is an indispensable step for successful pro-motion of a new technology, which documents the consensusamong multiple players and defines voluntary characteristicsand rules in a specific industry. Due to the availability ofstructured methods and reliable data, standardization helps topromote innovation and disseminate groupbreaking ideas andknowledge about cutting-edge techniques. More importantly,standardization can build customer trust in products, servicesand systems, which helps to develop favorable market condi-tion. The technical standards for MEC are being developedby ETSI, and a new industry specification group (ISG) was (cid:48)(cid:40)(cid:38)(cid:3)(cid:57)(cid:76)(cid:85)(cid:87)(cid:88)(cid:68)(cid:79)(cid:76)(cid:93)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:47)(cid:68)(cid:92)(cid:72)(cid:85)(cid:48)(cid:40)(cid:38)(cid:3)(cid:43)(cid:82)(cid:86)(cid:87)(cid:76)(cid:81)(cid:74)(cid:3)(cid:44)(cid:81)(cid:73)(cid:85)(cid:68)(cid:86)(cid:87)(cid:85)(cid:88)(cid:70)(cid:87)(cid:88)(cid:85)(cid:72)(cid:48)(cid:40)(cid:38)(cid:3)(cid:43)(cid:68)(cid:85)(cid:71)(cid:90)(cid:68)(cid:85)(cid:72)(cid:3)(cid:53)(cid:72)(cid:86)(cid:82)(cid:88)(cid:85)(cid:70)(cid:72)(cid:86)(cid:48)(cid:40)(cid:38)(cid:3)(cid:57)(cid:76)(cid:85)(cid:87)(cid:88)(cid:68)(cid:79)(cid:76)(cid:93)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:48)(cid:68)(cid:81)(cid:68)(cid:74)(cid:72)(cid:85)(cid:15)(cid:3)(cid:44)(cid:68)(cid:68)(cid:54)(cid:48)(cid:40)(cid:38)(cid:3)(cid:36)(cid:83)(cid:83)(cid:79)(cid:76)(cid:70)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:51)(cid:79)(cid:68)(cid:87)(cid:73)(cid:82)(cid:85)(cid:80)(cid:48)(cid:40)(cid:38)(cid:3)(cid:36)(cid:83)(cid:83)(cid:79)(cid:76)(cid:70)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:51)(cid:79)(cid:68)(cid:87)(cid:73)(cid:82)(cid:85)(cid:80)(cid:3)(cid:54)(cid:72)(cid:85)(cid:89)(cid:76)(cid:70)(cid:72)(cid:86) (cid:55)(cid:85)(cid:68)(cid:73)(cid:73)(cid:76)(cid:70)(cid:3)(cid:82)(cid:73)(cid:73)(cid:79)(cid:82)(cid:68)(cid:71)(cid:76)(cid:81)(cid:74)(cid:3)(cid:73)(cid:88)(cid:81)(cid:70)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:11)(cid:55)(cid:50)(cid:41)(cid:12) (cid:53)(cid:68)(cid:71)(cid:76)(cid:82)(cid:3)(cid:49)(cid:72)(cid:87)(cid:90)(cid:82)(cid:85)(cid:78)(cid:44)(cid:81)(cid:73)(cid:82)(cid:85)(cid:80)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:54)(cid:72)(cid:85)(cid:89)(cid:76)(cid:70)(cid:72)(cid:3)(cid:11)(cid:53)(cid:49)(cid:44)(cid:54)(cid:12) (cid:38)(cid:82)(cid:80)(cid:80)(cid:88)(cid:81)(cid:76)(cid:70)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:54)(cid:72)(cid:85)(cid:89)(cid:76)(cid:70)(cid:72)(cid:86) (cid:54)(cid:72)(cid:85)(cid:89)(cid:76)(cid:70)(cid:72)(cid:3)(cid:53)(cid:72)(cid:74)(cid:76)(cid:86)(cid:87)(cid:85)(cid:92) (cid:48)(cid:40)(cid:38)(cid:3)(cid:36)(cid:83)(cid:83) (cid:57)(cid:48) (cid:48)(cid:40)(cid:38)(cid:3)(cid:36)(cid:83)(cid:83) (cid:57)(cid:48) (cid:48)(cid:40)(cid:38)(cid:3)(cid:36)(cid:83)(cid:83) (cid:57)(cid:48) (cid:48)(cid:40)(cid:38)(cid:3)(cid:36)(cid:83)(cid:83) (cid:57)(cid:48) (cid:48)(cid:40)(cid:38)(cid:3)(cid:36)(cid:83)(cid:83) (cid:57)(cid:48) (cid:36)(cid:51)(cid:44) (cid:36)(cid:51)(cid:44) (cid:36)(cid:51)(cid:44) (cid:36)(cid:51)(cid:44) (cid:36)(cid:83)(cid:83)(cid:79)(cid:76)(cid:70)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:48)(cid:68)(cid:81)(cid:68)(cid:74)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87)(cid:3)(cid:54)(cid:92)(cid:86)(cid:87)(cid:72)(cid:80)(cid:86)(cid:48)(cid:40)(cid:38)(cid:3)(cid:51)(cid:79)(cid:68)(cid:87)(cid:73)(cid:82)(cid:85)(cid:80)(cid:86)(cid:3)(cid:48)(cid:68)(cid:81)(cid:68)(cid:74)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87)(cid:3)(cid:54)(cid:92)(cid:86)(cid:87)(cid:72)(cid:80)(cid:48)(cid:40)(cid:38)(cid:3)(cid:43)(cid:82)(cid:86)(cid:87)(cid:76)(cid:81)(cid:74)(cid:3)(cid:44)(cid:81)(cid:73)(cid:85)(cid:68)(cid:86)(cid:87)(cid:85)(cid:88)(cid:70)(cid:87)(cid:88)(cid:85)(cid:72)(cid:3)(cid:48)(cid:68)(cid:81)(cid:68)(cid:74)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87)(cid:3)(cid:54)(cid:92)(cid:86)(cid:87)(cid:72)(cid:80) (cid:22)(cid:42)(cid:51)(cid:51)(cid:3)(cid:53)(cid:68)(cid:71)(cid:76)(cid:82)(cid:3)(cid:49)(cid:72)(cid:87)(cid:90)(cid:82)(cid:85)(cid:78)(cid:3)(cid:40)(cid:79)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) Fig. 13. MEC platform overview [5]. established within ETSI by Huawei, IBM, Nokia Networks,NTT docomo and Vodafone. The aim of the ISG is to buildup a standardized and open environment, which will allow theefficient and seamless integration of applications from vendors,service providers, and third-parties across multi-vendor MECplatforms [229]. In September 2014, an introductory technicalwhite paper on MEC was published by ETSI, which definedthe concept of MEC, proposed the referenced MEC platform,as well as pointed out a set of technical requirements andchallenges for MEC [5]. Also, typical use scenarios and theirrelationships with MEC have been discussed. These aspectshave also been documented in the ETSI specifications in 2015[47], [230]–[232]. Most recently, ETSI has announced six
Proofs of Concepts (PoCs) that were accepted by the MEC ISGin MEC World Congress 2016, which will assist the strategicplanning and decision-making of organizations, as well ashelp to identify which MEC solutions may be viable in thenetwork [233]. This provides the community with confidencein MEC and will accelerate the pace of the standardization. Itis interesting to note that, in this congress, the ETSI MEC ISGhas renamed
Mobile Edge Computing as Multi-access EdgeComputing in order to reflect the growing interest in MECfrom non-cellular operators, which will take effects startingfrom 2017 [234]. Most recently, the (3GPP) shows a growing interest in includingMEC into its 5G standard, and functionality supports foredge computing has been identified and reported in a recenttechnical specification document [235]. In this section, wewill first introduce the recent standardization efforts from theindustry, including the referenced MEC server framework as well as the technical challenges and requirements of MECsystems. Typical use scenarios of MEC will be then elaborated.In addition, we will discuss MEC-related issues in 5G stan-dardizations, including the functionality supports for MEC,and the innovative features in 5G systems with the potentialto help realize MEC.
