Research ArticleTask Scheduling for Edge Computing with Agile VNFsOn-Demand Service Model toward 5G and Beyond

Chia-Wei Tseng 1 Fan-Hsun Tseng 2 Yao-Tsung Yang 1

Chien-Chang Liu1 and Li-Der Chou 1

1Department of Computer Science and Information Engineering National Central University Taoyuan 32001 Taiwan2Department of TechnologyApplication andHumanResourceDevelopment National TaiwanNormalUniversity Taipei 10610 Taiwan

Correspondence should be addressed to Li-Der Chou cldcsiencuedutw

Received 16 April 2018 Accepted 28 June 2018 Published 11 July 2018

Academic Editor Chun-Hsian Huang

Copyright copy 2018 Chia-Wei Tseng et alThis is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

The demand for satisfying service requests effectively allocating computing resources and providing service on-demandapplication continuously increases along with the rapid development of the Internet Edge computing is used to satisfy the lowlatency network connection and local data processing requirements and to alleviate the workload in the cloudThis paper proposesa gateway-based edge computing service model to reduce the latency of data transmission and the network bandwidth from andto the cloud An on-demand computing resource allocation can be achieved by adjusting the task schedule of the edge gateway viathe lightweight virtualization technology DockerThe edge gateway can also process the service requests in the local network Theproposed edge computing servicemodel not only eliminates the computation burden of the traditional cloud servicemodel but alsoimproves the operation efficiency of the edge computing nodes This model can also be used for various innovation applications inthe cloud-edge computing environment for 5G and beyond

1 Introduction

The rapid development of cloud computing mobile broad-band networks and Internet ofThings (IoT) has been accom-panied by a considerable demand for network resourcesallocation data processing and service management and hasinevitably changed the traditional network infrastructureBoth software-defined network (SDN) and network functionvirtualization (NFV) technologies not only transform thenetwork infrastructure from complicate physical entitiesinto virtual and programmable nodes but also introducesignificant changes to the development of the informationand communications technology (ICT) [1 2]

NFV represents a core structural change in how atelecommunication infrastructure is deployed [3] As shownin Figure 1 NFV aims to solve the problems related tothe distribution of network device resources SpecificallyNFV uses virtualization technology to integrate differenttypes of devices that are located on the data center networknodes and customer-premises equipment (CPE) including

routers switches firewalls and intrusion-detection systemsinto standard or general hardware equipment In this wayNFV can efficiently reduce the purchase cost of networkphysical devices and the related network maintenance costsAs the logical outcome of NFV virtual network functions(VNF) are virtualized tasks that are formerly performedby proprietary dedicated hardware VNFs move individualnetwork functions from dedicated hardware devices to asoftware that runs on commodity hardware MeanwhileNFV decouples the network functions from the proprietaryhardware appliances to run them by using a software andto subsequently accelerate service innovation and provision-ing especially within service provider environments Theemergence of NFV benefits the software-operated networksinstead of those operated by physical devices and enhancesthe flexibility and rapid deployment of each network servicefunction SDN is another technology that is used to solvethe traditional network rigidity problem and to satisfy therequirements of diversified and dynamic network services[4]Themain purpose of SDN is to solve the limitations of the

HindawiWireless Communications and Mobile ComputingVolume 2018 Article ID 7802797 13 pageshttpsdoiorg10115520187802797

2 Wireless Communications and Mobile Computing

Virtual Function

Virtual Function

Standard or GeneralHardware Equipment

Server CPE Storage

Virtual Function

FW IDS

Router Switch

NAT DHCP

FW

Figure 1 Concept of NFV

Cloud

Fog ComputingCloudlets

Data Center

Edge

IoT Devices

Small cellGWGW

Data Center Storage

MDCMDC

Mobile-Edge Computing

MDC

Cloud Massive Cloud

Figure 2 Concept of edge computing

traditional network architecture where the network is dividedinto control and data planes The software rearranges thenetwork architecture by granting the network administrationauthority to the software controller on the control planethrough a centralized control approach In this way thiscontrol mode which is withdrawn from the existing networkarchitecture and transformed into a programmable networkeffectively addresses the limitations of the traditional networkdesign A centralized control and flexible operation of thenetwork can be achieved by using a single logic point tocontrol the entire network SDN and NFV are networkingtechnologies that can work with each other to lead thetransformation deployment and management of the net-work design under the environment of data transmissionand telecommunication infrastructure The integration ofthese technologies also presents a direction for the futuredevelopment of the Internet especially after the emergenceof the 5G technology [5 6]

As the network becomes increasingly flexible softwaredefined and virtualized several standards organizationsare working to introduce the SDNNFV technology into

mobile networks to satisfy the low latency requirement ofthe 5GIoT network for data processing and informationtransmission and to indirectly drive the development ofthe edge computing technology [7ndash9] Edge computing isa distributed computing architecture that moves computingpower for applications data and services from a cloud datacenter server to the CPE that is located close to the edgeof the user network to be processed [10ndash12] Gateway isa type of CPE and common-edge computing device thatpossesses basic capabilities in computing analyzing andpreprocessing the data that are collected near hosts and otherIoT devices to accelerate the data processing and to reducethe transmission latency [13 14] In the gateway-based edgenetwork architecture the data are computed and processednear their sources thereby making this architecture suitablefor processing immediate service requests

The concept of edge computing is shown in Figure 2The edge network is located between IoT devices and thecloud [15] Edge computing extends the existing cloud com-puting paradigm to the network edge in order to satisfy therequirements of latency-critical and computation-intensive

Wireless Communications and Mobile Computing 3

IoT applications The distributed architecture of edge com-puting nodes satisfies the computing requirements for severalapplications data and services from the cloud data centerto the CPE or the micro data center Based on differentfields of development the edge computing network canbe divided into cloudlets mobile edge computing (MEC)and fog computing [16ndash18] Cloudlets are communicationtools that provide microservices to the network surroundingthe user and virtualize and shrink the computing resourceto deploy the computing resource near the end of theuser The application services and technologies of cloudletsare currently being promoted by Akamai and MicrosoftIntroduced by the European Telecommunications StandardsInstitute (ETSI) MEC is highly applicable in the field ofmobile communication and is operated and managed bycommunication companies [19]This technologymainly aimsto reduce the increasing pressure on network equipment andto help mobile companies create a unique mobile servicemodel Introduced by Cisco and currently promoted by theOpenFog Consortium Alliance fog computing is a conceptof extended cloud computing that focuses on the data pro-cessing function of the local network [20] By emphasizingnear-field data communication fog computing can be usedin different network equipment for personal or businessmanagement and for providing relevant IoT services inspecific areas The OpenFog Consortium aims to promotethe OpenFog computing technology by developing an openarchitecture that can identify certain core technologies andcapabilities such as distributed computing network andstorage which provide intelligent support at the edge of theIoT Keeping in line with the current trends in technologyintegration OpenFog has also cooperated with the ETSIIndustrial Specification Group to establish an applicationframework and development technology for the specificationand interoperability of ICT and for expanding the applicationscope of edge computing

The emergence of the 5G technology has increasedthe requirements for data movement data storage dataprocessing and data analysis for the traditional central-ized cloud computing architecture thereby increasing therequired number of hardware devices to improve the loadcapacity of the core network and the servers Given the largeamount of data generated from IoT the existing cloud servicemodel cannot satisfy the large and dispersed requests ofusers This model has been used for data computing anal-ysis and processing However the demand for low latencyand immediacy has gradually increased thereby renderingthis model slightly insufficient The introduction of edgecomputing helps process data in advance and reduce datatraffic and transmission time this technology also allows thetransfer of computing power to the CPE or terminal nodesto enhance the immediate feedback of the edge operations tothe environment [21] The Cisco Internet Business SolutionsGroup [22] predicts that the IoT will have more than 50billion terminals and devices by 2020 while the InternationalData Corporation estimates that more than 40 of the datain the future need to be analyzed processed and stored atthe edge of the network [23] To effectively respond to thechallenges brought upon by the aforementioned changes the

service model in the edge of the network must be redefinedto optimize the network operation and service processingefficiency to reduce the workload in the cloud and to satisfythe latency requirements of 5G IoT

The rest of this paper is organized as follows Section 2presents the research background and the related worksSection 3 describes the design of the edge computing servicemodel Section 4 presents the experiment results Section 5concludes the paper

2 Background and Related Works

Network operators must deploy various types of networkequipment and software platforms to support increasinglydiverse Internet requirements However the current networkdesign is based on hardware concept various hardwaredevices are often incompatible or unable to communicatewith one another and network maintenance managementhas a high complexity and cost NFV has been recentlyproposed as a solution to the difficulties being faced bynetwork operators Through NFV network and commu-nication operators can decrease the related costs flexiblyand efficiently process the customer network deploymentrequirements reduce the waste of hardware resources andincrease the resource utilization efficiency The current NFVrelated standards are mainly driven by ETSI Given theimpact and benefits brought upon by the virtualization ofnetwork functions operatorsmdashparticularly network serviceproviders and telecommunication operatorsmdashare consideredthe most active participants in the development of NFVstandards

As shown in Figure 3 hypervisor and container are thetwo main frameworks for VNFs [24] Hypervisor is themost common virtualization technology that uses a group ofvirtualization hardware devices to execute the host operatingsystem in a virtual manner and to abstract the physical hard-ware of the host operating systemThe virtual machine (VM)is the operating system deployed on the virtual managementprogram while the Linux container (LXC) is an operating-system-level virtualization technology that mainly packagesthe application service system into a software container thatstores the codes of the application service software requiredoperating system core and library As its most attractivefeature LXC has a simplified software structure Thereforethe image file of the container only includes a small part ofthose operating system components that are smaller than thegeneral VM which is a complete operating system Given itssmall size and lightweight features the container has a muchfaster deployment or movement speed compared with VMAccordingly most industry practitioners prefer the containerover the VM Docker is an open source project based onLXC technology [25] that is operated on the Linux operat-ing system This technology creates an additional softwareabstraction layer for virtualizing the service application inresponse to the development of microservices [26] Asidefrom being a famous product of the Git Hub communityDocker has also been favored by Google and Red Hat and iscurrently the fastest growing NFV virtualization technologyMany studies have also been conducted to examine the

4 Wireless Communications and Mobile Computing

Hardware

Host OS

Hypervisor(VMM)

GuestOS

BinariesLibraries

App

System Virtualization

GuestOS

BinariesLibraries

App

GuestOS

BinariesLibraries

App

Hardware

Host OS

BinariesLibraries

App

BinariesLibraries

App

BinariesLibraries

App

Container Virtualization

namespace

Figure 3 Common approach for the hosting of VMs versus containers

implementation of Docker as an edge computing platform[27 28]

The development of virtualization technologies hasdriven several innovations in the IoT architecture Edgecomputing has also emerged as a new concept that satisfiesthe requirements of IoT for distributed data processing andnumber of devices CPE devices such as the edge gatewayuse virtualization technology to provide local operations thatare particularly suitable for IoT applications with low latencyand immediacy Edge computing can completely disassemblelarge-scale services that are originally handled by cloud datacenters cut these services into small and easy-to-manageparts and distribute them among the edge network nodes ormicro data centers to accelerate their resolution and to reducethe workload in the cloud In this case the service requestsof users are processed on the gateway of the edge networkwhile those requests that exceed the computing capabilitiesor cannot be handled by the edge network are forwarded tothe cloud for processing Edge computing can also be used toreduce the amount of data being sent back to the cloudThesedata are initially processed at the edge node and are thenpreprocessed or filtered before they are transmitted back tothe cloud to ensure that only useful data will be transmittedEdge computing not only reduces network connection delaysand meets the demand of 5GIoT for improving delays butalso promotes the convergence of ICT industry technologiesUsing lightweight virtual technology also helps networkoperators install applications and control software on edgegateways provide localization services and create innovativeservice models Reference [29] proposes a SDN-based edgecomputing architecture called software-defined infrastruc-ture (SDI) which usesOpenFlow andOpenStack to virtualizeservice resources for building smart applications on virtualinfrastructure that can flexibly schedule network resourcesTo introduce the concept of resource sharing [30] proposesa cloud-edge framework where each edge node can use thelocal processing platform to share computing resources toreduce the computational burden on the cloud that processes

the data and to improve the operation of the edge networkTomeet the low latency and fast data processing requirementsof mobile networks [12] examines mobile edge operationsidentifies the applicable situations and reference scenarios fordifferentMECs and proposesmarginalities forMEC offload-ing decisions computing resource allocation and mobilitymanagement The computational needs of the Internet canalso be used as reference for examining edge network servicemanagement and application innovation

Given the resource constraints (eg CPU memory andnetwork-intensive service requests) of the local gatewayhardware the task scheduling mechanism for handling theservice requests of users plays an important role in improvingthe service capabilities of edge computing Several algorithmsfor task scheduling are available in a computer system [31] In-depth studies on several task scheduling algorithms for cloudcomputing have also been conducted [32 33] Reference[34] proposes a task scheduling approach called HealthEdgewhich sets various processing priorities for different tasksbased on the collected human health data and determineswhether a task must be operated on a local device or a remotecloud in order to reduce the total processing time as muchas possible Reference [35] adopts a Markov decision processapproach to solve the stochastic optimization problem in theMEC system In this approach the computation tasks arescheduled based on the queuing state of the task buffer theexecution state of the local processing unit and the stateof the transmission unit Reference [36] proposes a greedybest availability (GBA) mechanism to identify the idealizedtask scheduling policy and to reduce the queuing time ofservices by scheduling the tasks according to their workcompletion time Reference [37] analyzes the performanceof the round-robin (RR) algorithm in the cloud computingenvironment and reveals that RR scheduling fairly allocatescomputing resources among tasks of the same priority byusing the time slicing approach The results show that theRR algorithm demonstrates a better response time and loadbalancing compared with the other algorithms

Wireless Communications and Mobile Computing 5

Virtual Network Functions

VNF

Hardware Resource

Com

putin

g Re

sour

ce

Service RequestVNF

VNF

VNF

R

R

R

R

Figure 4 Network function scheduling problem in CPE

3 Problem Definition

NFV is currently the most valued solution to the prob-lems associated with the operation cost and efficiency ofnetwork services Accordingly many ICT companies havebegun to examine virtualization technology Despite thehigh expectations toward this technology satisfying servicerequests effectively allocating virtual computing resourcesand providing service on-demand application still challengethe deployment of NFV and play key roles in the futuredevelopment of 5GIoT services NFV also allows the estab-lishment of multiple independent heterogeneous virtualnetworks based on common underlying network resourcesthereby enabling service providers to provide customizedservices according to the demands of users Virtual networkembedding is a process of mapping the virtual networkto the underlying network (substrate network) through themapping algorithm and according to the current resourcesituation of the infrastructure providerThis process is anNP-Hard problem and has been investigated in NFV resourceallocation research [38 39] Previous studies [40 41] havealso systematically discussed the problem of virtual networkmapping and provided good references for research

This paper aims to optimize task scheduling and resourceallocation by using the proposed edge computing servicemodel Scheduling tasks in edge computing aremore complexthan that in cloud computing An edge computing operationis typically spread over the device of the client the edgegateways and occasionally a broker of the cloud networkTherefore deciding where to schedule computational tasksremains a key problem in edge computing Given that thegateway-based IoT architecture is currently the mainstreamthe task scheduling mechanism discussed in this paperfocuses on the edge gateway Task scheduling and resourceallocation are the main problems to be solved in this gatewaygiven the limited amount of available computing resourcesLightweight NFV technology plays an indispensable role inthe rapid deployment of the edge gateway The relationshipbetween task scheduling and resource allocation for hostservice requests is plotted in Figure 4 As each service requestarrives at a different time the required VNF service andprocessing time also differ The edge gateway can adjust thetask scheduling and resource configuration according to the

differences among the service requests As the basic idea forthe service on-demand model only one VNF is dedicated toa single service request This service model problem can beanalyzed by using queuing theory which regards the requestand react processing as a waiting-line system the input of awaiting-line system as the service request the service counteras the gateway scheduling function and the output as therequested VNF resource Although many queuing modelsmay be used in operations management [42] this paperprimarily focuses on the task scheduling approach that allowsa certain edge gateway to processmultipleVNFs fromaqueueone after another Given its limited capacity the edge gatewaycan only schedule a limited number of service requests whilethe subsequent service requests need to be forwarded to thecloud for processing With these considerations the researchproblem is formulated as follows Assume that the servicerequest is Poissonrsquos ratio within one edge gateway and thatthe service time is an exponential distribution The servicerequest set can be denoted as 119877 = 1198771 1198772 1198773 119877119899 wheren denotes the number of service requests in the systemEquations (1) and (2) are used to determine the probabilityof n service requests in the system In these equations Ndenotes the maximum number of service requests that canbe scheduled in the gateway 120582 is the average number ofincoming user requests in one unit of time 120583 is the serviceefficiency (the ability of the service counter) 120588 = 120582120583 is theratio that the request can be met within one unit of time P0denotes the initial condition L denotes the total number ofservice requests arriving in the systemwithin a planned timeand 119871119902 denotes the total number of server requests queued inthe system L and 119871119902 can be computed by using (3) and (4)respectively Equation (5) computes the waiting and servicetime of a service in the system while (6) computes119882119902 or theaverage waiting time for a user request

1198750 =

1 minus 1205881 minus 120588119873+1 120588 = 11119873 + 1 120588 = 1

(1)

119875119899 =

1205881198991198750 = 1205881198991 minus 1205881 minus 120588119873+1 120588 = 1

1198750 =1119873 + 1 120588 = 1

(2)

6 Wireless Communications and Mobile Computing

Binaries Libraries

Host OS

Hardware Resource

VNF

Docker Engine

ResourceEstimation Scheduler VNF

Configuration

VN

F Manager

Container images

R

R

R

R

R

Edge Computing Node VNFs Resource PoolService Request

VNF

VNF

PublicRegistry

PrivateRegistry

SDN Controller

Figure 5 Service on-demand edge computing model

119871 =

1205881 minus 120588 minus

(119873 + 1) 120588119873+1

1 minus 120588119873+1 120588 = 11198732 120588 = 1

(3)

119871119902 = L minus (1 minus 1198750) (4)

119882 = 119871120582 (1 minus 119875119899)

(5)

119882119902 =119871119902120582

(6)

The results reveal that the waiting time for the userrequest can be reduced in two ways namely by reducing thenumber of service requests queued in the edge gateway andby increasing the processing speed of task scheduling in theedge computing gateway A large number of tasks must becompleted within a short period to achieve an efficient edgecomputing

4 Design of the Edge ComputingService Model

How to construct an elastic and cost-effective edge com-puting service model improve management efficiency meetuser service requests achieve centralized management anddevelop flexible configuration service models are the currentresearch trends in the development of a 5G SDNNFV net-work This paper proposes a gateway-based edge computingservicemodel (Figure 5) to improve the operational efficiencyof the edge computing node to accelerate the processingof user service requests and to increase the utilizationefficiency of a limited number of computing resources In thismodel when different user requests enter the edge gatewaythis gateway determines whether the requested services canbe processed or not If the edge gateway itself lacks thecomputing capacity or resources then the controller forwardsthe service request to the cloud to reduce the data processinglatency

The proposed edge computing service model can bedivided into resource estimation scheduler and lightweightVNF configuration Figure 6 presents the flowchart of theoperations in the proposed edge computing service modeland these three parts are further described in the followingsections

Resource estimation checkswhether the edge gateway hasa sufficient amount of computing resources to provide edgecomputing services For the user request set R the resourceallocation must abide by the following rules which are alsoused by the edge gateway

(1) For a single service request Ri any resource of Vi(eg CPU memory and disk) is less than the totalresources of P forall119877119862119894ltPC forall119877119872119894ltPM and forall119877119863119894ltPD119877119862119894 119877119872119894 and 119877119863119894 denote the CPU memory and diskspace requests sent to Ri respectively

(2) The sum of computing resources that are allocated tothe VNF in the gateway is less than the total numberof resources of the physical machine P 119881119862119894 lt PC 119881119872119894lt PM 119881119863119894 lt PD 119881119862119894 119881119872119894 and 119881119863119894 denote the sum ofCPU memory and disk space resources allocated toVi respectively

As shown in Figure 6 after receiving the user servicerequest the edge gateway checks whether a sufficient amountof computing resources is available to satisfy such requestThe following situations may be encountered in this case

(1) If the edge gateway has a sufficient amount ofresources then the user service request is processedthrough the scheduler and queued in the systembased on the result of the task scheduling algorithm

(2) If the edge gateway does not have a sufficient amountof available resources then the user service request isdirectly transferred to the cloud instead

Task scheduling aims to increase the operational effi-ciency of the edge gateway Given that one service requestdiffers from another task scheduling examines how the

Wireless Communications and Mobile Computing 7

Forward Ri to Cloud DC

Start

Download from private storage

RCilt(PCiminusVCi)

RMilt(PMiminusVMi)

RDilt(PDiminusVDi)

No

Yes

Yes

Yes

Resource Estimation for Ri

VNF allocation

Download from public storage

End

No

No

No

Yes

Task Scheduling Docker Image is exist

Riconfigured to

the scheduler No

PCi = (PCiminusRCi)PMi = (PMiminusRMi)PDi =(PDiminusRDi)

Service on-demand VNF Matching

Yes

Figure 6 Edge computing service model flowchart

edge gateway can meet the requirements of different servicerequests and accomplish task scheduling in the system As itsprimary purpose scheduling attempts to reduce the amountof time spent on dealing with the most demanding servicerequirements asmuch as possible For this purpose this paperconstructs the Greedy Available Fit (GAF) task schedulingmechanism to enhance the operational efficiency of edgecomputing services

Assume that each service request Ri configures a VNFvirtualization service resource Vi ti denotes the processingtime of the ith service request di denotes its deadline andj denotes its completion time in the system The Ri in thesystem must be completed at schedule and before the basictime limit Otherwise this service request must be forwardeddirectly to the cloud As each service request arrives at adifferent time the time required for the operation processingand the deadline time also differ In this case deadline isselected as a priority parameter for the task scheduling Thedeadline can be used to define the processing priority forservices that is the deadline for those services that requirereal-time processing may be set according to their processingand precedence requirements on different VNFs This paperaims to insert as many tasks as possible into the schedulebefore completing the largest task Ri of the deadline Toaccomplish this objective a task with the maximum deadlineis selected as the baseline and the remaining tasks aresequentially inserted into the queue based on their processing

time119873[119894 119895] denotes the maximumnumber of j tasks selectedfrom the front i tasks and can be formulated as

119873[119894 119895] = max

119873[119894 minus 1 119895]

119873 [119894 minus 1 119895 minus 119905119894] + 1 119894119891 119895 le 119889119894119873[119894 minus 1 119895 minus 119905119894] 119894119891 119895 gt 119889119894

(7)

Equation (8) shows the initial condition of119873[119894 119895]

119873[1 119895] =

minusinfin 119894119891 119895 = 11990511 119894119891 119895 = 1199051 le 11988910 119894119891 119895 = 1199051 gt 1198891

(8)

Assume that there is no time gap betweenRi and119877119894+1 thatR1 starts from time 0 and that N = 1 If t1 le d1 then N = 0because the deadline is exceeded