A. Referenced MEC Server Framework
In the MEC introductory technical white paper [5], theETSI MEC ISG has defined a referenced framework for MECservers (a.k.a. MEC platforms), where each server consists of ahosting infrastructure and an application platform as shown inFig. 13. The hosting infrastructure includes the hardware com-ponents (such as the computation, memory, and networkingresources) and an MEC virtualization layer (which abstractsthe detailed hardware implementation to the MEC applicationplatform). Also, the MEC host infrastructure provides theinterface to the host infrastructure management system as wellas the radio network elements, which, however, are beyond thescope of the MEC initiative due to the availability of multipleimplementation options.The MEC application platform includes an MEC virtual-ization manager together with an
Infrastructure as a Service(IaaS) controller, and provides multiple MEC applicationplatform services. The MEC virtualization manager supportsa hosting environment by providing IaaS facilities, while theIaaS controller provides a security and resource sandbox (i.e.,a virtual environment) for both the applications and MECplatform. The MEC application platform offers four maincategories of services, i.e., traffic offloading function (TOF), radio network information services (RNIS), communicationservices, and service registry. An MEC application platformmanagement interface is used by the operators for MECapplication platform management, supporting the applicationconfiguration and life cycle control, as well as VM operationmanagement.On top of the MEC application platform, the MEC appli-cations are deployed and executed within the VMs, whichare managed by their related application management systemsand agnostic to the MEC server/platform and other MECapplications. B. Technical Challenges and Requirements
In this subsection, we will briefly summarize the technicalchallenges and requirements specified in [5], [232]. Network Integration : As MEC is a new type of servicedeployed on top of the communication networks, the MECplatform is supposed to be transparent to the 3GPP networkarchitectures, i.e., the existing 3GPP specifications should notbe largely affected by the introduction of MEC. Application Portability : Application portability requiresMEC applications to be seamlessly loaded and executed by theMEC servers deployed by multiple vendors. This eliminatesthe need for dedicated development or integration efforts foreach MEC platform, and provides more freedom on optimizingthe location and execution of MEC applications. It requiresthe consistency of the MEC application platform managementsystems, as well as mechanisms used to package, deploy andmanage applications from different platforms and vendors. Security : The MEC systems face more security chal-lenges than communication networks due to the integrationof computing and IT services. Hence, the security require-ments for the 3GPP networks and the IT applications (e.g.,isolating different applications as much as possible) shouldbe simultaneously satisfied. Besides, because of the natureof proximity, the physical security of the MEC servers ismore vulnerable compared to conventional data centers. Thus,the MEC platforms need to be designed in a way that bothlogical intrusions and physical intrusions are well protected.Moreover, authorization is an important aspect to prevent theunauthorized/untrusted third-party applications from destroy-ing MEC hosts as well as the valued radio access network. Performance : As mentioned previously, the telecomoperators expect that introducing MEC will have minimalimpacts on the network performance, e.g., the throughput,latency, and packet loss. Thus, sufficient capacity should beprovisioned to process the user traffic in the system deploy-ment stage. Also, because of the highly-virtualized nature, theprovided performance may be impaired especially for thoseapplications that require intensive use of hardware resourcesor have low latency requirements. As a result, how to improvethe efficiency of virtualized environments becomes a bigchallenge. Resilience : The MEC systems should offer certainlevel of resilience and meet the high-availability requirementsdemanded by their network operators. The MEC platformsand applications should have fault-tolerant abilities to prevent them from adversely affecting other normal operations of thenetwork. Operation : The virtualization and Cloud technologiesmake it possible for various parties to participate in themanagement of MEC systems. Thus, the implementation ofthe management framework should also consider the diversityof potential deployments. Regulatory and Legal Considerations : The develop-ment of MEC systems should meet the regulatory and legalrequirements, e.g., the privacy and charging.Besides the aforementioned challenges and requirements,there still exist more aspects that should be considered in thefinal MEC standards, such as the support for user mobility,applications/traffic migration, and requirements on the con-nectivity and storage. However, currently, the standardizationefforts and even efforts from the research communities are stillon their infant stages.
C. Use Scenarios
MEC will enable numerous mobile applications. In thissubsection, we will introduce four typical use scenarios thathave been documented by ETSI MEC ISG in [47]. Video Stream Analysis Service : Video stream analysishas a broad range of applications such as the vehicularlicense plate recognition, face recognition, and home securitysurveillance, for which, the basic operations include objectdetection and classification. The video analysis algorithmsnormally have a high computation complexity, and thus it ispreferable to move the analysis jobs away from the video-capturing devices (e.g., the camera) to simplify the devicedesign and reduce the cost. If these processing tasks arehandled in the central cloud, the video stream should be routedto the core network [236], which will consume a great amountof network bandwidth due to the nature of video stream.By performing the video analysis in the place close to edgedevices, the system can not only enjoy the benefits of lowlatency, but also avoid the problem of network congestioncaused by the video stream uploading. The MEC-based videoanalysis system is shown in Fig. 14, where the edge servershould have the ability to conduct video management andanalysis, and only the valuable video clips (screenshots) willbe backed up to the cloud data centers. Augmented Reality Service : AR is a live direct orindirect view of a physical, real-world environment whoseelements are augmented (or supplemented) by computer-generated sensory inputs such as sound, video, graphics,or GPS data . Upon analyzing such information, the ARapplications can provide additional information in real-time.The AR applications are highly localized and require lowlatency as well as intensive data processing. One of themost popular applications is the museum video guides, i.e., ahandheld mobile device that provides the detailed informationof some exhibits that cannot be easily shown on the scene.Online games, such as the Pok´emon Go , is another important https://en.wikipedia.org/wiki/Augmented reality Fig. 14. MEC for video stream analysis [5]. (cid:36)(cid:53)(cid:3)(cid:50)(cid:69)(cid:77)(cid:72)(cid:70)(cid:87)(cid:18)(cid:39)(cid:68)(cid:87)(cid:68)(cid:3)(cid:38)(cid:68)(cid:70)(cid:75)(cid:72) (cid:48)(cid:40)(cid:38)(cid:3)(cid:54)(cid:72)(cid:85)(cid:89)(cid:72)(cid:85) (cid:38)(cid:72)(cid:81)(cid:87)(cid:85)(cid:68)(cid:79)(cid:3)(cid:36)(cid:53)(cid:3)(cid:38)(cid:68)(cid:70)(cid:75)(cid:72) (cid:44)(cid:81)(cid:87)(cid:72)(cid:85)(cid:81)(cid:72)(cid:87)(cid:3)(cid:38)(cid:82)(cid:81)(cid:87)(cid:72)(cid:81)(cid:87)(cid:3)(cid:54)(cid:72)(cid:85)(cid:89)(cid:72)(cid:85) (cid:50)(cid:69)(cid:77)(cid:72)(cid:70)(cid:87)(cid:3)(cid:44)(cid:39) (cid:38)(cid:82)(cid:85)(cid:72)(cid:3)(cid:49)(cid:72)(cid:87)(cid:90)(cid:82)(cid:85)(cid:78)
TrackerTracker MapperMapper Object (cid:3)
RecognizerObject (cid:3)
Recognizer (cid:57)(cid:76)(cid:71)(cid:72)(cid:82) (cid:48)(cid:82)(cid:69)(cid:76)(cid:79)(cid:72)(cid:3)(cid:39)(cid:72)(cid:89)(cid:76)(cid:70)(cid:72)
Fig. 15. MEC for AR services [5]. application that AR techniques play a critical role. An MEC-based AR application system is shown in Fig. 15, wherethe MEC server should be able to distinguish the requestedcontents by accurately analyzing the input data, and thentransmit the AR data back to the end user. Much attentionhas been paid on the MEC-enabled AR systems recently, andone demo has been implemented by Intel and roadshowed inthe Mobile World Congress 2016 [237]. IoT Applications : To simplify the hardware complexityof IoT devices and prolong their battery lives, it is promising tooffload the computation-intensive tasks for remote processingand retrieve the results (required action) once the processingis completed. Also, some IoT applications need to obtaindistributed information for computation, which might be dif-ficult for an IoT device without the aid of an external entity.Since the MEC servers host high-performance computationcapabilities and are able to collect distributed information,their deployment will significantly simplify the design of IoTdevices, without the need to have strong processing powerand capability to receive information from multiple sourcesfor performing meaningful computation. Another importantfeature of IoT is the heterogeneity of the devices runningdifferent forms of protocols, and their management should beaccomplished by a low-latency aggregation point (gateway),which could be the MEC server. Connected Vehicles : The connected vehicle technologycan enhance safety, reduce traffic congestion, sense vehicles’behaviors, as well as provide opportunities for numerousvalue-added services such as the car finder and parking lo-cation [238]–[240]. However, the maturity of such technology is yet to come as the latency requirement cannot be met withthe existing connected car clouds, which contributes to an end-to-end latency between 100ms to 1s. MEC is a key enablingtechnique for connected vehicles by adding computation andgeo-distributed services to roadside BSs. By receiving andanalyzing the messages from proximate vehicles and roadsidesensors, the connected vehicle cloudlets are able to propagatethe hazard warnings and latency-sensitive messages withina 20ms end-to-end delay, allowing the drivers to react im-mediately (as shown in Fig. 16) and make it possible forautonomous driving. The connected vehicle technology hasalready attracted extensive attention from the automobile man-ufacturers (e.g., Volvo, Peugeot), automotive supplier (e.g.,BOSCH), telecom operators (e.g., Orange, Vodafone, NTTdocomo), telecom vendors (e.g., QualComm, Nokia, Huawei),as well as many research institutes. In November 9 2015,Nokia presented two use cases for connected vehicles onan automotive driving testbed, including the emergency brakelight and cooperative passing assistance.In addition to connected vehicle systems with automobiles,MEC will also be applicable for enabling connected unmannedaerial vehicles (UAVs), which play an increasingly importantrole in various scenarios such as photography, disaster re-sponse, inspection and monitoring, precision agriculture, etc.In 2016, Nokia proposed the UAV traffic management (UTM)based MEC architecture for connected UAVs in [241], wherethe UTM unit provides functions of fleet management, auto-mated UAV missions, 3D navigation, and collision avoidance.However, as existing mobile networks are mainly designed for https://networks.nokia.com/solutions/mobile-edge-computing Fig. 16. MEC for connected vehicles. users on the ground, UAVs will have very limited connectivityand bandwidth. Therefore, reconfiguring the mobile networksto guarantee the connectivity and low latency between theUAVs and the infrastructure becomes a critical task for de-signing MEC systems for connected UAVs.Due to limited space, we omit the description of some otherinteresting application scenarios, such as active device track-ing, RAN-aware content optimization, distributed content and
Domain Name System (DNS) caching, enterprise networks, aswell as safe-and-smart cities. Interested readers may refer tothe white papers on MEC [5], [21], [242] for details.
D. MEC in 5G Standardizations
The 5G standard is currently under development, which is toenable the connectivity of a broad range of applications withnew functionality, characteristics, and requirements [77]. Toachieve these visions, the network features and functionalityin 5G networks are foreseen to be migrated from hardware tosoftware, thanks to the recent development of SDN and NFVtechniques. Since 2015, MEC (together with SDN and VFN) isrecognized by the European
5G infrastructure Public PrivatePartnership (5GPPP) research body as one of the key emergingtechnologies for 5G networks as it is a natural developmentin the evolution of mobile BSs and the convergence of IT andtelecommunication networking [15]. In April 2017, 3GPP hasincluded supporting edge computing as one of the high levelfeatures in 5G systems in the technical specification document[235], which will be introduced in this subsection. We havealso identified some innovative features of 5G systems, whichwould pave the way for the realization, standardization andcommercialization of MEC.