Thewhole decision process is summarized inAlgorithm 1which starts with an empty schedule and inserts the availabletasks into this schedule in three different cases In case 1 if thecurrent time step j exceeds the deadline of Ri then N[i j] =N[iminus1 jminusti] In case (2) if there is not enough time to finishrequest Ri by the current time j then N[i j] = N[iminus1 j] Incase (3) if time j does not exceed the deadline of Ri and thereis enough time to finish request Ri by time j then N[i j] =max(N[iminus1 j] N[iminus1 jminusti ] + 1)

8 Wireless Communications and Mobile Computing

Input 119877 = 1198771 1198772 119877119899 ti di jOutput119873[119894 119895](1) Start(2) Set the largest task di as the baseline of the scheduler(3) for 119894 larr997888 1 to n initialize the maximum N at time 0 to be zero for each service

request in queue(4) 119873[119894 0] = 0(5) for 119895 larr997888 1 to 119889119899 determine the maximum N obtained by R1 at each time step(6) if ((j == t1) and (j lt= d1))(7) 119873[1 119895] = 1(8) else(9) 119873[1 119895] =119873[1 119895 minus 1](10) for 119894 larr997888 2 to n(11) for 119895 larr997888 1 to 119889119894(12) if (j gt di) case (1) the current time step j already exceeds 119877119894rsquos deadline(13) then119873[119894 119895] = 119873[119894 minus 1 119895 minus 119905119894](14) else if (j lt ti) case (2) there is not enough time to finish request Ri by the

current time j(15) then119873[119894 119895] = 119873[119894 minus 1 119895](16) else case (3) time j does not exceed 119877119894rsquos deadline and there is enough time

to finish request 119877119894 by time j(17) then119873[119894 119895] = max(119873[119894 minus 1 119895]119873[119894 minus 1 119895 minus 119905119894] + 1)(18) Schedule Completed(19) End

Algorithm 1 Greedy Available Fit algorithm

Virtualization technology can be applied on the CPE inmany ways such as by using the VM container or VM inte-gration of container However a traditional VM consumesmany system resources and cannot meet the requirementsfor light weight and service on-demand deployment In thiscase this research adopts container technology instead Acontainer handles only one service request at a time and stopsits operation after completing the delivery of a service TheVNF manager on the gateway is responsible for configuringand allocating each VNF The VNF template image that isrequired by different services is placed in the VNF resourcepool (with Docker Hub as the default other public or privateregistrations can also be specified) The edge gateway notonly controls the amount of gateway resources used byeach container but also manages several resources such asCPU and RAM to ensure that the container can obtain therequired resources without affecting the performance of theother executing containers on the edge gateway

Linux containers adopt a hierarchical structure to rapidlydeploy VNFs and to manage NFV flexible scheduling Theunderlying structure uses the file archiving mechanism ofDocker namely the advanced multilayered unification filesystem to incorporatemany different VNF imagesThe layersare stacked up andwhen someVNF service functions need tobe accessed the container retrieves the VNF images throughthe VNFmanager and uses these images directly To integratethe open resources of Docker Hub and to expand the privatewarehouse of the VNF service manager the system can storethe image files obtained from the public warehouse in aprivate warehouse for future use The judgment mechanismof the VNF service manager is described as follows

(1) If the image requested by the host is already in thelocal container then the VNF image can be taken outdirectly by the Docker daemon of the VNF servicemanager No complex configurations are required

(2) If the VNF image to be used by the host is not in thelocal container then the VNF manager connects tothe common warehouse Docker Hub on the Internetto find and store the desired service image file ina private warehouse via pushpull Afterward theacquired VNF image is controlled by the Dockerdaemon

By using the OS kernel for resource isolation the con-tainer technology does not need to rely on a virtual softwarelayer and does not require the VM to install a guest OSTherefore the capacity of the image file is much larger thanthat of the virtual machine The image file is small and canbe rapidly deployed through network transmission therebysaving network resources By using Linux container technol-ogy we can configure various VNF images to be executed onthe same edge gateway thereby replacing the virtual machineand achieving the goal of lightweight virtualization

5 Experiment

An experiment is conducted to evaluate the proposed taskscheduling algorithmand to test its performance in deployinglightweight VNF The experiment environment is shown inFigure 7 A simulation is initially performed to evaluate thetask scheduling performance of the algorithm by changingthe number of service requirements The service request is

Wireless Communications and Mobile Computing 9

Edge Network

Internet

Edge Gateway

DataCenter

Cloud Broker

Docker HubDocker Hub

Figure 7 The experiment environment

characterized as a Poisson process There we showed that theaverage number of service requests in the system is given by120582120583 where 120582 is the average arrival rate and 120583 is the averageservice rate The input parameter 120582120583 is 045 First comefirst service (FCFS) priority task scheduling [34] RR [37]and GBA [36] are then compared with the proposed GAFmechanism

The effect of the service requests on the average waitingtime average response time and task scheduling of differentscheduling mechanisms is evaluated in the simulation FCFSis based on the service requests that enter the edge gatewayqueuing system and schedules the tasks according to thesequence of these requests Priority task scheduling is anunfair scheduling algorithm where the service requests aresorted according to their priority and where the high-prioritytasks are performedfirstThose tasks having the same priorityare sorted by using the FIFO scheduling mechanism TheGBA sorts the tasks based on the completion time of theservice requests in queuing systemTheRR algorithm is basedon the conventional RR scheduling performed in the processscheduling RR scheduling fairly allocates the computingresources to those tasks with the same priority by using thetime slicing approach (time quantum)

Figure 8 shows the experiment result for average waitingtime where the x-axis refers to the number of service requestsand the y main axis refers to the average waiting time thatis the time a user spends to complete the service schedulingand to determine the resource allocation after receiving aservice request and a successful reply If the demand revertsto an errortimeout or if the edge gateway does not haveenough computing resources to support such demand thenthe sample is excluded from the calculation of the averagewaiting time As can be seen in the experimental resultthe RR algorithm obtains the longest average waiting timebecause of its application of time slice rules in order for eachtask to be processed within a fixed amount of time The timequantum affects the overall performance of the operation

0100200300400500600

10 20 30 40 50 60 70 80 90 100

Aver

age W

aitin

g Ti

me (

ms)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 8 The experiment result for average waiting time

Having a very long time quantum leads to a very long waitingtime while having a very short time quantum results intask schedule conversion a poor execution efficiency and anextended waiting time For FCFS given that the time spentdiffers across each task the waiting time for the next taskmust be determined based on the schedule of the previoustask Therefore if the last task schedule is too long thesystem cannot quickly process the subsequent task schedulesthereby affecting the overall task scheduling efficiency Giventhat its average waiting time is relatively shorter than thatof the FCFS mechanism priority task scheduling can satisfythe task requirements and flexibly perform the schedulingHowever prioritizing a high-priority program will delay allof the low-priority requests thereby creating an indefinitesituation in the low-priority program and extending thewaiting time GBA is based on the earliest completion time ofthe taskWhen the service demand is low the average waitingtime of the GBA algorithm is similar to that of priority taskscheduling However when the service demand increases theaverage waiting time of the GBAmechanism becomes longerthan that of priority task scheduling Compared with FCFS

10 Wireless Communications and Mobile Computing

050

100150200250300350

10 20 30 40 50 60 70 80 90 100

Ave

rage

Res

pons

e Tim

e (m

s)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 9 The experiment result for average response time

01020304050607080

10 20 30 40 50 60 70 80 90 100

Num

ber o

f Sch

edul

ed T

asks

Number of Service RequestsGAFFCFSPriority

GBARR

Figure 10 The experiment result for task scheduling efficiency

priority task scheduling RR and GBA the proposed GAFalgorithm selects the largest deadline task as the baseline andthen sequentially inserts the remaining tasks into the queuebased on their processing timeTherefore the GAF algorithmobtains the shortest average waiting time

Figure 9 shows the results for average response time Asexpected the RR algorithm outperforms all other approachesin terms of average response time because this algorithm isespecially designed for time sharing Using a fair strategy toallocate VNF resources ensures that each service request hasa fixed time If the service processing time exceeds the timequantum allocated by the system then the sample is excludedfrom the calculation of the average response time Given thatFCFS is processed in sequence the average response timewill be affected by the time spent by the previous 119877119894minus1 taskin the scheduler For example if the previous service requesthas a very long task schedule then the average responsetime increases Given that the proposed GAF algorithmsequentially inserts tasks into the queue its average responsetime is slightly longer than that of FCFS and GBA Althoughthe priority algorithm can sort the tasks according to thepriorities of the service requests the low-priority tasks areeasily delayed thereby increasing the average response time

Figure 10 shows the task scheduling efficiency of thecompared algorithms If the service request number is lessthan 60 then each algorithm can complete the task schedul-ing However when this number is exceeded the number

of tasks that can be scheduled varies across each algorithmGAF outperforms all the other algorithms as it ranks thelargest deadline task at the end and then sequentially sortsthe remaining tasks according to their processing time In thisway this algorithm ensures that most tasks will be completedwithin a minimum processing time Those service requeststhat cannot be scheduled will bypass the gateway and will bedirectly forwarded to the cloud data center When the servicerequest number is 80 GAF can complete 68 of the taskscheduling Meanwhile GBA performs the scheduling basedon the earliest task completion time Although GAF andGBAcan schedule the same number of tasks in the experimentalenvironment the latter has a longer average waiting timecompared with the former GAF also outperforms FCFCand priority task scheduling by 96 and 62 in terms oftask scheduling efficiency respectively because this approachranks the largest deadline task at the end and then sorts theremaining tasks according to their required processing timeGAF also outperforms theRRmechanismby 70because thelatter applies the time-sharing rule to minimize the numberof scheduled tasks Based on these results the GAF algorithmcan schedule more tasks in the edge gateway within a shortaverage waiting time and improve the operational efficiencyof edge computing services

In addition to task scheduling the VNF configuration isanother main factor that affects the operational efficiency ofthe edge computing service model This paper utilizes theDocker technology to achieve an agile deployment of VNFsSuch technology can provide a centralized management andflexible configuration of VNF services Depending on theservice requests of each user the deployment of differentVNFs can meet the application requirements of serviceson-demand The VNF image of Docker can be mainlydivided into four categories namely system images (egUbuntu CentOS and Debian) tool images (eg GolangFlask and Tomcat) service images (eg MySQL MongoDBand RabbitMQ) and application images (eg WordPressand DRUPAL) To test the practical deployment efficiency oflightweight VNF for the edge gateway a pull performancetest is performed for each type of Docker VNF image Theedge gateway is simulated by using a standalone PC with 34GHz dual-core CPU 16GBmemory and 5400 rpmHD500GThe experimental network bandwidth is set to 100 Mbps Tocompare the pull performance with the local edge gateway atest VM in the Google Cloud platform is used to simulate thecloud network broker Table 1 presents the result of the pullperformance test

As shown in Table 1 each VNF image size has differentphysical gateway and cloud network broker costs due to thelimitations in the network bandwidth A larger VNF imagetakes a longer time to be downloaded from Docker Hub Asexpected the cloud network broker has a better computingarchitecture than the local edge gateway because the locationof the edge gateway affects the performance of the VNFdeployment Although the edge computing architecture ofthe cloud network broker deploys VNF more efficiently thanthe local gateway the user service requestsmust still be sent tothe Internet for processing and this methodmakes it difficultto demonstrate the technical advantages of edge computing

Wireless Communications and Mobile Computing 11

Table 1 Pull performance test results for different VNFs

Image Types VNFs Image Size (MB) Physical GatewayCost (seconds)

Cloud NetworkBroker Cost(seconds)

System imagesubuntu 112 22095 7918centos 207 24319 8926debian 207 21635 5646

Tool imagesgolang 779 86793 1851flask 458 6222 1633

prometheus 112 2414 4737

Service images

httpd 177 20753 1031nginx 109 13216 4778bind 337 8438 15105squid 215 150031 9973mysql 374 24852 10449

mongodb 503 61347 17116

Application images wordpress 407 32866 13213drupal 452 8139 14405

11095

1593

0

20

40

60

80

100

120

140

160

180

Docker VM

Tim

e (s)

Figure 11 VNF boot time test for the VM and container

In the VNF boot time test we set the image to Ubuntu-16-04-x64 VM and Docker are used to perform the boottime test on the edge gateway We test the time of openingfive VMs and Docker containers VMware is used as the VMhypervisor As shown in Figure 11 the five VMs are openedin 1593 seconds while the Docker containers are openedin 11095 seconds In sum the VMs have a much longerdeployment time than Docker More time can be saved byusingDocker in the edge gateway that requires a large amountof VNFs

To test the average VNF boot time we boot a VM waitfor its activation shut it down and repeat the same steps for14 other VMs The test results are shown in Figure 12 Whenusing VMware the average time to boot the VMs is about3186 seconds However when using Docker these VMs arebooted in only 22 seconds In sum the Docker containerperforms approximately 15 times faster than the VM

The results of the VNF deployment test highlight thesignificant advantages of lightweight containers over tradi-tional virtualizationmethodsThese containers can be started

2219

3186

0

5

10

15

20

25

30

35

Docker VM

Tim

e (s)

Figure 12 Average VNF boot time test for the VM and container

within a few seconds because they adopt a common host OStherebymaking it unnecessary to execute the guestOS in eachcontainer The container also does not need to wait for theOS to start up thereby saving a few seconds Fast start is farfaster than traditional VMs Meanwhile given the design oflightweight VNF images and its highly efficient virtualizationthe application-centric virtualization technology Docker canautomatically obtain container images from the Docker Hubfor flexible deployment and management and can meet theapplication requirements of the edge network for the serviceon-demand of VNF deployment

6 Conclusions

This paper proposed a gateway-based edge computing servicemodel and a GAF task scheduling mechanism that allowsthe edge gateway to schedule more tasks within a short aver-age waiting time The resource estimation and lightweightVNF configuration technologies are used to improve theoperational efficiency of edge computing services to increase

12 Wireless Communications and Mobile Computing

resource utilization to achieve a rapid deployment of VNFand to meet service on-demand requests The combinationof resource estimation task scheduling and lightweight VNFconfiguration design provides an integrated solution that cansatisfy the service on-demand requests for 5G networks Thesimulation results showed that the proposedGAFmechanismoutperforms the other scheduling algorithmsMeanwhile thecomparison revealed that using the lightweight virtualizationtechnology in edge gateways ismore efficient and competitivethan using traditional VMs

In the future we plan to study the multiqueue taskscheduling problem for edge computing the possible coop-eration between edge gateway and cloud servers and the useof SD-WAN technology to achieve a seamless operation of thecloud-edge network

Data Availability

The data used to support the findings of this study areincluded within the article

Disclosure

The funder had no role in the study design data collectionand analysis decision to publish or preparation of thismanuscript

Conflicts of Interest

The authors declare no conflicts of interest regarding thepublication of this paper

Acknowledgments

The work described in this paper was supported in part bythe Ministry of Science and Technology of the Republic ofChina (Grants nos 104-2221-E-008-039 105-2221-E-008-071107-2623-E-008-002 and 107-2636-E-003-001)

References

[1] J Matias J Garay N Toledo J Unzilla and E Jacob ldquoTowardan SDN-enabled NFV architecturerdquo IEEE CommunicationsMagazine vol 53 no 4 pp 187ndash193 2015

[2] Q Duan N Ansari and M Toy ldquoSoftware-defined networkvirtualization An architectural framework for integrating SDNand NFV for service provisioning in future networksrdquo IEEENetwork vol 30 no 5 pp 10ndash16 2016

[3] TSI ldquoNetwork Functions Virtualization (NFV) architecturalframeworkrdquo[Online] Available httpswwwetsiorgdeliveretsi gsNFV001 099002010101 60gs NFV002v010101ppdf

[4] ONF ldquoSDN Architecturerdquo [Online] Available httpwwwopennetworkingorgimagesstoriesdownloadssdn-resourcestechnical-reportsTR SDN ARCH 10 06062014pdf

[5] A Basta A Blenk K Hoffmann H J Morper M Hoffmannand W Kellerer ldquoTowards a Cost Optimal Design for a 5GMobile Core Network Based on SDN and NFVrdquo IEEE Trans-actions on Network and Service Management vol 14 no 4 pp1061ndash1075 2017

[6] F Z Yousaf M Bredel S Schaller and F Schneider ldquoNFV andSDN-Key technology enablers for 5G networksrdquo IEEE Journal

on Selected Areas in Communications vol 35 no 11 pp 2468ndash2478 2017

[7] A C Baktir A Ozgovde and C Ersoy ldquoHow Can EdgeComputing Benefit FromSoftware-DefinedNetworking A Sur-vey Use Cases and Future Directionsrdquo EEE CommunicationsSurveys amp Tutorials vol 19 no 4 pp 2359ndash2391 2017

[8] Guangshun Li Jiping Wang Junhua Wu and Jianrong SongldquoData Processing Delay Optimization in Mobile Edge Comput-ingrdquoWireless Communications andMobile Computing vol 2018Article ID 6897523 9 pages 2018

[9] Zhimin Wang Qinglin Zhao Fangxin Xu Hongning Daiand Yujun Zhang ldquoDetection Performance of Packet Arrivalunder Downclocking for Mobile Edge Computingrdquo WirelessCommunications and Mobile Computing vol 2018 Article ID9641712 7 pages 2018

[10] W Yu F Liang X He et al ldquoA Survey on the Edge Computingfor the Internet of Thingsrdquo IEEE Access vol 6 pp 6900ndash69192018

[11] P Mach and Z Becvar ldquoMobile Edge Computing A Survey onArchitecture and Computation Offloadingrdquo IEEE Communica-tions Surveys amp Tutorials vol 19 no 3 pp 1628ndash1656 2017

[12] N Abbas Y Zhang A Taherkordi and T Skeie ldquoMobile EdgeComputing A Surveyrdquo IEEE Internet of Things Journal vol 5no 1 pp 450ndash465 2018

[13] Guangshun Li Jianrong Song Junhua Wu and Jiping WangldquoMethod of Resource Estimation Based on QoS in Edge Com-putingrdquo Wireless Communications and Mobile Computing vol2018 Article ID 7308913 9 pages 2018

[14] D Happ and A Wolisz ldquoTowards gateway to Cloud offloadingin IoT publishsubscribe systemsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 101ndash106 Valencia Spain May 2017

[15] G Premsankar M Di Francesco and T Taleb ldquoEdge Comput-ing for the Internet of Things A Case Studyrdquo IEEE Internet ofThings Journal vol 5 no 2 pp 1275ndash1284 2018

[16] L Zhao W Sun Y Shi and J Liu ldquoOptimal Placementof Cloudlets for Access Delay Minimization in SDN-BasedInternet of Things Networksrdquo IEEE Internet of Things Journalvol 5 no 2 pp 1334ndash1344 2018

[17] SWang X Zhang Y Zhang LWang J Yang andWWang ldquoASurvey onMobile Edge Networks Convergence of ComputingCaching and Communicationsrdquo IEEE Access vol 5 pp 6757ndash6779 2017

[18] Y Liu J E Fieldsend and G Min ldquoA Framework of FogComputing ArchitectureChallenges and Optimizationrdquo IEEEAccess vol 5 pp 25445ndash25454 2017

[19] ETSI Multi-access Edge Computing [Online] Availablehttpwwwetsiorgtechnologies-clusterstechnologiesmulti-access-edge-computing

[20] OpenFog Consortium [Online] Available httpswwwopen-fogconsortiumorg

[21] Y Mao C You J Zhang K Huang and K B Letaief ldquoA SurveyonMobile Edge ComputingThe Communication PerspectiverdquoIEEE Communications Surveys amp Tutorials vol 19 no 4 pp2322ndash2358 2017

[22] Cisco ldquoThe Internet of Things How the Next Evolution ofthe Internet Is Changing Everything white paperrdquo [Online]Available httpswwwciscocomcdamen usIoT IBSG0411FINALpdf

[23] IDC ldquoFutureScape Worldwide Internet of Things 2017 Pre-dictionsrdquo November 2016 [Online] Available httpswwwidccomgetdocjspcontainerId=US41910716

Wireless Communications and Mobile Computing 3

IoT applications The distributed architecture of edge com-puting nodes satisfies the computing requirements for severalapplications data and services from the cloud data centerto the CPE or the micro data center Based on differentfields of development the edge computing network canbe divided into cloudlets mobile edge computing (MEC)and fog computing [16ndash18] Cloudlets are communicationtools that provide microservices to the network surroundingthe user and virtualize and shrink the computing resourceto deploy the computing resource near the end of theuser The application services and technologies of cloudletsare currently being promoted by Akamai and MicrosoftIntroduced by the European Telecommunications StandardsInstitute (ETSI) MEC is highly applicable in the field ofmobile communication and is operated and managed bycommunication companies [19]This technologymainly aimsto reduce the increasing pressure on network equipment andto help mobile companies create a unique mobile servicemodel Introduced by Cisco and currently promoted by theOpenFog Consortium Alliance fog computing is a conceptof extended cloud computing that focuses on the data pro-cessing function of the local network [20] By emphasizingnear-field data communication fog computing can be usedin different network equipment for personal or businessmanagement and for providing relevant IoT services inspecific areas The OpenFog Consortium aims to promotethe OpenFog computing technology by developing an openarchitecture that can identify certain core technologies andcapabilities such as distributed computing network andstorage which provide intelligent support at the edge of theIoT Keeping in line with the current trends in technologyintegration OpenFog has also cooperated with the ETSIIndustrial Specification Group to establish an applicationframework and development technology for the specificationand interoperability of ICT and for expanding the applicationscope of edge computing

The emergence of the 5G technology has increasedthe requirements for data movement data storage dataprocessing and data analysis for the traditional central-ized cloud computing architecture thereby increasing therequired number of hardware devices to improve the loadcapacity of the core network and the servers Given the largeamount of data generated from IoT the existing cloud servicemodel cannot satisfy the large and dispersed requests ofusers This model has been used for data computing anal-ysis and processing However the demand for low latencyand immediacy has gradually increased thereby renderingthis model slightly insufficient The introduction of edgecomputing helps process data in advance and reduce datatraffic and transmission time this technology also allows thetransfer of computing power to the CPE or terminal nodesto enhance the immediate feedback of the edge operations tothe environment [21] The Cisco Internet Business SolutionsGroup [22] predicts that the IoT will have more than 50billion terminals and devices by 2020 while the InternationalData Corporation estimates that more than 40 of the datain the future need to be analyzed processed and stored atthe edge of the network [23] To effectively respond to thechallenges brought upon by the aforementioned changes the

service model in the edge of the network must be redefinedto optimize the network operation and service processingefficiency to reduce the workload in the cloud and to satisfythe latency requirements of 5G IoT

The rest of this paper is organized as follows Section 2presents the research background and the related worksSection 3 describes the design of the edge computing servicemodel Section 4 presents the experiment results Section 5concludes the paper

2 Background and Related Works

Network operators must deploy various types of networkequipment and software platforms to support increasinglydiverse Internet requirements However the current networkdesign is based on hardware concept various hardwaredevices are often incompatible or unable to communicatewith one another and network maintenance managementhas a high complexity and cost NFV has been recentlyproposed as a solution to the difficulties being faced bynetwork operators Through NFV network and commu-nication operators can decrease the related costs flexiblyand efficiently process the customer network deploymentrequirements reduce the waste of hardware resources andincrease the resource utilization efficiency The current NFVrelated standards are mainly driven by ETSI Given theimpact and benefits brought upon by the virtualization ofnetwork functions operatorsmdashparticularly network serviceproviders and telecommunication operatorsmdashare consideredthe most active participants in the development of NFVstandards