1) Functionality Supports Offered by 5G Networks:
Fromthe 5G network operators’ point of view, reducing the end-to-end latency and load on the transport networks are twodominant design targets, which could possibly be achievedwith MEC as operators and third part applications couldbe hosted close to the user equipment’s (UE’s) associatedwireless AP. To integrate MEC in 5G systems, the recent 5Gtechnical specifications have explicitly pointed out necessaryfunctionality supports that should be offered by 5G networksfor edge computing, as listed below: • The 5G core network should select the traffic to be routedto the applications in the local data networks. • The 5G core network selects a user plane function (UPF)in proximity to the UE to route and execute the trafficsteering from the local data networks via the interface,which should be based on the UE’s subscription data,UE location, and the data from the application function (AF). • The 5G network should guarantee the session and servicecontinuity to enable UE and application mobility. • The 5G core network and AF should provide informationto each other via the network exposure function (NEF) . • The policy control function (PCF) provides rules for QoScontrol and charging for the traffic routed to the local datanetwork. The NEF supports external exposure of capabilities of network functions,which can be categorized into monitoring capability, provisioning capability,and policy/charging capability [235]. The PCF was defined as a stand-alone functional part of the 5G corenetwork that allows to shape the network behaviour based on the operatorpolicies [243].
2) Innovative Features in 5G to Facilitate MEC:
Comparedto previous generations of wireless networks, 5G networkspossess various innovative features that are beneficial to therealization, standardization, and commercialization of MEC.Three of them will be detailed in this subsection, including the support service requirement , mobility management strategy ,and capability of network slicing . • Support Service Requirement:
In 5G systems, theQoS characteristics (in terms of resource type, prioritylevel, packet delay budget, and packet error rate), whichdescribe the packet forwarding treatment that a QoS flowreceives edge-to-edge between the UE and the UPF, areassociated with the
5G QoS Indicator (5QI). In [235],a standardized 5QI to QoS mapping table is provided,showing a broad range of services that can be supportedin 5G systems. In particular, 5G systems are able to caterthe requirements of latency-sensitive applications (e.g.,real-time gaming and vehicular-to-everything (V2X) mes-sages, which have a stringent packet budget delay require-ment, i.e., < < − ), and mission-critical services (e.g., push-to-talk signaling that has both low delay ( < < − ) requirements). Theseapplications coincide with typical MEC applications asmentioned in Section V-C, i.e., 5G network is a viablechoice for wireless communications in MEC systems. • Advanced Mobility Management Strategy:
The con-cept of mobility pattern was introduced for designing mo-bility management strategy for 5G systems. Such strate-gies may be used by the 5G core network to characterizeand optimize UE mobility. Specifically, the mobilitypattern could be determined, monitored, and updated bythe 5G core network based on the subscription of the UE,statistics of UE mobility, network local policy, and UEassisted information [235]. The mobility pattern not onlyplays a central role on designing advanced transmissionschemes in wireless communication systems, but alsobecomes a non-negligible design consideration for manyMEC applications discussed in Section V-C, e.g., the ARservices and connected vehicular applications. Thus, inte-gration of advanced mobility management strategies thatmake full use of the mobility pattern in 5G network canhelp to develop an efficient wireless interface for MECsystems. Besides, the mobility pattern obtained from the5G core network can be further leveraged to design jointradio-and-computational resource management strategiesfor MEC systems. • Capability of Network Slicing:
Network slicing is aform of agile and virtual network architecture that allowsmultiple network instances to be created on top of acommon shared physical infrastructure . Each of thenetwork instances is optimized for a specific service,enabling resource isolation and customized network oper-ations [244]. Due to the heterogeneous types of servicesthat 5G systems need to support (different requirements interms of functionality and performance), network slicing https://5g.co.uk/guides/what-is-network-slicing/ is regarded as an indispensable feature in 5G systems tosupport different services running across a single radioaccess network. Existing studies found that network slic-ing is of supreme need for three use scenarios, including ultra-reliable and low latency communication (URLLC),massive machine type communication (mMTC), and en-hanced mobile broadband (eMBB) [245]. With the capa-bility of network slicing in 5G systems, MEC applicationscould be provisioned with optimized and dedicated net-work resources, which could help to reduce the latencyincurred by the access networks substantially and supportintense access of MEC service subscribers.VI. C ONCLUSION
MEC is an innovative network paradigm to cater for theunprecedented growth of computation demands and the ever-increasing computation quality of user experience require-ments. It aims at enabling Cloud Computing capabilities andIT services in close proximity to end users, by pushingabundant computational and storage resources towards thenetwork edges. The direct interaction between mobile devicesand edge servers through wireless communications brings thepossibility of supporting applications with ultra-low latencyrequirement, prolonging device battery lives and facilitatinghighly-efficient network operations. However, they come alongwith various new design considerations and unique challengesdue to reasons such as the complex wireless environments andthe inherent limited computation capacities of MEC servers.In this survey, we presented a comprehensive overview andresearch outlook of MEC from the communication perspective.To this end, we first summarized the modeling methodologieson key components of MEC systems such as the computationtasks, communications, as well as mobile devices and MECservers computation. This help characterize the latency andenergy performance of MEC systems. Based upon the systemmodeling, we conducted a comprehensive literature reviewon recent research efforts on resource management for MECunder various system architectures, which exploit the conceptsof computation offloading, joint radio-and-computational re-source allocation, MEC server scheduling, as well as multi-server selection and cooperation. A number of potential re-search directions were then identified, including MEC de-ployment issues, cache-enabled MEC, mobility managementfor MEC, green MEC, as well as security-and-privacy issuesin MEC. Key research problems and preliminary solutionsfor each of these directions were elaborated. Finally, weintroduced the recent standardization efforts from industry,along with several typical use scenarios. The comprehensiveoverview and research outlook on MEC provided in thissurvey hopefully can serve as useful references and valuableguidelines for further in-depth investigations of MEC.R [2] Q. Zhang, L. Cheng, and R. Boutaba, “Cloud computing: State-of-the-art and research challenges,” Journal Internet Services appl. , vol. 1,no. 1, pp. 7–18, 2010.[3] N. Wingfield, “Amazon’s profits grow more than 800 percent,lifted by cloud services,”
The New York Times
IEEE Internet Things J. , vol. PP, no. 99, pp. 1–1, 2016.[5] ETSI, “Mobile-edge computing introductory technical whitepaper,”
White Paper, Mobile-edge Computing Industry Initiative .[Online]. Available: https://portal.etsi.org/portals/0/tbpages/mec/docs/mobile-edge computing - introductory technical white paper v1[6] G. P. Fettweis, “The tactile Internet: Applications and challenges,”
IEEE Veh. Techn. Mag. , vol. 9, no. 1, pp. 64–70, Mar. 2014.[7] A. A. Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, andM. Ayyash, “Internet of Things: A survey on enabling technologies,protocols, and applications,”
IEEE Commun. Surveys Tuts. , vol. 17,no. 4, pp. 2347–2376, 4th Quarter 2015.[8] Juniper, “Smart wireless devices and the Inter-net of me,”
White paper , Mar. 2015. [Online].Available: http://itersnews.com/wp-content/uploads/experts/2015/03/96079Smart-Wireless-Devices-and-the-Internet-of-Me.pdf[9] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. Soong,and J. C. Zhang, “What will 5G be?”
IEEE J. Sel. Areas Commun.
Proc. ACM 1st edition of the MCCworkshop on Mobile cloud computing , 2012, pp. 13–16.[12] S. Yi, C. Li, and Q. Li, “A survey of fog computing: Concepts,applications and issues,” in
Proc. ACM Workshop on Mobile Big Data
Computer , vol. 7,no. 6, pp. 34–45, 1974.[15] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young, “Mobileedge computing—A key technology towards 5G,”
ETSI White Paper ,vol. 11, 2015.[16] C.-Y. Chang, K. Alexandris, N. Nikaein, K. Katsalis, and T. Spyropou-los, “MEC architectural implications for LTE/LTE-A networks,” in
Proc. ACM Workshop on Mobility in the Evolving Internet Architecture(MobiArch) , New York, NY, Oct. 2016, pp. 13–18.[17] Z. Q. Jaber and M. I. Younis, “Design and implementation of real timeface recognition system (RTFRS),”
Int. J. Compt. Appl. , vol. 94, no. 12,pp. 15–22, May 2014.[18] T. Verbelen, P. Simoens, F. D. Turck, and B. Dhoedt, “Leveragingcloudlets for immersive collaborative applications,”
IEEE PervasiveComput. , vol. 12, no. 4, pp. 30–38, Oct.-Dec. 2013.[19] A. A. Shuwaili and O. Simeone, “Energy-efficient resource allocationfor mobile edge computing-based augmented reality applications,”
IEEE Wireless Commun. Lett. , vol. PP, no. 99, Apr. 2017.[20] A.-S. Ali and O. Simeone, “Energy-efficient resource allocation formobile edge computing-based augmented reality applications,”
IEEEWireless Commun. Lett.
Commun. ACM , vol. 53, no. 4, pp. 52–58, Apr.2010.[23] M. Othman, S. A. Madani, S. U. Khan et al. , “A survey of mobilecloud computing application models,”
IEEE Commun. Surveys Tuts. ,vol. 16, no. 1, pp. 393–413, 1st Quater 2014.[24] M. F. Bari, R. Boutaba, R. Esteves, L. Z. Granville, M. Podlesny,M. G. Rabbani, Q. Zhang, and M. F. Zhani, “Data center networkvirtualization: A survey,”
IEEE Commun. Surveys Tuts. , vol. 15, no. 2,pp. 909–928, 2nd Quarter 2013.[25] A. Ghiasi and R. Baca, “Overview of largest data centers,”
IEEE802.3bs Task Force Interim Meeting
Proc. IEEE Int.Conf. Pervasive Comput. Commun. (PerCom) , Lugano, Switzerland,Mar. 2012, pp. 122–127.[27] T. X. Tran, A. Hajisami, P. Pandey, and D. Pompili, “Collaborativemobile edge computing in 5G networks: New paradigms, scenarios,and challenges,”
IEEE Commun. Mag. , vol. 55, no. 4, Apr. 2017.[28] S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, and W. Wang, “Asurvey on mobile edge networks: Convergence of computing, cachingand communications,”
IEEE Access , to appear.[29] J. Zhang, W. Xie, F. Yang, and Q. Bi, “Mobile edge computing andfield trial results for 5G low latency scenario,”
China Commun. , vol. 13,no. 2 (Supplement), pp. 174–182, 2016.[30] E. Cuervo, A. Balasubramanian, D.-k. Cho, A. Wolman, S. Saroiu,R. Chandra, and P. Bahl, “Maui: Making smartphones last longerwith code offload,” in
Proc. ACM Int. Conf. Mobile Syst. Appl. Serv.(MobiSys) , San Francisco, California, USA, Jun. 2010, pp. 49–62.[31] M. Satyanarayanan, P. Bahl, R. Caceres, and N. Davies, “The case forVM-based cloudlets in mobile computing,”
IEEE Pervasive Comput. ,vol. 8, no. 4, pp. 14–23, 2009.[32] 5GPPP, “5g automotive vision,”
White Paper . [On-line]. Available: https://5g-ppp.eu/wp-content/uploads/2014/02/5G-PPP-White-Paper-on-Automotive-Vertical-Sectors.pdf[33] O. Khalid, M. Khan, S. Khan, and A. Zomaya, “OmniSuggest: Aubiquitous cloud based context aware recommendation system formobile social networks,”
IEEE Trans. Serv. Comput. , vol. 7, no. 3,pp. 401–414, Dec. 2014.[34] K. Goel and M. Goel, “Cloud computing based e-commerce model,”in
Proc. IEEE Int. Conf. Recent Trends in Electron., Info. & Commun.Techn. (RTEICT) , Banglore, India, May 2016, pp. 27–30.[35] G. Riah, “E-learning systems based on cloud computing: A review,”
ELSEVIER Proc. Comput. Sci. , vol. 62, pp. 352–359, Sep. 2015.[36] A. Abbas and S. U. Khan, “A review on the state-of-the-art privacy-preserving approaches in the e-health clouds,”
IEEE J. Biomed. HealthInform. , vol. 18, no. 4, pp. 1431–1441, Apr. 2014.[37] G. Intelligence, “Understanding 5G: Perspectives on future technolog-ical advancements in mobile,”
London, UK , 2014.[38] A. Somov and R. Giaffreda, “Powering IoT devices: Technologiesand opportunities,”
Available: http://iot.ieee.org/newsletter/november-2015/powering-iot-devices-technologies-and-opportunities.html .[39] R. Kemp, N. Palmer, T. Kielmann, F. Seinstra, N. Drost, J. Maassen,and H. Bal, “EyeDentify: Multimedia cyber foraging from a smart-phone,” in
Proc. IEEE Int. Symp. Multimedia , San Diego, CA, USA,Dec. 2009, pp. 392–399.[40] B. Shi, J. Yang, Z. Huang, and P. Hui, “Offloading guidelines foraugmented reality applications on wearable devices,” in
Proc. ACMInt. Symp. Multimedia , Brisbane, Australia, Oct. 2015, pp. 1271–1274.[41] W. N. Schilit, “A system architecture for context-aware mobile com-puting,” Ph.D. dissertation, Columbia University, 1995.[42] C. Perera, A. Zaslavsky, P. Christen, and D. Georgakopoulos, “Contextaware computing for the Internet of Things: A survey,”
IEEE Commun.Surveys Tuts. , vol. 16, no. 1, pp. 414–454, 1st Quater 2014.[43] S. Nunna, A. Kousaridas, M. Ibrahim, M. Dillinger, C. Thuemmler,H. Feussner, and A. Schneider, “Enabling real-time context-awarecollaboration through 5G and mobile edge computing,” in
Proc. IEEEInt. Conf. Inf. Techn. New Generations (ITNG) , Las Vegas, NV, Apr.2015, pp. 601–605.[44] X. Luo, “From augmented reality to augmented computing: A lookat cloud-mobile convergence,” in
Proc. IEEE Int. Symp. UbiquitousVirtual Reality , Gwangju, South Korea, Jul. 2009, pp. 29–32.[45] A. Thiagarajan, L. Ravindranath, H. Balakrishnan, S. Madden, andL. Girod, “Accurate, low-energy trajectory mapping for mobile de-vices.” in
Proc. USENIX Symp. Networked Systems Design and Imple-mentation (NSDI) , Boston, MA, Mar. 2011, pp. 1–14.[46] H. Suo, Z. Liu, J. Wan, and K. Zhou, “Security and privacy inmobile cloud computing,” in
Proc. IEEE Int. Wireless Commun. MobileComput. Conf. (IWCMC)
Comput ,vol. 49, no. 5, pp. 78–81, May 2016.[49] O. Salman, I. Elhajj, A. Kayssi, and A. Chehab, “Edge computingenabling the Internet of Things,” in
Proc. IEEE World Forum Internetof Things (WFIOT) , Dec. 2015, pp. 603–608. [50] A. Ahmed and E. Ahmed, “A survey on mobile edge computing,” in Proc. IEEE Int. Conf. Intell. Syst. Control (ISCO) , Coimbatore, India,Jan. 2016, pp. 1–8.[51] T. Taleb, K. Samdanis, B. Mada, H. Flinck, S. Dutta, and D. Sabella,“On multi-access edge computing: A survey of the emerging 5g net-work edge architecture & orchestration,”
IEEE Commun. SurveysTuts. , to appear.[52] H. Liu, F. Eldarrat, H. Alqahtani, A. Reznik, X. de Foy, andY. Zhang, “Mobile edge cloud system: Architectures, challenges, andapproaches,”
IEEE Syst. J. , to appear.[53] M. T. Beck, M. Werner, S. Feld, and S. Schimper, “Mobile edgecomputing: A taxonomy,” in
Proc. Int. Conf. Advances Future Internet(AFIN) , Lisbon, Portugal, Nov 2014, pp. 48–54.[54] P. Mach and Z. Becvar, “Mobile edge computing: A survey on archi-tecture and computation offloading,”
IEEE Commun. Surveys Tuts. , toappear.[55] D. Sabella, A. Vaillant, P. Kuure, U. Rauschenbach, and F. Giust,“Mobile-edge computing architecture: The role of MEC in the Internetof Things.”
IEEE Consum. Electron. Mag. , vol. 5, no. 4, pp. 84–91,Oct. 2016.[56] E. Ahmed and M. H. Rehmani, “Mobile edge computing: Opportu-nities, solutions, and challenges,”
Future Generation Comput. Syst. ,vol. 70, pp. 59–63, May 2017.[57] A. u. R. Khan, M. Othman, S. A. Madani, and S. U. Khan, “A survey ofmobile cloud computing application models,”
IEEE Commun. SurveysTuts. , vol. 16, no. 1, pp. 393–413, 1st Quater 2014.[58] A. P. Miettinen and J. K. Nurminen, “Energy efficiency of mobileclients in cloud computing,” in
Proc. USENIX Conf. Hot Topics CloudComput. (HotCloud) , Boston, MA, Jun. 2010, pp. 1–7.[59] S. Melendez and M. P. McGarry, “Computation offloading decisionsfor reducing completion time,” 2016. [Online]. Available: http://arxiv.org/pdf/1608.05839.pdf[60] W. Yuan and K. Nahrstedt, “Energy-efficient soft real-time CPUscheduling for mobile multimedia systems,” in
Proc. ACM Symp.Operat. Syst. Principles (SOSP) , Bolton Landing, NY, USA, Oct. 2003,pp. 149–163.[61] M. Jia, J. Cao, and L. Yang, “Heuristic offloading of concurrent tasksfor computation-intensive applications in mobile cloud computing,”in
Proc. IEEE Int. Conf. Comput. Commun. (INFOCOM WKSHPS) ,Toronto, Canada, Apr. 2014, pp. 352–357.[62] S. E. Mahmoodi, R. N. Uma, and K. P. Subbalakshmi, “Optimal jointscheduling and cloud offloading for mobile applications,”
IEEE Trans.Cloud Comput. , vol. PP, no. 99, pp. 1–13, 2016.[63] A. Goldsmith,
Wireless Communications . New York, NY, USA:Cambridge University Press, 2005.[64] D. Gesbert, S. Hanly, H. Huang, S. S. Shitz, O. Simeone, and W. Yu,“Multi-cell MIMO cooperative networks: A new look at interference,”
IEEE J. Sel. Areas Commun. , vol. 28, no. 9, pp. 1380–1408, Dec. 2010.[65] S. A. Jafar, “Topological interference management through indexcoding,”
IEEE Trans. Inf. Theory , vol. 60, no. 1, pp. 529–568, Jan2014.[66] C. Li, J. Zhang, M. Haenggi, and K. B. Letaief, “User-centric intercellinterference nulling for downlink small cell networks,”
IEEE Trans.Commun. , vol. 63, no. 4, pp. 1419–1431, Apr. 2015.[67] E. Torkildson, U. Madhow, and M. Rodwell, “Indoor millimeter waveMIMO: Feasibility and performance,”