As shown in Figure 3 hypervisor and container are thetwo main frameworks for VNFs [24] Hypervisor is themost common virtualization technology that uses a group ofvirtualization hardware devices to execute the host operatingsystem in a virtual manner and to abstract the physical hard-ware of the host operating systemThe virtual machine (VM)is the operating system deployed on the virtual managementprogram while the Linux container (LXC) is an operating-system-level virtualization technology that mainly packagesthe application service system into a software container thatstores the codes of the application service software requiredoperating system core and library As its most attractivefeature LXC has a simplified software structure Thereforethe image file of the container only includes a small part ofthose operating system components that are smaller than thegeneral VM which is a complete operating system Given itssmall size and lightweight features the container has a muchfaster deployment or movement speed compared with VMAccordingly most industry practitioners prefer the containerover the VM Docker is an open source project based onLXC technology [25] that is operated on the Linux operat-ing system This technology creates an additional softwareabstraction layer for virtualizing the service application inresponse to the development of microservices [26] Asidefrom being a famous product of the Git Hub communityDocker has also been favored by Google and Red Hat and iscurrently the fastest growing NFV virtualization technologyMany studies have also been conducted to examine the

4 Wireless Communications and Mobile Computing

Hardware

Host OS

Hypervisor(VMM)

GuestOS

BinariesLibraries

App

System Virtualization

GuestOS

BinariesLibraries

App

GuestOS

BinariesLibraries

App

Hardware

Host OS

BinariesLibraries

App

BinariesLibraries

App

BinariesLibraries

App

Container Virtualization

namespace

Figure 3 Common approach for the hosting of VMs versus containers

implementation of Docker as an edge computing platform[27 28]

The development of virtualization technologies hasdriven several innovations in the IoT architecture Edgecomputing has also emerged as a new concept that satisfiesthe requirements of IoT for distributed data processing andnumber of devices CPE devices such as the edge gatewayuse virtualization technology to provide local operations thatare particularly suitable for IoT applications with low latencyand immediacy Edge computing can completely disassemblelarge-scale services that are originally handled by cloud datacenters cut these services into small and easy-to-manageparts and distribute them among the edge network nodes ormicro data centers to accelerate their resolution and to reducethe workload in the cloud In this case the service requestsof users are processed on the gateway of the edge networkwhile those requests that exceed the computing capabilitiesor cannot be handled by the edge network are forwarded tothe cloud for processing Edge computing can also be used toreduce the amount of data being sent back to the cloudThesedata are initially processed at the edge node and are thenpreprocessed or filtered before they are transmitted back tothe cloud to ensure that only useful data will be transmittedEdge computing not only reduces network connection delaysand meets the demand of 5GIoT for improving delays butalso promotes the convergence of ICT industry technologiesUsing lightweight virtual technology also helps networkoperators install applications and control software on edgegateways provide localization services and create innovativeservice models Reference [29] proposes a SDN-based edgecomputing architecture called software-defined infrastruc-ture (SDI) which usesOpenFlow andOpenStack to virtualizeservice resources for building smart applications on virtualinfrastructure that can flexibly schedule network resourcesTo introduce the concept of resource sharing [30] proposesa cloud-edge framework where each edge node can use thelocal processing platform to share computing resources toreduce the computational burden on the cloud that processes

the data and to improve the operation of the edge networkTomeet the low latency and fast data processing requirementsof mobile networks [12] examines mobile edge operationsidentifies the applicable situations and reference scenarios fordifferentMECs and proposesmarginalities forMEC offload-ing decisions computing resource allocation and mobilitymanagement The computational needs of the Internet canalso be used as reference for examining edge network servicemanagement and application innovation

Given the resource constraints (eg CPU memory andnetwork-intensive service requests) of the local gatewayhardware the task scheduling mechanism for handling theservice requests of users plays an important role in improvingthe service capabilities of edge computing Several algorithmsfor task scheduling are available in a computer system [31] In-depth studies on several task scheduling algorithms for cloudcomputing have also been conducted [32 33] Reference[34] proposes a task scheduling approach called HealthEdgewhich sets various processing priorities for different tasksbased on the collected human health data and determineswhether a task must be operated on a local device or a remotecloud in order to reduce the total processing time as muchas possible Reference [35] adopts a Markov decision processapproach to solve the stochastic optimization problem in theMEC system In this approach the computation tasks arescheduled based on the queuing state of the task buffer theexecution state of the local processing unit and the stateof the transmission unit Reference [36] proposes a greedybest availability (GBA) mechanism to identify the idealizedtask scheduling policy and to reduce the queuing time ofservices by scheduling the tasks according to their workcompletion time Reference [37] analyzes the performanceof the round-robin (RR) algorithm in the cloud computingenvironment and reveals that RR scheduling fairly allocatescomputing resources among tasks of the same priority byusing the time slicing approach The results show that theRR algorithm demonstrates a better response time and loadbalancing compared with the other algorithms

Wireless Communications and Mobile Computing 5

Virtual Network Functions

VNF

Hardware Resource

Com

putin

g Re

sour

ce

Service RequestVNF

VNF

VNF

R

R

R

R

Figure 4 Network function scheduling problem in CPE

3 Problem Definition

NFV is currently the most valued solution to the prob-lems associated with the operation cost and efficiency ofnetwork services Accordingly many ICT companies havebegun to examine virtualization technology Despite thehigh expectations toward this technology satisfying servicerequests effectively allocating virtual computing resourcesand providing service on-demand application still challengethe deployment of NFV and play key roles in the futuredevelopment of 5GIoT services NFV also allows the estab-lishment of multiple independent heterogeneous virtualnetworks based on common underlying network resourcesthereby enabling service providers to provide customizedservices according to the demands of users Virtual networkembedding is a process of mapping the virtual networkto the underlying network (substrate network) through themapping algorithm and according to the current resourcesituation of the infrastructure providerThis process is anNP-Hard problem and has been investigated in NFV resourceallocation research [38 39] Previous studies [40 41] havealso systematically discussed the problem of virtual networkmapping and provided good references for research

This paper aims to optimize task scheduling and resourceallocation by using the proposed edge computing servicemodel Scheduling tasks in edge computing aremore complexthan that in cloud computing An edge computing operationis typically spread over the device of the client the edgegateways and occasionally a broker of the cloud networkTherefore deciding where to schedule computational tasksremains a key problem in edge computing Given that thegateway-based IoT architecture is currently the mainstreamthe task scheduling mechanism discussed in this paperfocuses on the edge gateway Task scheduling and resourceallocation are the main problems to be solved in this gatewaygiven the limited amount of available computing resourcesLightweight NFV technology plays an indispensable role inthe rapid deployment of the edge gateway The relationshipbetween task scheduling and resource allocation for hostservice requests is plotted in Figure 4 As each service requestarrives at a different time the required VNF service andprocessing time also differ The edge gateway can adjust thetask scheduling and resource configuration according to the

differences among the service requests As the basic idea forthe service on-demand model only one VNF is dedicated toa single service request This service model problem can beanalyzed by using queuing theory which regards the requestand react processing as a waiting-line system the input of awaiting-line system as the service request the service counteras the gateway scheduling function and the output as therequested VNF resource Although many queuing modelsmay be used in operations management [42] this paperprimarily focuses on the task scheduling approach that allowsa certain edge gateway to processmultipleVNFs fromaqueueone after another Given its limited capacity the edge gatewaycan only schedule a limited number of service requests whilethe subsequent service requests need to be forwarded to thecloud for processing With these considerations the researchproblem is formulated as follows Assume that the servicerequest is Poissonrsquos ratio within one edge gateway and thatthe service time is an exponential distribution The servicerequest set can be denoted as 119877 = 1198771 1198772 1198773 119877119899 wheren denotes the number of service requests in the systemEquations (1) and (2) are used to determine the probabilityof n service requests in the system In these equations Ndenotes the maximum number of service requests that canbe scheduled in the gateway 120582 is the average number ofincoming user requests in one unit of time 120583 is the serviceefficiency (the ability of the service counter) 120588 = 120582120583 is theratio that the request can be met within one unit of time P0denotes the initial condition L denotes the total number ofservice requests arriving in the systemwithin a planned timeand 119871119902 denotes the total number of server requests queued inthe system L and 119871119902 can be computed by using (3) and (4)respectively Equation (5) computes the waiting and servicetime of a service in the system while (6) computes119882119902 or theaverage waiting time for a user request

1198750 =

1 minus 1205881 minus 120588119873+1 120588 = 11119873 + 1 120588 = 1

(1)

119875119899 =

1205881198991198750 = 1205881198991 minus 1205881 minus 120588119873+1 120588 = 1

1198750 =1119873 + 1 120588 = 1

(2)

6 Wireless Communications and Mobile Computing

Binaries Libraries

Host OS

Hardware Resource

VNF

Docker Engine

ResourceEstimation Scheduler VNF

Configuration

VN

F Manager

Container images

R

R

R

R

R

Edge Computing Node VNFs Resource PoolService Request

VNF

VNF

PublicRegistry

PrivateRegistry

SDN Controller

Figure 5 Service on-demand edge computing model

119871 =

1205881 minus 120588 minus

(119873 + 1) 120588119873+1

1 minus 120588119873+1 120588 = 11198732 120588 = 1

(3)

119871119902 = L minus (1 minus 1198750) (4)

119882 = 119871120582 (1 minus 119875119899)

(5)

119882119902 =119871119902120582

(6)

The results reveal that the waiting time for the userrequest can be reduced in two ways namely by reducing thenumber of service requests queued in the edge gateway andby increasing the processing speed of task scheduling in theedge computing gateway A large number of tasks must becompleted within a short period to achieve an efficient edgecomputing

4 Design of the Edge ComputingService Model

How to construct an elastic and cost-effective edge com-puting service model improve management efficiency meetuser service requests achieve centralized management anddevelop flexible configuration service models are the currentresearch trends in the development of a 5G SDNNFV net-work This paper proposes a gateway-based edge computingservicemodel (Figure 5) to improve the operational efficiencyof the edge computing node to accelerate the processingof user service requests and to increase the utilizationefficiency of a limited number of computing resources In thismodel when different user requests enter the edge gatewaythis gateway determines whether the requested services canbe processed or not If the edge gateway itself lacks thecomputing capacity or resources then the controller forwardsthe service request to the cloud to reduce the data processinglatency

The proposed edge computing service model can bedivided into resource estimation scheduler and lightweightVNF configuration Figure 6 presents the flowchart of theoperations in the proposed edge computing service modeland these three parts are further described in the followingsections

Resource estimation checkswhether the edge gateway hasa sufficient amount of computing resources to provide edgecomputing services For the user request set R the resourceallocation must abide by the following rules which are alsoused by the edge gateway

(1) For a single service request Ri any resource of Vi(eg CPU memory and disk) is less than the totalresources of P forall119877119862119894ltPC forall119877119872119894ltPM and forall119877119863119894ltPD119877119862119894 119877119872119894 and 119877119863119894 denote the CPU memory and diskspace requests sent to Ri respectively

(2) The sum of computing resources that are allocated tothe VNF in the gateway is less than the total numberof resources of the physical machine P 119881119862119894 lt PC 119881119872119894lt PM 119881119863119894 lt PD 119881119862119894 119881119872119894 and 119881119863119894 denote the sum ofCPU memory and disk space resources allocated toVi respectively

As shown in Figure 6 after receiving the user servicerequest the edge gateway checks whether a sufficient amountof computing resources is available to satisfy such requestThe following situations may be encountered in this case