IEEE Trans. Wireless Commun. ,vol. 10, no. 12, pp. 4150–4160, Dec. 2011.[68] X. Yu, J. C. Shen, J. Zhang, and K. B. Letaief, “Alternating mini-mization algorithms for hybrid precoding in millimeter wave MIMOsystems,”
IEEE J. Sel. Topics Signal Process. , vol. 10, no. 3, pp. 485–500, Apr. 2016.[69] S. M. Alamouti, “A simple transmit diversity technique for wirelesscommunications,”
IEEE J. Sel. Areas Commun. , vol. 16, no. 8, pp.1451–1458, Aug. 1998.[70] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, “Capacitylimits of MIMO channels,”
IEEE J. Sel. Areas Commun. , vol. 51, no. 6,pp. 684–195, Jun. 2003.[71] E. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta, “Massive MIMOfor next generation wireless systems,”
IEEE Commun. Mag. , vol. 52,no. 2, pp. 186–195, Feb. 2014.[72] J. G. Andrews, H. Claussen, M. Dohler, S. Rangan, and M. C. Reed,“Femtocells: Past, present, and future,”
IEEE Commun. Mag. , vol. 30,no. 3, pp. 497–508, Mar. 2012.[73] H. S. Dhillon, R. K. Ganti, F. Baccelli, and J. G. Andrews, “Modelingand analysis of K-tier downlink heterogeneous cellular networks,”
IEEETrans. Commun. , vol. 30, no. 3, pp. 550–560, Mar. 2012. [74] S. Han, Y.-C. Liang, and B.-H. Soong, “Spectrum refarming: A newparadigm of spectrum sharing for cellular networks,”
IEEE Trans.Commun. , vol. 63, no. 5, pp. 1895–1906, May 2016.[75] Q. Chen, G. Yu, and Z. Ding, “Optimizing unlicensed spectrumsharing for LTE-U and WiFi network coexistence,”
IEEE J. Sel. AreasCommun. , vol. 34, no. 10, pp. 2562–2574, Oct. 2016.[76] P. Kryszkiewicz, A. Kliks, and H. Bogucka, “Small-scale spectrumaggregation and sharing,”
IEEE J. Sel. Areas Commun. , vol. 34, no. 10,pp. 2630–2641, Oct. 2016.[77] ERICSSON, “5G radio access - Capabilities and technologies,”
White paper
Kluwer J. VLSI Signal Process. Syst. , vol. 13, no. 2/3, pp. 203–221,Aug. 1996.[79] W. Yuan and K. Nahrstedt, “Energy-efficient CPU scheduling formultimedia applications,”
ACM Trans. Compt. Syst. , vol. 24, no. 3,pp. 292–331, Aug 2006.[80] K. D. Vogeleer, G. Memmi, P. Jouvelot, and F. Coelho, “The en-ergy/frequency convexity rule: Modeling and experimental validationon mobile devices,” in
Proc. Springer Int. Conf. Parallel Process. Appl.Mathematics (PPAM) , Warsaw, Poland, Sep. 2013, pp. 793–803.[81] W. Zhang, Y. Wen, K. Guan, D. Kilper, H. Luo, and D. O. Wu, “Energy-optimal mobile cloud computing under stochastic wireless channel,”
IEEE Trans. Wireless Commun. , vol. 12, no. 9, pp. 4569–4581, Sep.2013.[82] A. Carroll and G. Heiser, “An analysis of power consumption in asmartphone,” in
Proc. USENIX Annual Technical Conf. , Boston, MA,Jun. 2010, pp. 1–14.[83] C. You, K. Huang, and H. Chae, “Energy efficient mobile cloudcomputing powered by wireless energy transfer,”
IEEE J. Sel. AreasCommun. , vol. 34, no. 5, pp. 1757–1771, May 2016.[84] C. You, K. Huang, H. Chae, and B.-H. Kim, “Energy-efficient resourceallocation for mobile-edge computation offloading,”
IEEE Trans. Wire-less Commun. , vol. PP, no. 99, 2016.[85] P. Barham, B. Dragovic, K. Fraser, S. Hand, T. Harris, A. Ho,R. Neugebauer, I. Pratt, and A. Warfield, “Xen and the art of virtual-ization,” in
Proc. ACM Symp. Operat. Syst. Principles (SOSP) , BoltonLanding, NY, Oct. 2003, pp. 164–177.[86] S. Barbarossa, S. Sardellitti, and P. Di Lorenzo, “Joint allocation ofcomputation and communication resources in multiuser mobile cloudcomputing,” in
Proc. IEEE Int. Workshop Signal Process. AdvancesWireless Commun. (SPAWC) , Darmstadt, Germany, Jun. 2013, pp. 26–30.[87] X. Chen, L. Jiao, W. Li, and X. Fu, “Efficient multi-user computationoffloading for mobile-edge cloud computing,”
IEEE Trans. Netw. ,vol. 24, pp. 2795–2808, Oct. 2016.[88] X. Lyu, H. Tian, P. Zhang, and C. Sengul, “Multi-user joint taskoffloading and resources optimization in proximate clouds,”
IEEETrans. Veh. Techn. , vol. PP, no. 99, pp. 1–13, Jul. 2016.[89] S. Vakilinia, M. M. Ali, and D. Qiu, “Modeling of the resourceallocation in cloud computing centers,”
Elsevier Comput. Netw. , vol. 91,pp. 453–470, Nov. 2015.[90] D. Bruneo, “A stochastic model to investigate data center performanceand QoS in IaaS cloud computing systems,”
IEEE Trans. ParallelDistrib. Syst. , vol. 25, no. 3, pp. 560–569, 2014.[91] X. Fan, W.-D. Weber, and L. A. Barroso, “Power provisioning for awarehouse-sized computer,” in
ACM SIGARCH Comput. Archit. News ,vol. 35, no. 2, 2007, pp. 13–23.[92] C.-C. Lin, P. Liu, and J.-J. Wu, “Energy-efficient virtual machineprovision algorithms for cloud systems,” in
Proc. IEEE Utility andCloud Computing (UCC) , Melbourne, Australia, Dec. 2011, pp. 81–88.[93] A. Beloglazov, J. Abawajy, and R. Buyya, “Energy-aware resourceallocation heuristics for efficient management of data centers for cloudcomputing,”
Elsevier Future Generation Comput. Syst. , vol. 28, no. 5,pp. 755–768, May 2012.[94] K. Kumar and Y. H. Lu, “Cloud computing for mobile users: Canoffloading computation save energy?”
Comput. , vol. 43, no. 4, pp. 51–56, Apr. 2010.[95] K. Kumar, J. Liu, Y.-H. Lu, and B. Bhargava, “A survey of computationoffloading for mobile systems,”
Mobile Netw. Appl. , vol. 18, no. 1, pp.129–140, Feb. 2013.[96] S. Barbarossa, S. Sardellitti, and P. D. Lorenzo, “Communicating whilecomputing: Distributed mobile cloud computing over 5G heterogeneousnetworks,”
IEEE Signal Process. Mag. , vol. 31, no. 6, pp. 45–55, Nov.2014. [97] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li, “Mobile-edge com-puting: Partial computation offloading using dynamic voltage scaling,” IEEE Trans. Commun. , vol. 64, no. 10, pp. 4268–4282, Oct. 2016.[98] Y. H. Kao, B. Krishnamachari, M. R. Ra, and F. Bai, “Hermes: Latencyoptimal task assignment for resource-constrained mobile computing,”in
Proc. IEEE Int. Conf. Comput. Commun. (INFOCOM) , Hong Kong,China, Apr. 2015, pp. 1894–1902.[99] W. Zhang, Y. Wen, and D. O. Wu, “Collaborative task execution inmobile cloud computing under a stochastic wireless channel,”
IEEETrans. Wireless Commun. , vol. 14, no. 1, pp. 81–93, Jan. 2015.[100] S. Khalili and O. Simeone, “Inter-layer per-mobile optimization ofcloud mobile computing: A message-passing approach.” [Online].Available: http://arxiv.org/abs/1509.01596[101] P. D. Lorenzo, S. Barbarossa, and S. Sardellitti, “Joint optimizationof radio resources and code partitioning in mobile edge computing.”[Online]. Available: http://arxiv.org/abs/1307.3835v3[102] S. E. Mahmoodi, K. P. Subbalakshmi, and V. Sagar, “Cloud offloadingfor multi-radio enabled mobile devices,” in
Proc. IEEE Int. Conf.Commun. (ICC) , London, UK, Jun. 2015, pp. 5473–5478.[103] D. Huang, P. Wang, and D. Niyato, “A dynamic offloading algorithmfor mobile computing,”
IEEE Trans. Wireless Commun. , vol. 11, no. 6,pp. 1991–1995, Jun. 2012.[104] J. Liu, Y. Mao, J. Zhang, and K. B. Letaief, “Delay-optimal compu-tation task scheduling for mobile-edge computing systems,” in
Proc.IEEE Int. Symp. Inf. Theory (ISIT) , Barcelona, Spain, Jul 2016, pp.1451–1455.[105] S. Chen, Y. Wang, and M. Pedram, “A semi-Markovian decisionprocess based control method for offloading tasks from mobile devicesto the cloud,” in
Proc. IEEE Global Commun. Conf. (GLOBECOM) ,Atlanta, GA, Dec 2013, pp. 2885–2890.[106] S.-T. Hong and H. Kim, “QoE-aware computation offloading schedul-ing to capture energy-latency tradeoff in mobile clouds,” in
Proc. IEEEInt. Conf. Sensing, Commun. Netw. (SECON) , London, UK, Jun. 2016,pp. 1–9.[107] J. Kwak, Y. Kim, J. Lee, and S. Chong, “Dream: Dynamic resourceand task allocation for energy minimization in mobile cloud systems,”
IEEE J. Sel. Areas Commun. , vol. 33, no. 12, pp. 2510–2523, Dec2015.[108] Z. Jiang and S. Mao, “Energy delay tradeoff in cloud offloading formulti-core mobile devices,”
IEEE Access , vol. 3, pp. 2306–2316, Nov.2015.[109] D. T. Hoang, D. Niyato, and P. Wang, “Optimal admission controlpolicy for mobile cloud computing hotspot with cloudlet,” in
Proc.IEEE Wireless Commun. Networking Conf. (WCNC) , Paris, France,Apr. 2012, pp. 3145–3149.[110] Y. Mao, J. Zhang, S. Song, and K. B. Letaief, “Power-delay tradeoffin multi-user mobile-edge computing systems,” in
Proc. IEEE GlobalCommun. Conf. (GLOBECOM) , Washington, DC, Dec. 2016, pp. 1–6.[111] K. Wang, K. Yang, and C. Magurawalage, “Joint energy minimizationand resource allocation in C-RAN with mobile cloud,”
IEEE Trans.Cloud Comput. , vol. PP, no. 99, pp. 1–11, Jan. 2016.[112] J. Ren, G. Yu, Y. Cai, and Y. He, “Latency optimization forresource allocation in mobile-edge computation offloading.” [Online].Available: https://arxiv.org/pdf/1704.00163.pdf[113] M.-H. Chen, B. Liang, and M. Dong, “Joint offloading decision andresource allocation for multi-user multi-task mobile cloud,” in
Proc.IEEE Int. Conf. Commun. (ICC) , Kuala Lumpur, Malaysia, May 2016,pp. 1–6.[114] M.-H. Chen, B. Liang, and D. Ming, “Joint offloading and resourceallocation for computation and communication in mobile cloud withcomputing access point,” in
Proc. IEEE Int. Conf. Comput. Commun.(INFOCOM) , Atlanta, GA, USA, Apr. 2017, pp. 1863–1871.[115] X. Chen, “Decentralized computation offloading game for mobile cloudcomputing,”
IEEE Trans. Parallel Distrib. Syst. , vol. 26, no. 4, pp. 974–983, Apr. 2016.[116] M.-H. Chen, B. Liang, and M. Dong, “Multi-user mobile cloudoffloading game with computing access point,” in
Proc. IEEE Int. Conf.Cloud Networking (Cloudnet) , Pisa, Italy, Oct. 2016, pp. 64–69.[117] X. Ma, C. Lin, X. Xiang, and C. Chen, “Game-theoretic analysis ofcomputation offloading for cloudlet-based mobile cloud computing,” in
Proc. ACM Int. Conf. Modeling, Anal. and Simulation of Wireless andMobile Syst. (MSWiM) , Cancun, Mexico, Nov. 2015, pp. 271–278.[118] S. Guo, B. Xiao, Y. Yang, and Y. Yang, “Energy-efficient dynamicoffloading and resource scheduling in mobile cloud computing,” in
Proc. IEEE Int. Conf. Comput. Commun. (INFOCOM) , San Francisco,CA, Apr. 2016, pp. 1–9. [119] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of radioand computational resources for multicell mobile-edge computing,”
IEEE Trans. Signal Info. Process. Over Networks , vol. 1, no. 2, pp.89–103, Jun. 2015.[120] M. Molina, O. Mu˜noz, A. Pascual-Iserte, and J. Vidal, “Joint schedulingof communication and computation resources in multiuser wirelessapplication offloading,” in
Proc. IEEE Int. Symp. on Personal Indoorand Mobile Radio Comm. (PIMRC) , Washington, DC, Sep. 2014, pp.1093–1098.[121] Y. Yu, J. Zhang, and K. B. Letaief, “Joint subcarrier and CPU timeallocation for mobile edge computing,” in
Proc. IEEE Global Commun.Conf. (GLOBECOM) , Washington, DC, Dec. 2016, pp. 1–6.[122] L. Yang, J. Cao, H. Cheng, and Y. Ji, “Multi-user computationpartitioning for latency sensitive mobile cloud applications,”
IEEETrans. Comput. , vol. 64, no. 8, pp. 2253–2266, Aug. 2015.[123] Y. Li, L. Sun, and W. Wang, “Exploring device-to-device communica-tion for mobile cloud computing,” in
Proc. IEEE Int. Conf. Commun.(ICC) , Sydney, Australia, Jun. 2014, pp. 2239–2244.[124] M. Jo, T. Maksymyuk, B. Strykhalyuk, and C.-H. Cho, “Device-to-device-based heterogeneous radio access network architecture formobile cloud computing,”
IEEE Wireless Commun. , vol. 22, no. 3, pp.50–58, Mar. 2015.[125] Z. Sheng, C. Mahapatra, V. Leung, M. Chen, and P. Sahu, “Energyefficient cooperative computing in mobile wireless sensor networks,”
IEEE Trans. Cloud Comput. , vol. PP, no. 99, pp. 1–13, Jul. 2015.[126] J. Song, Y. Cui, M. Li, J. Qiu, and R. Buyya, “Energy-traffic trade-off cooperative offloading for mobile cloud computing,” in
Proc.IEEE/ACM Int. Symp. Quality of Service (IWQoS) , Hong Kong, China,May 2014, pp. 284–289.[127] X. Cao, F. Wang, J. Xu, R. Zhang, and S. Cui, “Joint computationand communication cooperation for mobile edge computing,” Apr.2017. [Online]. Available: https://arxiv.org/pdf/1704.06777.pdf[128] C. You and K. Huang, “Mobile cooperative computing: Energy-efficientpeer-to-peer computation offloading,” 2017. [Online]. Available:https://arxiv.org/abs/1704.04595[129] L. Chen, S. Zhou, and J. Xu, “Computation peer offloading forenergy-constrained mobile edge computing in small-cell networks,”Mar. 2017. [Online]. Available: https://arxiv.org/pdf/1703.06058.pdf[130] L. Lei, Z. Zhong, K. Zheng, J. Chen, and H. Meng, “Challenges onwireless heterogeneous networks for mobile cloud computing,”
IEEEWireless Commun. , vol. 20, no. 3, pp. 34–44, 2013.[131] T. Zhao, S. Zhou, X. Guo, Y. Zhao, and Z. Niu, “A cooperativescheduling scheme of local cloud and Internet cloud for delay-awaremobile cloud computing,” in
Proc. IEEE Global Commun. Conf.Worshops (GC WKSHPS) , San Diego, CA, Dec. 2015, pp. 1–6.[132] Y. Ge, Y. Zhang, Q. Qiu, and Y.-H. Lu, “A game theoretic resourceallocation for overall energy minimization in mobile cloud computingsystem,” in
Proc. ACM/IEEE Int. Symp. Low Power Electron. Design ,Redondo Beach, CA, Jul.-Aug. 2012, pp. 279–284.[133] T. Q. Dinh, J. Tang, Q. D. La, and T. Q. Quek, “Offloading in mobileedge computing: Task allocation and computational frequency scaling,”
Commun. ACM , vol. PP, no. 99, pp. 1–14, Apr. 2017.[134] R. Kaewpuang, D. Niyato, P. Wang, and E. Hossain, “A framework forcooperative resource management in mobile cloud computing,”
IEEEJ. Sel. Areas Commun. , vol. 31, no. 12, pp. 2685–2700, Dec. 2013.[135] R. Yu, J. Ding, S. Maharjan, S. Gjessing, Y. Zhang, and D. Tsang,“Decentralized and optimal resource cooperation in geo-distributedmobile cloud computing,”
IEEE Trans. Emerg. Topics Comput. , vol. PP,no. 99, pp. 1–13, Sep. 2015.[136] M. S. Elbamby, M. Bennis, and W. Saad, “Proactive edgecomputing in latency-constrained fog networks.” [Online]. Available:https://arxiv.org/pdf/1704.06749.pdf[137] S. Wang, R. Urgaonkar, T. He, M. Zafer, K. Chan, and K. K. Leung,“Mobility-induced service migration in mobile micro-clouds,” in
Proc.IEEE Military Commun. Conf. (MILCOM) , Baltimore, MD, Oct. 2014,pp. 835–840.[138] R. Urgaonkar, S. Wang, T. He, M. Zafer, K. Chan, and K. K. Leung,“Dynamic service migration and workload scheduling in edge-clouds,”
Performance Evaluation , vol. 91, pp. 205–228, 2015.[139] M.-H. Chen, M. Dong, and B. Liang, “Joint offloading decision andresource allocation for mobile cloud with computing access point,”in
Proc. IEEE Int. Conf. Accoustic, Speech, and Signal Processing(ICASSP) , Shanghai, China, Mar. 2016, pp. 3516–3520.[140] O. Munoz, A. P-Iserte, and J. Vidal, “Optimization of radio andcomputational resources for energy efficiency in latency-constrainedapplication offloading,”
IEEE Trans. Veh. Techn. , vol. 64, no. 10, pp.497–508, Oct. 2015. [141] S. Wang and S. Dey, “Modeling and characterizing user experience ina cloud server based mobile gaming approach,” in Proc. IEEE GlobalCommun. Conf. (GLOBECOM) , Honolulu, HI, Nov.-Dec. 2009, pp.1–7.[142] S. Wang, M. Zafer, and K. K. Leung, “Online placement of multi-component applications in edge computing environments,”
IEEE Ac-cess , vol. 5, Feb. 2017.[143] Y. Shi, J. Zhang, B. O’Donoghue, and K. B. Letaief, “Large-scaleconvex optimization for dense wireless cooperative networks,”
IEEETrans. Signal Process. , vol. 63, no. 18, pp. 4729–4743, Sep. 2013.[144] C. Vallati, A. Virdis, E. Mingozzi, and G. Stea, “Mobile-edge com-puting come home connecting things in future smart homes usingLTE device-to-device communications,”
IEEE Consum. Electron. Mag. ,vol. 5, no. 4, pp. 77–83, Oct. 2016.[145] T. H. Luan, L. Gao, Z. Li, Y. Xiang, G. Wei, and L. Sun, “Fogcomputing: Focusing on mobile users at the edge.” [Online]. Available:https://arxiv.org/pdf/1502.01815v3.pdf[146] L. Tong, Y. Li, and W. Gao, “A hierarchical edge cloud architecturefor mobile computing,” in
Proc. IEEE Int. Conf. Comput. Commun.(INFOCOM) , San Francisco, CA, USA, Apr. 2016, pp. 1–9.[147] G. Kirby, A. Dearle, A. Macdonald, and A. Fernandes, “Anapproach to ad hoc cloud computing.” [Online]. Available: https://arxiv.org/pdf/1002.4738v1.pdf[148] T. T. Huu, C. K. Tham, and D. Niyato, “A stochastic workloaddistribution approach for an ad-hoc mobile cloud,” in