(1) If the edge gateway has a sufficient amount ofresources then the user service request is processedthrough the scheduler and queued in the systembased on the result of the task scheduling algorithm

(2) If the edge gateway does not have a sufficient amountof available resources then the user service request isdirectly transferred to the cloud instead

Task scheduling aims to increase the operational effi-ciency of the edge gateway Given that one service requestdiffers from another task scheduling examines how the

Wireless Communications and Mobile Computing 7

Forward Ri to Cloud DC

Start

Download from private storage

RCilt(PCiminusVCi)

RMilt(PMiminusVMi)

RDilt(PDiminusVDi)

No

Yes

Yes

Yes

Resource Estimation for Ri

VNF allocation

Download from public storage

End

No

No

No

Yes

Task Scheduling Docker Image is exist

Riconfigured to

the scheduler No

PCi = (PCiminusRCi)PMi = (PMiminusRMi)PDi =(PDiminusRDi)

Service on-demand VNF Matching

Yes

Figure 6 Edge computing service model flowchart

edge gateway can meet the requirements of different servicerequests and accomplish task scheduling in the system As itsprimary purpose scheduling attempts to reduce the amountof time spent on dealing with the most demanding servicerequirements asmuch as possible For this purpose this paperconstructs the Greedy Available Fit (GAF) task schedulingmechanism to enhance the operational efficiency of edgecomputing services

Assume that each service request Ri configures a VNFvirtualization service resource Vi ti denotes the processingtime of the ith service request di denotes its deadline andj denotes its completion time in the system The Ri in thesystem must be completed at schedule and before the basictime limit Otherwise this service request must be forwardeddirectly to the cloud As each service request arrives at adifferent time the time required for the operation processingand the deadline time also differ In this case deadline isselected as a priority parameter for the task scheduling Thedeadline can be used to define the processing priority forservices that is the deadline for those services that requirereal-time processing may be set according to their processingand precedence requirements on different VNFs This paperaims to insert as many tasks as possible into the schedulebefore completing the largest task Ri of the deadline Toaccomplish this objective a task with the maximum deadlineis selected as the baseline and the remaining tasks aresequentially inserted into the queue based on their processing

time119873[119894 119895] denotes the maximumnumber of j tasks selectedfrom the front i tasks and can be formulated as

119873[119894 119895] = max

119873[119894 minus 1 119895]

119873 [119894 minus 1 119895 minus 119905119894] + 1 119894119891 119895 le 119889119894119873[119894 minus 1 119895 minus 119905119894] 119894119891 119895 gt 119889119894

(7)

Equation (8) shows the initial condition of119873[119894 119895]

119873[1 119895] =

minusinfin 119894119891 119895 = 11990511 119894119891 119895 = 1199051 le 11988910 119894119891 119895 = 1199051 gt 1198891

(8)

Assume that there is no time gap betweenRi and119877119894+1 thatR1 starts from time 0 and that N = 1 If t1 le d1 then N = 0because the deadline is exceeded

Thewhole decision process is summarized inAlgorithm 1which starts with an empty schedule and inserts the availabletasks into this schedule in three different cases In case 1 if thecurrent time step j exceeds the deadline of Ri then N[i j] =N[iminus1 jminusti] In case (2) if there is not enough time to finishrequest Ri by the current time j then N[i j] = N[iminus1 j] Incase (3) if time j does not exceed the deadline of Ri and thereis enough time to finish request Ri by time j then N[i j] =max(N[iminus1 j] N[iminus1 jminusti ] + 1)

8 Wireless Communications and Mobile Computing

Input 119877 = 1198771 1198772 119877119899 ti di jOutput119873[119894 119895](1) Start(2) Set the largest task di as the baseline of the scheduler(3) for 119894 larr997888 1 to n initialize the maximum N at time 0 to be zero for each service

request in queue(4) 119873[119894 0] = 0(5) for 119895 larr997888 1 to 119889119899 determine the maximum N obtained by R1 at each time step(6) if ((j == t1) and (j lt= d1))(7) 119873[1 119895] = 1(8) else(9) 119873[1 119895] =119873[1 119895 minus 1](10) for 119894 larr997888 2 to n(11) for 119895 larr997888 1 to 119889119894(12) if (j gt di) case (1) the current time step j already exceeds 119877119894rsquos deadline(13) then119873[119894 119895] = 119873[119894 minus 1 119895 minus 119905119894](14) else if (j lt ti) case (2) there is not enough time to finish request Ri by the

current time j(15) then119873[119894 119895] = 119873[119894 minus 1 119895](16) else case (3) time j does not exceed 119877119894rsquos deadline and there is enough time

to finish request 119877119894 by time j(17) then119873[119894 119895] = max(119873[119894 minus 1 119895]119873[119894 minus 1 119895 minus 119905119894] + 1)(18) Schedule Completed(19) End

Algorithm 1 Greedy Available Fit algorithm

Virtualization technology can be applied on the CPE inmany ways such as by using the VM container or VM inte-gration of container However a traditional VM consumesmany system resources and cannot meet the requirementsfor light weight and service on-demand deployment In thiscase this research adopts container technology instead Acontainer handles only one service request at a time and stopsits operation after completing the delivery of a service TheVNF manager on the gateway is responsible for configuringand allocating each VNF The VNF template image that isrequired by different services is placed in the VNF resourcepool (with Docker Hub as the default other public or privateregistrations can also be specified) The edge gateway notonly controls the amount of gateway resources used byeach container but also manages several resources such asCPU and RAM to ensure that the container can obtain therequired resources without affecting the performance of theother executing containers on the edge gateway

Linux containers adopt a hierarchical structure to rapidlydeploy VNFs and to manage NFV flexible scheduling Theunderlying structure uses the file archiving mechanism ofDocker namely the advanced multilayered unification filesystem to incorporatemany different VNF imagesThe layersare stacked up andwhen someVNF service functions need tobe accessed the container retrieves the VNF images throughthe VNFmanager and uses these images directly To integratethe open resources of Docker Hub and to expand the privatewarehouse of the VNF service manager the system can storethe image files obtained from the public warehouse in aprivate warehouse for future use The judgment mechanismof the VNF service manager is described as follows

(1) If the image requested by the host is already in thelocal container then the VNF image can be taken outdirectly by the Docker daemon of the VNF servicemanager No complex configurations are required

(2) If the VNF image to be used by the host is not in thelocal container then the VNF manager connects tothe common warehouse Docker Hub on the Internetto find and store the desired service image file ina private warehouse via pushpull Afterward theacquired VNF image is controlled by the Dockerdaemon

By using the OS kernel for resource isolation the con-tainer technology does not need to rely on a virtual softwarelayer and does not require the VM to install a guest OSTherefore the capacity of the image file is much larger thanthat of the virtual machine The image file is small and canbe rapidly deployed through network transmission therebysaving network resources By using Linux container technol-ogy we can configure various VNF images to be executed onthe same edge gateway thereby replacing the virtual machineand achieving the goal of lightweight virtualization

5 Experiment

An experiment is conducted to evaluate the proposed taskscheduling algorithmand to test its performance in deployinglightweight VNF The experiment environment is shown inFigure 7 A simulation is initially performed to evaluate thetask scheduling performance of the algorithm by changingthe number of service requirements The service request is

Wireless Communications and Mobile Computing 9

Edge Network

Internet

Edge Gateway

DataCenter

Cloud Broker

Docker HubDocker Hub

Figure 7 The experiment environment

characterized as a Poisson process There we showed that theaverage number of service requests in the system is given by120582120583 where 120582 is the average arrival rate and 120583 is the averageservice rate The input parameter 120582120583 is 045 First comefirst service (FCFS) priority task scheduling [34] RR [37]and GBA [36] are then compared with the proposed GAFmechanism

The effect of the service requests on the average waitingtime average response time and task scheduling of differentscheduling mechanisms is evaluated in the simulation FCFSis based on the service requests that enter the edge gatewayqueuing system and schedules the tasks according to thesequence of these requests Priority task scheduling is anunfair scheduling algorithm where the service requests aresorted according to their priority and where the high-prioritytasks are performedfirstThose tasks having the same priorityare sorted by using the FIFO scheduling mechanism TheGBA sorts the tasks based on the completion time of theservice requests in queuing systemTheRR algorithm is basedon the conventional RR scheduling performed in the processscheduling RR scheduling fairly allocates the computingresources to those tasks with the same priority by using thetime slicing approach (time quantum)

Figure 8 shows the experiment result for average waitingtime where the x-axis refers to the number of service requestsand the y main axis refers to the average waiting time thatis the time a user spends to complete the service schedulingand to determine the resource allocation after receiving aservice request and a successful reply If the demand revertsto an errortimeout or if the edge gateway does not haveenough computing resources to support such demand thenthe sample is excluded from the calculation of the averagewaiting time As can be seen in the experimental resultthe RR algorithm obtains the longest average waiting timebecause of its application of time slice rules in order for eachtask to be processed within a fixed amount of time The timequantum affects the overall performance of the operation

0100200300400500600

10 20 30 40 50 60 70 80 90 100

Aver

age W

aitin

g Ti

me (

ms)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 8 The experiment result for average waiting time

Having a very long time quantum leads to a very long waitingtime while having a very short time quantum results intask schedule conversion a poor execution efficiency and anextended waiting time For FCFS given that the time spentdiffers across each task the waiting time for the next taskmust be determined based on the schedule of the previoustask Therefore if the last task schedule is too long thesystem cannot quickly process the subsequent task schedulesthereby affecting the overall task scheduling efficiency Giventhat its average waiting time is relatively shorter than thatof the FCFS mechanism priority task scheduling can satisfythe task requirements and flexibly perform the schedulingHowever prioritizing a high-priority program will delay allof the low-priority requests thereby creating an indefinitesituation in the low-priority program and extending thewaiting time GBA is based on the earliest completion time ofthe taskWhen the service demand is low the average waitingtime of the GBA algorithm is similar to that of priority taskscheduling However when the service demand increases theaverage waiting time of the GBAmechanism becomes longerthan that of priority task scheduling Compared with FCFS

10 Wireless Communications and Mobile Computing

050

100150200250300350

10 20 30 40 50 60 70 80 90 100

Ave

rage

Res

pons

e Tim

e (m

s)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 9 The experiment result for average response time

01020304050607080

10 20 30 40 50 60 70 80 90 100

Num

ber o

f Sch

edul

ed T

asks

Number of Service RequestsGAFFCFSPriority

GBARR

Figure 10 The experiment result for task scheduling efficiency

priority task scheduling RR and GBA the proposed GAFalgorithm selects the largest deadline task as the baseline andthen sequentially inserts the remaining tasks into the queuebased on their processing timeTherefore the GAF algorithmobtains the shortest average waiting time

Figure 9 shows the results for average response time Asexpected the RR algorithm outperforms all other approachesin terms of average response time because this algorithm isespecially designed for time sharing Using a fair strategy toallocate VNF resources ensures that each service request hasa fixed time If the service processing time exceeds the timequantum allocated by the system then the sample is excludedfrom the calculation of the average response time Given thatFCFS is processed in sequence the average response timewill be affected by the time spent by the previous 119877119894minus1 taskin the scheduler For example if the previous service requesthas a very long task schedule then the average responsetime increases Given that the proposed GAF algorithmsequentially inserts tasks into the queue its average responsetime is slightly longer than that of FCFS and GBA Althoughthe priority algorithm can sort the tasks according to thepriorities of the service requests the low-priority tasks areeasily delayed thereby increasing the average response time