Proc. IEEE Int.Conf. Cloud Comput. Techn. Sci. (CloudCom) , Singapore, Dec. 2014,pp. 174–181.[149] D. M. Shila, W. Shen, Y. Cheng, X. Tian, and X. Shen, “AMCloud:Toward a secure autonomic mobile ad hoc cloud computing system,”
IEEE Wireless Commun. , vol. PP, no. 99, pp. 1–8, Oct. 2016.[150] X. Hou, Y. Li, M. Chen, D. Wu, D. Jin, and S. Chen, “Vehicular fogcomputing: A viewpoint of vehicles as the infrastructures,”
IEEE Trans.Veh. Techn. , vol. 65, no. 6, pp. 3860–3873, Jun. 2016.[151] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, andM. Franceschetti, “Stochastic geometry and random graphs for theanalysis and design of wireless networks,”
IEEE J. Sel. Areas Commun. ,vol. 27, no. 7, pp. 1029–1046, Sep. 2009.[152] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach tocoverage and rate in cellular networks,”
IEEE Trans. Commun. , vol. 59,no. 11, pp. 3122–3134, Nov. 2011.[153] M. Haenggi,
Stochastic Geometry for Wireless Networks . New York,NY, USA: Cambridge University Press, 2012.[154] C. Li, J. Zhang, J. G. Andrews, and K. B. Letaief, “Success probabilityand area spectral efficiency in multiuser MIMO HetNets,”
IEEE Trans.Commun. , vol. 64, no. 4, pp. 1544–1556, Apr. 2016.[155] N. Vastardis and K. Yang, “An enhanced community-based mobilitymodel for distributed mobile social networks,”
J. Ambient Intelligenceand Humanized Comput. , vol. 5, no. 1, pp. 65–75, Feb. 2014.[156] CISCO, “Cisco visual networking index: Global mobiledata traffic forecast update, 2015-2020,”
White paper
IEEE Commun. Mag. , vol. 52, no. 2, pp. 131–139, Feb. 2014.[158] E. Bastug, M. Bennis, and M. Debbah, “Living on the edge: The roleof proactive caching in 5G wireless networks,”
IEEE Commun. Mag. ,vol. 52, no. 8, pp. 82–89, Aug. 2014.[159] N. Golrezaei, K. Shanmugam, A. G. Dimakis, A. F. Molisch, andG. Caire, “Femtocaching: Wireless video content delivery throughdistributed caching helpers,” in
Proc. IEEE Int. Conf. Compt. Commun.(INFOCOM) , Orlando, FL, Mar. 2012, pp. 1107–1115.[160] A. S. Gomes, B. Sousa, D. Palma, V. Fonseca, Z. Zhao, E. Monteiro,T. Braun, P. Simoes, and L. Cordeiro, “Edge caching with mobilityprediction in virtualized LTE mobile networks,”
Future GenerationComput. Syst. , vol. 70, pp. 148–162, 2017.[161] L. Yang, J. Cao, G. Liang, and X. Han, “Cost aware service placementand load dispatching in mobile cloud systems,”
IEEE Trans. Comput. ,vol. 65, no. 5, pp. 1440–1452, May 2016.[162] J. Tordsson, R. S. Montero, R. Moreno-Vozmediano, and I. M. Llorente,“Cloud brokering mechanisms for optimized placement of virtualmachines across multiple providers,”
Future Generation Comput. Syst. ,vol. 28, no. 2, pp. 358–367, Feb. 2012.[163] B. Li, J. Li, J. Huai, T. Wo, Q. Li, and L. Zhong, “EnaCloud: Anenergy-saving application live placement approach for cloud computingenvironments,” in
Proc. IEEE Int. Conf. Cloud Comput. (CLOUD) ,Bangalore, India, Sep. 2009, pp. 17–24. [164] Y. Gao, H. Guan, Z. Qi, Y. Hou, and L. Liu, “A multi-objectiveant colony system algorithm for virtual machine placement in cloudcomputing,”
J. Comput. Syst. Sci. , vol. 79, no. 8, pp. 1230–1242, Dec.2013.[165] J. L. Lucas-Simarro, R. Moreno-Vozmediano, R. S. Montero, and I. M.Llorente, “Scheduling strategies for optimal service deployment acrossmultiple clouds,”
Future Generation Comput. Syst. , vol. 29, no. 6, pp.1431–1441, Aug. 2013.[166] H. Rheingold,
Virtual Reality: Exploring the Brave New Technologies .New York, NY, USA: Simon & Schuster Adult Publishing Group, 1991.[167] S. Wang and S. Dey, “Modeling and characterizing user experience ina cloud server based mobile gaming approach,” in
Proc. IEEE GlobalCommun. Conf. (GLOBECOM) , Honolulu, HI, Nov. 2009, pp. 1–7.[168] T. X. Tran, P. Pandey, A. Hajisami, and D. Pompili, “Collaborativemulti-bitrate video caching and processing in mobile-edge computingnetworks,” in
Proc. IEEE/IFIP Conf. Wireless On-demand NetworkSystems and Services (WONS) , Jackson Hole, WY, USA, Feb. 2017,pp. 1–8. [Online]. Available: https://arxiv.org/pdf/1612.01436v2.pdf[169] E. Bas¸tuˇg, M. Bennis, M. Kountouris, and M. Debbah, “Cache-enabledsmall cell networks: Modeling and tradeoffs,”
EURASIP J. WirelessCommun. Networking , vol. 2015, no. 1, pp. 1–11, Feb. 2015.[170] Y. Cui, Y. Wu, and D. Jiang, “Analysis and optimization of cachingand multicasting in large-scale cache-enabled information-centric net-works,” in
Proc. IEEE Global Commun. Conf. (GLOBECOM) , SanDiego, CA, Dec. 2015, pp. 1–7.[171] V. Suryaprakash, J. Møller, and G. Fettweis, “On the modeling andanalysis of heterogeneous radio access networks using a poisson clusterprocess,”
IEEE Trans. Wireless Commun. , vol. 14, no. 2, pp. 1035–1047, Feb. 2015.[172] D. Lopez-Perez, I. Guvenc, and X. Chu, “Mobility managementchallenges in 3GPP heterogeneous networks,”
IEEE Commun. Mag. ,vol. 50, no. 12, pp. 70–78, 2012.[173] A. Damnjanovic, J. Montojo, Y. Wei, T. Ji, T. Luo, M. Vajapeyam,T. Yoo, O. Song, and D. Malladi, “A survey on 3GPP heterogeneousnetworks,”
IEEE Wireless Commun. , vol. 18, no. 3, pp. 10–21, Mar.2011.[174] M. Kassar, B. Kervella, and G. Pujolle, “An overview of verticalhandover decision strategies in heterogeneous wireless networks,”
ELSEVIER Comput. Commun. , vol. 31, no. 10, pp. 2607–2620, Oct.2008.[175] C. Wang, Y. Li, and D. Jin, “Mobility-assisted opportunistic computa-tion offloading,”
IEEE Commun. Lett. , vol. 18, no. 10, pp. 1779–1782,Oct. 2014.[176] Y. Zhang, D. Niyato, and P. Wang, “Offloading in mobile cloudletsystems with intermittent connectivity,”
IEEE Trans. Mobile Comput. ,vol. 14, no. 12, pp. 2516–2529, Dec. 2015.[177] K. Lee and I. Shin, “User mobility model based computation offloadingdecision for mobile cloud,”
J. Comput. Sci. Eng. , vol. 9, no. 3, pp. 155–162, Sep. 2015.[178] M. R. Rahimi, N. Venkatasubramanian, and A. V. Vasilakos, “Music:Mobility-aware optimal service allocation in mobile cloud computing,”in
Proc. IEEE Int. Conf. Cloud Comput. (CLOUD) , Santa ClaraMarriott, CA, Jun. 2013, pp. 75–82.[179] A. Prasad, P. Lund´en, M. Moisio, M. A. Uusitalo, and Z. Li, “Efficientmobility and traffic management for delay tolerant cloud data in 5Gnetworks networks,” in
Proc. IEEE Int. Symp. on Personal Indoor andMobile Radio Comm. (PIMRC) , 2015, pp. 1740–1745.[180] R. Wang, X. Peng, J. Zhang, and K. B. Letaief, “Mobility-awarecaching for content-centric wireless networks: Modeling and method-ology,”
IEEE Commun. Mag. , vol. 54, no. 8, pp. 77–83, Aug. 2016.[181] S.-W. Ko, K. Huang, S.-L. Kim, and H. Chae, “Online prefetchingfor mobile computation offloading,” 2016. [Online]. Available:https://arxiv.org/abs/1608.04878[182] K. Doppler, M. Rinne, C. Wijting, C. B. Ribeiro, and K. Hugl, “Device-to-device communication as an underlay to LTE-advanced networks,”
IEEE Commun. Mag. , vol. 47, no. 12, pp. 42–49, Dec 2009.[183] C.-A. Chen, M. Won, R. Stoleru, and G. G. Xie, “Energy-efficient fault-tolerant data storage and processing in mobile cloud,”
IEEE Trans.Cloud Comput. , vol. 3, no. 1, pp. 28–41, Jan. 2015.[184] C. A. Chen, R. Stoleru, and G. G. Xie, “Energy-efficient and fault-tolerant mobile cloud storage,” in
Proc. IEEE Int. Conf. Could Net-working (Cloudnet) , Pisa, Italy, Oct. 2016, pp. 51–57.[185] D. Satria, D. Park, and M. Jo, “Recovery for overloaded mobile edgecomputing,”
Future Generation Comput. Syst. , vol. 70, pp. 138–147,2017. [186] S. Chaisiri, B.-S. Lee, and D. Niyato, “Optimization of resourceprovisioning cost in cloud computing,” IEEE Trans. Serv. Comput. ,vol. 5, no. 2, pp. 164–177, Apr. 2012.[187] Y. Zhang, J. Yan, and X. Fu, “Reservation-based resource schedulingand code partition in mobile cloud computing,” in
Proc. IEEE Int. Conf.Compt. Commun. Workshops (INFOCOM WKSHPS) , San Francisco,CA, Apr. 2016, pp. 962–967.[188] X. Jin, F. Zhang, A. V. Vasilakos, and Z. Liu, “Green data centers:A survey, perspectives, and future directions.” [Online]. Available:http://arxiv.org/abs/1608.00687[189] X. Sun and N. Ansari, “Green cloudlet network: A distributed greenmobile cloud network,”
IEEE Netw. , to appear.[190] L. A. Barroso and U. Holzle, “The case for energy-proportionalcomputing,”
Comput. , vol. 40, no. 12, pp. 651–664, Dec 2007.[191] M. Lin, A. Wierman, L. L. H. Andrew, and E. Thereska, “Dynamicright-sizing for power-proportional data centers,”
IEEE/ACM Trans.Netw. , vol. 21, no. 5, pp. 1378–1391, Oct. 2013.[192] M. Lin, Z. Liu, A. Wierman, and L. L. H. Andrew, “Online algorithmsfor geographical load balancing,” in
Proc. IEEE Int. Green Comput.Conf. (IGCC) , San Jose, CA, Jun 2012, pp. 1–10.[193] H. Xu, C. Feng, and B. Li, “Temperature aware workload managementin geo-distributed data centers,”
IEEE Trans. Parallel Distrib. Syst. ,vol. 26, no. 6, pp. 1743–1753, Jun. 2015.[194] A. Beloglazov and R. Buyya, “Energy efficient resource management invirtualized cloud data centers,” in
Proc. IEEE/ACM Int. Conf. Cluster,Cloud and Grid Computing (CCGrid) , Melbourne, Australia, May2010, pp. 826–831.[195] X. Li, J. Wu, S. Tang, and S. Lu, “Let’s stay together: Towards trafficaware virtual machine placement in data centers,” in
Proc. IEEE Int.Conf. Comput. Commun. (INFOCOM) , Toronto, Canada, Apr. 2014,pp. 1842–1850.[196] L. Chen and H. Shen, “Consolidating complementary VMs withspatial/temporal-awareness in cloud datacenters,” in
Proc. IEEE Int.Conf. Comput. Commun. (INFOCOM) , Apr. 2014, pp. 1033–1041.[197] Z. Han, H. Tan, G. Chen, R. Wang, Y. Chen, and F. C. M. Lau, “Dy-namic virtual machine management via approximate Markov decisionprocess,” in
Proc. IEEE Int. Conf. Comput. Commun. (INFOCOM) ,San Francisco, CA, Apr. 2016, pp. 1–9.[198] S. Sudevalayam and P. Kulkarni, “Energy harvesting sensor nodes:Survey and implications,”
IEEE Commun. Surveys Tuts. , vol. 13, no. 3,pp. 443–461, 3rd Quater 2011.[199] S. Ulukus, A. Yener, E. Erkip, O. Simeone, M. Zorzi, P. Grover, andK. Huang, “Energy harvesting wireless communications: A review ofrecent advances,”
IEEE J. Sel. Areas Commun. , vol. 33, no. 3, pp.360–381, Mar. 2015.[200] Y. Mao, Y. Luo, J. Zhang, and K. B. Letaief, “Energy harvesting smallcell networks: Feasibility, deployment, and operation,”
IEEE Commun.Mag. , vol. 53, no. 6, pp. 94–101, Jun 2015.[201] CNN, “Battery life concerns mobile users,” Sep. 2005. [Online]. Avail-able: http://edition.cnn.com/2005/TECH/ptech/09/22/phone.study/[202] J. Xu and S. Ren, “Online learning for offloading and autoscaling inrenewable-powered mobile edge computing,” in
Proc. IEEE GlobalCommun. Conf. (GLOBECOM) , Washington, DC, Dec. 2016, pp. 1–6.[203] Y. Mao, J. Zhang, and K. B. Letaief, “Dynamic computation offloadingfor mobile-edge computing with energy harvesting devices,”
IEEE J.Sel. Areas Commun. , vol. 34, no. 12, pp. 3590–3605, Dec. 2016.[204] C. Chen, B. He, and X. Tang, “Green-aware workload scheduling ingeographically distributed data centers,” in
Proc. IEEE Int. Conf. CloudComput. Tech. Sci. (CloudCom) , Taipei, Taiwan, Nov. 2012, pp. 82–89.[205] C. Dong, F. Kong, X. Liu, and H. Zeng, “Green power analysis forgeographical load balancing based datacenters,” in
Proc. IEEE Int.Green Comput. Conf. (IGCC) , Arlington, VA, Jun. 2013, pp. 1–8.[206] X. Sun, N. Ansari, and Q. Fan, “Green energy aware avatar migrationstrategy in green cloudlet networks,” in
Proc. IEEE Int. Conf. CloudComput. Tech. Sci. (CloudCom) , Vancouver, Canada, Nov. 2015, pp.139–146.[207] T. Chen, Y. Zhang, X. Wang, and G. B. Giannakis, “Robust workloadand energy management for sustainable data centers,”
IEEE J. Sel.Areas Commun. , vol. 34, no. 3, pp. 651–664, Mar. 2016.[208] Y. Luo, J. Zhang, and K. B. Letaief, “Transmit power minimizationfor wireless networks with energy harvesting relays,”
IEEE Trans.Commun. , vol. 64, no. 3, pp. 987–1000, Mar. 2016.[209] J. Gong, S. Zhou, and Z. Niu, “Optimal power allocation for energyharvesting and power grid coexisting wireless communication systems,”
IEEE Trans. Commun. , vol. 61, no. 7, pp. 3040 – 3049, Jul. 2013. [210] T. Han and N. Ansari, “On optimizing green energy utilization forcellular networks with hybrid energy supplies,”
IEEE Trans. WirelessCommun. , vol. 12, no. 8, pp. 3872–3882, 2013.[211] Y. Mao, J. Zhang, and K. B. Letaief, “Grid energy consumption andQoS tradeoff in hybrid energy supply wireless networks,”
IEEE Trans.Wireless Commun. , vol. 15, no. 5, pp. 3573 – 3586, May 2016.[212] W. C. Brown, “The history of power transmission by radio waves,”
IEEE Trans. Microw. Theory Techn. , vol. 32, no. 9, pp. 1230–1242,Sep. 1984.[213] H. Ju and R. Zhang, “Throughput maximization in wireless poweredcommunication networks,”
IEEE Trans. Wireless Commun. , vol. 13,no. 1, pp. 418–428, Jan. 2014.[214] K. Huang and V. K. N. Lau, “Enabling wireless power transferin cellular networks: Architecture, modeling and deployment,”
IEEETrans. Wireless Commun. , vol. 13, no. 2, pp. 902–912, May 2014.[215] Z. Chang, J. Gong, Y. Li, Z. Zhou, T. Ristaniemi, G. Shi, Z. Han, andZ. Niu, “Energy efficient resource allocation for wireless power transferenabled collaborative mobile clouds,”
IEEE J. Sel. Areas Commun. ,vol. 34, no. 12, pp. 3438–3450, Dec. 2016.[216] R. Roman, J. Lopez, and M. Mambo, “Mobile edge computing, fog etal.: A survey and analysis of security threats and challenges,”
ElsevierFuture Generation Comput. Syst. , vol. PP, no. 99, Nov. 2016.[217] S. Yi, Z. Qin, and Q. Li, “Security and privacy issues of fog computing:A survey,” in
Proc. Int. Conf. Wireless Algorithms, Systems, andApplications (WASA) , Qufu, China, Aug. 2015, pp. 1–10.[218] M. M. Fouda, Z. M. Fadlullah, N. Kato, R. Lu, and X. S. Shen, “Alightweight message authentication scheme for smart grid communi-cations,”
IEEE Trans. Smart Grid , vol. 2, no. 4, pp. 675–685, Dec.2011.[219] A. M. Y. Ahmed and D. Qian, “An optimization of security andtrust management in distributed systems,” in
Proc. IEEE Int. AdvanceComputing Conf. (IACC) , Ghaziabad, India, Feb. 2013, pp. 120–126.[220] X. Huang, Y. Xiang, E. Bertino, J. Zhou, and L. Xu, “Robust multi-factor authentication for fragile communications,”
IEEE Trans. De-pendable Secure Comput. , vol. 11, no. 6, pp. 568–581, Nov.-Dec. 2014.[221] M. C. Gorantla, C. Boyd, and J. M. G. Nieto, “Attribute-basedauthenticated key exchange,” in
Proc. Australian Conf. Info. Securityand Privacy (ACISP) , Sydney, Australia, Jul. 2010, pp. 1–25.[222] H. M. Pimentel, S. Kopp, M. A. S. Jr., R. M. Silveira, and G. Bressan,“OCP: A protocol for secure communication in federated contentnetworks,”
Compt. Commun. , vol. 68, pp. 47–60, Sep. 2015.[223] M. Liyanage, A. B. Abro, M. Ylianttila, and A. Gurtov, “Opportuni-ties and challenges of software-defined mobile networks in networksecurity,”
IEEE Security Privacy , vol. 14, no. 4, pp. 34–44, Jul. 2016.[224] W. Yang and C. Fung, “A survey on security in network functionsvirtualization,” in
Proc. IEEE NetSoft Conf. Workshops (NetSoft) ,Seoul, Korea, Jun. 2016, pp. 15–19.[225] B. Liang, “Mobile edge computing,” in
Key Technologies for 5GWireless Systems ∼ IEEE Trans. Comput. , vol. 65, no. 1, pp. 216–229, Jan. 2016.[228] R. Gennaro, G. Craig, and P. Bryan, “Non-interactive verifiable com-puting: Outsourcing computation to untrusted workers,” in
Annu. Conf.Adv. Cryptol. , Apr. 2017. [On-line]. Available: https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.aspx?specificationId=3144[236] A. Anjum, T. Abdullah, M. Tariq, Y. Baltaci, and N. Antonopoulos,“Video stream analysis in clouds: An object detection and classificationframework for high performance video analytics,” IEEE Trans. CloudComput.
IEEECommun. Mag. , vol. 47, no. 11, pp. 84–95, Nov. 2009.[239] N. Lu, N.Cheng, N. Zhang, X. Shen, and J. W. Mark, “Connectedvehicles: Solutions and challenges,”
IEEE Internet Things J. , vol. 1,no. 4, pp. 289–299, Apr. 2015.[240] E. Uhlemann, “Introducing connected vehicles,”