Figure 10 shows the task scheduling efficiency of thecompared algorithms If the service request number is lessthan 60 then each algorithm can complete the task schedul-ing However when this number is exceeded the number

of tasks that can be scheduled varies across each algorithmGAF outperforms all the other algorithms as it ranks thelargest deadline task at the end and then sequentially sortsthe remaining tasks according to their processing time In thisway this algorithm ensures that most tasks will be completedwithin a minimum processing time Those service requeststhat cannot be scheduled will bypass the gateway and will bedirectly forwarded to the cloud data center When the servicerequest number is 80 GAF can complete 68 of the taskscheduling Meanwhile GBA performs the scheduling basedon the earliest task completion time Although GAF andGBAcan schedule the same number of tasks in the experimentalenvironment the latter has a longer average waiting timecompared with the former GAF also outperforms FCFCand priority task scheduling by 96 and 62 in terms oftask scheduling efficiency respectively because this approachranks the largest deadline task at the end and then sorts theremaining tasks according to their required processing timeGAF also outperforms theRRmechanismby 70because thelatter applies the time-sharing rule to minimize the numberof scheduled tasks Based on these results the GAF algorithmcan schedule more tasks in the edge gateway within a shortaverage waiting time and improve the operational efficiencyof edge computing services

In addition to task scheduling the VNF configuration isanother main factor that affects the operational efficiency ofthe edge computing service model This paper utilizes theDocker technology to achieve an agile deployment of VNFsSuch technology can provide a centralized management andflexible configuration of VNF services Depending on theservice requests of each user the deployment of differentVNFs can meet the application requirements of serviceson-demand The VNF image of Docker can be mainlydivided into four categories namely system images (egUbuntu CentOS and Debian) tool images (eg GolangFlask and Tomcat) service images (eg MySQL MongoDBand RabbitMQ) and application images (eg WordPressand DRUPAL) To test the practical deployment efficiency oflightweight VNF for the edge gateway a pull performancetest is performed for each type of Docker VNF image Theedge gateway is simulated by using a standalone PC with 34GHz dual-core CPU 16GBmemory and 5400 rpmHD500GThe experimental network bandwidth is set to 100 Mbps Tocompare the pull performance with the local edge gateway atest VM in the Google Cloud platform is used to simulate thecloud network broker Table 1 presents the result of the pullperformance test

As shown in Table 1 each VNF image size has differentphysical gateway and cloud network broker costs due to thelimitations in the network bandwidth A larger VNF imagetakes a longer time to be downloaded from Docker Hub Asexpected the cloud network broker has a better computingarchitecture than the local edge gateway because the locationof the edge gateway affects the performance of the VNFdeployment Although the edge computing architecture ofthe cloud network broker deploys VNF more efficiently thanthe local gateway the user service requestsmust still be sent tothe Internet for processing and this methodmakes it difficultto demonstrate the technical advantages of edge computing

Wireless Communications and Mobile Computing 11

Table 1 Pull performance test results for different VNFs

Image Types VNFs Image Size (MB) Physical GatewayCost (seconds)

Cloud NetworkBroker Cost(seconds)

System imagesubuntu 112 22095 7918centos 207 24319 8926debian 207 21635 5646

Tool imagesgolang 779 86793 1851flask 458 6222 1633

prometheus 112 2414 4737

Service images

httpd 177 20753 1031nginx 109 13216 4778bind 337 8438 15105squid 215 150031 9973mysql 374 24852 10449

mongodb 503 61347 17116

Application images wordpress 407 32866 13213drupal 452 8139 14405

11095

1593

0

20

40

60

80

100

120

140

160

180

Docker VM

Tim

e (s)

Figure 11 VNF boot time test for the VM and container

In the VNF boot time test we set the image to Ubuntu-16-04-x64 VM and Docker are used to perform the boottime test on the edge gateway We test the time of openingfive VMs and Docker containers VMware is used as the VMhypervisor As shown in Figure 11 the five VMs are openedin 1593 seconds while the Docker containers are openedin 11095 seconds In sum the VMs have a much longerdeployment time than Docker More time can be saved byusingDocker in the edge gateway that requires a large amountof VNFs

To test the average VNF boot time we boot a VM waitfor its activation shut it down and repeat the same steps for14 other VMs The test results are shown in Figure 12 Whenusing VMware the average time to boot the VMs is about3186 seconds However when using Docker these VMs arebooted in only 22 seconds In sum the Docker containerperforms approximately 15 times faster than the VM

The results of the VNF deployment test highlight thesignificant advantages of lightweight containers over tradi-tional virtualizationmethodsThese containers can be started

2219

3186

0

5

10

15

20

25

30

35

Docker VM

Tim

e (s)

Figure 12 Average VNF boot time test for the VM and container

within a few seconds because they adopt a common host OStherebymaking it unnecessary to execute the guestOS in eachcontainer The container also does not need to wait for theOS to start up thereby saving a few seconds Fast start is farfaster than traditional VMs Meanwhile given the design oflightweight VNF images and its highly efficient virtualizationthe application-centric virtualization technology Docker canautomatically obtain container images from the Docker Hubfor flexible deployment and management and can meet theapplication requirements of the edge network for the serviceon-demand of VNF deployment

6 Conclusions

This paper proposed a gateway-based edge computing servicemodel and a GAF task scheduling mechanism that allowsthe edge gateway to schedule more tasks within a short aver-age waiting time The resource estimation and lightweightVNF configuration technologies are used to improve theoperational efficiency of edge computing services to increase

12 Wireless Communications and Mobile Computing

resource utilization to achieve a rapid deployment of VNFand to meet service on-demand requests The combinationof resource estimation task scheduling and lightweight VNFconfiguration design provides an integrated solution that cansatisfy the service on-demand requests for 5G networks Thesimulation results showed that the proposedGAFmechanismoutperforms the other scheduling algorithmsMeanwhile thecomparison revealed that using the lightweight virtualizationtechnology in edge gateways ismore efficient and competitivethan using traditional VMs

In the future we plan to study the multiqueue taskscheduling problem for edge computing the possible coop-eration between edge gateway and cloud servers and the useof SD-WAN technology to achieve a seamless operation of thecloud-edge network

Data Availability

The data used to support the findings of this study areincluded within the article

Disclosure

The funder had no role in the study design data collectionand analysis decision to publish or preparation of thismanuscript

Conflicts of Interest

The authors declare no conflicts of interest regarding thepublication of this paper

Acknowledgments

The work described in this paper was supported in part bythe Ministry of Science and Technology of the Republic ofChina (Grants nos 104-2221-E-008-039 105-2221-E-008-071107-2623-E-008-002 and 107-2636-E-003-001)

References

[1] J Matias J Garay N Toledo J Unzilla and E Jacob ldquoTowardan SDN-enabled NFV architecturerdquo IEEE CommunicationsMagazine vol 53 no 4 pp 187ndash193 2015

[2] Q Duan N Ansari and M Toy ldquoSoftware-defined networkvirtualization An architectural framework for integrating SDNand NFV for service provisioning in future networksrdquo IEEENetwork vol 30 no 5 pp 10ndash16 2016

[3] TSI ldquoNetwork Functions Virtualization (NFV) architecturalframeworkrdquo[Online] Available httpswwwetsiorgdeliveretsi gsNFV001 099002010101 60gs NFV002v010101ppdf

[4] ONF ldquoSDN Architecturerdquo [Online] Available httpwwwopennetworkingorgimagesstoriesdownloadssdn-resourcestechnical-reportsTR SDN ARCH 10 06062014pdf

[5] A Basta A Blenk K Hoffmann H J Morper M Hoffmannand W Kellerer ldquoTowards a Cost Optimal Design for a 5GMobile Core Network Based on SDN and NFVrdquo IEEE Trans-actions on Network and Service Management vol 14 no 4 pp1061ndash1075 2017

[6] F Z Yousaf M Bredel S Schaller and F Schneider ldquoNFV andSDN-Key technology enablers for 5G networksrdquo IEEE Journal

on Selected Areas in Communications vol 35 no 11 pp 2468ndash2478 2017

[7] A C Baktir A Ozgovde and C Ersoy ldquoHow Can EdgeComputing Benefit FromSoftware-DefinedNetworking A Sur-vey Use Cases and Future Directionsrdquo EEE CommunicationsSurveys amp Tutorials vol 19 no 4 pp 2359ndash2391 2017

[8] Guangshun Li Jiping Wang Junhua Wu and Jianrong SongldquoData Processing Delay Optimization in Mobile Edge Comput-ingrdquoWireless Communications andMobile Computing vol 2018Article ID 6897523 9 pages 2018

[9] Zhimin Wang Qinglin Zhao Fangxin Xu Hongning Daiand Yujun Zhang ldquoDetection Performance of Packet Arrivalunder Downclocking for Mobile Edge Computingrdquo WirelessCommunications and Mobile Computing vol 2018 Article ID9641712 7 pages 2018

[10] W Yu F Liang X He et al ldquoA Survey on the Edge Computingfor the Internet of Thingsrdquo IEEE Access vol 6 pp 6900ndash69192018

[11] P Mach and Z Becvar ldquoMobile Edge Computing A Survey onArchitecture and Computation Offloadingrdquo IEEE Communica-tions Surveys amp Tutorials vol 19 no 3 pp 1628ndash1656 2017

[12] N Abbas Y Zhang A Taherkordi and T Skeie ldquoMobile EdgeComputing A Surveyrdquo IEEE Internet of Things Journal vol 5no 1 pp 450ndash465 2018

[13] Guangshun Li Jianrong Song Junhua Wu and Jiping WangldquoMethod of Resource Estimation Based on QoS in Edge Com-putingrdquo Wireless Communications and Mobile Computing vol2018 Article ID 7308913 9 pages 2018

[14] D Happ and A Wolisz ldquoTowards gateway to Cloud offloadingin IoT publishsubscribe systemsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 101ndash106 Valencia Spain May 2017

[15] G Premsankar M Di Francesco and T Taleb ldquoEdge Comput-ing for the Internet of Things A Case Studyrdquo IEEE Internet ofThings Journal vol 5 no 2 pp 1275ndash1284 2018

[16] L Zhao W Sun Y Shi and J Liu ldquoOptimal Placementof Cloudlets for Access Delay Minimization in SDN-BasedInternet of Things Networksrdquo IEEE Internet of Things Journalvol 5 no 2 pp 1334ndash1344 2018

[17] SWang X Zhang Y Zhang LWang J Yang andWWang ldquoASurvey onMobile Edge Networks Convergence of ComputingCaching and Communicationsrdquo IEEE Access vol 5 pp 6757ndash6779 2017

[18] Y Liu J E Fieldsend and G Min ldquoA Framework of FogComputing ArchitectureChallenges and Optimizationrdquo IEEEAccess vol 5 pp 25445ndash25454 2017

[19] ETSI Multi-access Edge Computing [Online] Availablehttpwwwetsiorgtechnologies-clusterstechnologiesmulti-access-edge-computing

[20] OpenFog Consortium [Online] Available httpswwwopen-fogconsortiumorg

[21] Y Mao C You J Zhang K Huang and K B Letaief ldquoA SurveyonMobile Edge ComputingThe Communication PerspectiverdquoIEEE Communications Surveys amp Tutorials vol 19 no 4 pp2322ndash2358 2017

[22] Cisco ldquoThe Internet of Things How the Next Evolution ofthe Internet Is Changing Everything white paperrdquo [Online]Available httpswwwciscocomcdamen usIoT IBSG0411FINALpdf

[23] IDC ldquoFutureScape Worldwide Internet of Things 2017 Pre-dictionsrdquo November 2016 [Online] Available httpswwwidccomgetdocjspcontainerId=US41910716

4 Wireless Communications and Mobile Computing

Hardware

Host OS

Hypervisor(VMM)

GuestOS

BinariesLibraries

App

System Virtualization

GuestOS

BinariesLibraries

App

GuestOS

BinariesLibraries

App

Hardware

Host OS

BinariesLibraries

App

BinariesLibraries

App

BinariesLibraries

App

Container Virtualization

namespace

Figure 3 Common approach for the hosting of VMs versus containers

implementation of Docker as an edge computing platform[27 28]

The development of virtualization technologies hasdriven several innovations in the IoT architecture Edgecomputing has also emerged as a new concept that satisfiesthe requirements of IoT for distributed data processing andnumber of devices CPE devices such as the edge gatewayuse virtualization technology to provide local operations thatare particularly suitable for IoT applications with low latencyand immediacy Edge computing can completely disassemblelarge-scale services that are originally handled by cloud datacenters cut these services into small and easy-to-manageparts and distribute them among the edge network nodes ormicro data centers to accelerate their resolution and to reducethe workload in the cloud In this case the service requestsof users are processed on the gateway of the edge networkwhile those requests that exceed the computing capabilitiesor cannot be handled by the edge network are forwarded tothe cloud for processing Edge computing can also be used toreduce the amount of data being sent back to the cloudThesedata are initially processed at the edge node and are thenpreprocessed or filtered before they are transmitted back tothe cloud to ensure that only useful data will be transmittedEdge computing not only reduces network connection delaysand meets the demand of 5GIoT for improving delays butalso promotes the convergence of ICT industry technologiesUsing lightweight virtual technology also helps networkoperators install applications and control software on edgegateways provide localization services and create innovativeservice models Reference [29] proposes a SDN-based edgecomputing architecture called software-defined infrastruc-ture (SDI) which usesOpenFlow andOpenStack to virtualizeservice resources for building smart applications on virtualinfrastructure that can flexibly schedule network resourcesTo introduce the concept of resource sharing [30] proposesa cloud-edge framework where each edge node can use thelocal processing platform to share computing resources toreduce the computational burden on the cloud that processes

the data and to improve the operation of the edge networkTomeet the low latency and fast data processing requirementsof mobile networks [12] examines mobile edge operationsidentifies the applicable situations and reference scenarios fordifferentMECs and proposesmarginalities forMEC offload-ing decisions computing resource allocation and mobilitymanagement The computational needs of the Internet canalso be used as reference for examining edge network servicemanagement and application innovation

Given the resource constraints (eg CPU memory andnetwork-intensive service requests) of the local gatewayhardware the task scheduling mechanism for handling theservice requests of users plays an important role in improvingthe service capabilities of edge computing Several algorithmsfor task scheduling are available in a computer system [31] In-depth studies on several task scheduling algorithms for cloudcomputing have also been conducted [32 33] Reference[34] proposes a task scheduling approach called HealthEdgewhich sets various processing priorities for different tasksbased on the collected human health data and determineswhether a task must be operated on a local device or a remotecloud in order to reduce the total processing time as muchas possible Reference [35] adopts a Markov decision processapproach to solve the stochastic optimization problem in theMEC system In this approach the computation tasks arescheduled based on the queuing state of the task buffer theexecution state of the local processing unit and the stateof the transmission unit Reference [36] proposes a greedybest availability (GBA) mechanism to identify the idealizedtask scheduling policy and to reduce the queuing time ofservices by scheduling the tasks according to their workcompletion time Reference [37] analyzes the performanceof the round-robin (RR) algorithm in the cloud computingenvironment and reveals that RR scheduling fairly allocatescomputing resources among tasks of the same priority byusing the time slicing approach The results show that theRR algorithm demonstrates a better response time and loadbalancing compared with the other algorithms

Wireless Communications and Mobile Computing 5

Virtual Network Functions

VNF

Hardware Resource

Com

putin

g Re

sour

ce

Service RequestVNF

VNF

VNF

R

R

R

R

Figure 4 Network function scheduling problem in CPE

3 Problem Definition

NFV is currently the most valued solution to the prob-lems associated with the operation cost and efficiency ofnetwork services Accordingly many ICT companies havebegun to examine virtualization technology Despite thehigh expectations toward this technology satisfying servicerequests effectively allocating virtual computing resourcesand providing service on-demand application still challengethe deployment of NFV and play key roles in the futuredevelopment of 5GIoT services NFV also allows the estab-lishment of multiple independent heterogeneous virtualnetworks based on common underlying network resourcesthereby enabling service providers to provide customizedservices according to the demands of users Virtual networkembedding is a process of mapping the virtual networkto the underlying network (substrate network) through themapping algorithm and according to the current resourcesituation of the infrastructure providerThis process is anNP-Hard problem and has been investigated in NFV resourceallocation research [38 39] Previous studies [40 41] havealso systematically discussed the problem of virtual networkmapping and provided good references for research

This paper aims to optimize task scheduling and resourceallocation by using the proposed edge computing servicemodel Scheduling tasks in edge computing aremore complexthan that in cloud computing An edge computing operationis typically spread over the device of the client the edgegateways and occasionally a broker of the cloud networkTherefore deciding where to schedule computational tasksremains a key problem in edge computing Given that thegateway-based IoT architecture is currently the mainstreamthe task scheduling mechanism discussed in this paperfocuses on the edge gateway Task scheduling and resourceallocation are the main problems to be solved in this gatewaygiven the limited amount of available computing resourcesLightweight NFV technology plays an indispensable role inthe rapid deployment of the edge gateway The relationshipbetween task scheduling and resource allocation for hostservice requests is plotted in Figure 4 As each service requestarrives at a different time the required VNF service andprocessing time also differ The edge gateway can adjust thetask scheduling and resource configuration according to the

differences among the service requests As the basic idea forthe service on-demand model only one VNF is dedicated toa single service request This service model problem can beanalyzed by using queuing theory which regards the requestand react processing as a waiting-line system the input of awaiting-line system as the service request the service counteras the gateway scheduling function and the output as therequested VNF resource Although many queuing modelsmay be used in operations management [42] this paperprimarily focuses on the task scheduling approach that allowsa certain edge gateway to processmultipleVNFs fromaqueueone after another Given its limited capacity the edge gatewaycan only schedule a limited number of service requests whilethe subsequent service requests need to be forwarded to thecloud for processing With these considerations the researchproblem is formulated as follows Assume that the servicerequest is Poissonrsquos ratio within one edge gateway and thatthe service time is an exponential distribution The servicerequest set can be denoted as 119877 = 1198771 1198772 1198773 119877119899 wheren denotes the number of service requests in the systemEquations (1) and (2) are used to determine the probabilityof n service requests in the system In these equations Ndenotes the maximum number of service requests that canbe scheduled in the gateway 120582 is the average number ofincoming user requests in one unit of time 120583 is the serviceefficiency (the ability of the service counter) 120588 = 120582120583 is theratio that the request can be met within one unit of time P0denotes the initial condition L denotes the total number ofservice requests arriving in the systemwithin a planned timeand 119871119902 denotes the total number of server requests queued inthe system L and 119871119902 can be computed by using (3) and (4)respectively Equation (5) computes the waiting and servicetime of a service in the system while (6) computes119882119902 or theaverage waiting time for a user request

1198750 =

1 minus 1205881 minus 120588119873+1 120588 = 11119873 + 1 120588 = 1

(1)

119875119899 =

1205881198991198750 = 1205881198991 minus 1205881 minus 120588119873+1 120588 = 1

1198750 =1119873 + 1 120588 = 1

(2)

6 Wireless Communications and Mobile Computing

Binaries Libraries

Host OS

Hardware Resource

VNF

Docker Engine

ResourceEstimation Scheduler VNF

Configuration

VN

F Manager

Container images

R

R

R

R

R

Edge Computing Node VNFs Resource PoolService Request

VNF

VNF

PublicRegistry

PrivateRegistry

SDN Controller

Figure 5 Service on-demand edge computing model

119871 =

1205881 minus 120588 minus

(119873 + 1) 120588119873+1

1 minus 120588119873+1 120588 = 11198732 120588 = 1

(3)

119871119902 = L minus (1 minus 1198750) (4)

119882 = 119871120582 (1 minus 119875119899)

(5)

119882119902 =119871119902120582

(6)

The results reveal that the waiting time for the userrequest can be reduced in two ways namely by reducing thenumber of service requests queued in the edge gateway andby increasing the processing speed of task scheduling in theedge computing gateway A large number of tasks must becompleted within a short period to achieve an efficient edgecomputing

4 Design of the Edge ComputingService Model

How to construct an elastic and cost-effective edge com-puting service model improve management efficiency meetuser service requests achieve centralized management anddevelop flexible configuration service models are the currentresearch trends in the development of a 5G SDNNFV net-work This paper proposes a gateway-based edge computingservicemodel (Figure 5) to improve the operational efficiencyof the edge computing node to accelerate the processingof user service requests and to increase the utilizationefficiency of a limited number of computing resources In thismodel when different user requests enter the edge gatewaythis gateway determines whether the requested services canbe processed or not If the edge gateway itself lacks thecomputing capacity or resources then the controller forwardsthe service request to the cloud to reduce the data processinglatency

The proposed edge computing service model can bedivided into resource estimation scheduler and lightweightVNF configuration Figure 6 presents the flowchart of theoperations in the proposed edge computing service modeland these three parts are further described in the followingsections

Resource estimation checkswhether the edge gateway hasa sufficient amount of computing resources to provide edgecomputing services For the user request set R the resourceallocation must abide by the following rules which are alsoused by the edge gateway

(1) For a single service request Ri any resource of Vi(eg CPU memory and disk) is less than the totalresources of P forall119877119862119894ltPC forall119877119872119894ltPM and forall119877119863119894ltPD119877119862119894 119877119872119894 and 119877119863119894 denote the CPU memory and diskspace requests sent to Ri respectively

(2) The sum of computing resources that are allocated tothe VNF in the gateway is less than the total numberof resources of the physical machine P 119881119862119894 lt PC 119881119872119894lt PM 119881119863119894 lt PD 119881119862119894 119881119872119894 and 119881119863119894 denote the sum ofCPU memory and disk space resources allocated toVi respectively

As shown in Figure 6 after receiving the user servicerequest the edge gateway checks whether a sufficient amountof computing resources is available to satisfy such requestThe following situations may be encountered in this case

(1) If the edge gateway has a sufficient amount ofresources then the user service request is processedthrough the scheduler and queued in the systembased on the result of the task scheduling algorithm

(2) If the edge gateway does not have a sufficient amountof available resources then the user service request isdirectly transferred to the cloud instead

Task scheduling aims to increase the operational effi-ciency of the edge gateway Given that one service requestdiffers from another task scheduling examines how the

Wireless Communications and Mobile Computing 7

Forward Ri to Cloud DC

Start

Download from private storage

RCilt(PCiminusVCi)

RMilt(PMiminusVMi)

RDilt(PDiminusVDi)

No

Yes

Yes

Yes

Resource Estimation for Ri

VNF allocation

Download from public storage

End

No

No

No

Yes

Task Scheduling Docker Image is exist

Riconfigured to

the scheduler No

PCi = (PCiminusRCi)PMi = (PMiminusRMi)PDi =(PDiminusRDi)

Service on-demand VNF Matching

Yes

Figure 6 Edge computing service model flowchart

edge gateway can meet the requirements of different servicerequests and accomplish task scheduling in the system As itsprimary purpose scheduling attempts to reduce the amountof time spent on dealing with the most demanding servicerequirements asmuch as possible For this purpose this paperconstructs the Greedy Available Fit (GAF) task schedulingmechanism to enhance the operational efficiency of edgecomputing services

Assume that each service request Ri configures a VNFvirtualization service resource Vi ti denotes the processingtime of the ith service request di denotes its deadline andj denotes its completion time in the system The Ri in thesystem must be completed at schedule and before the basictime limit Otherwise this service request must be forwardeddirectly to the cloud As each service request arrives at adifferent time the time required for the operation processingand the deadline time also differ In this case deadline isselected as a priority parameter for the task scheduling Thedeadline can be used to define the processing priority forservices that is the deadline for those services that requirereal-time processing may be set according to their processingand precedence requirements on different VNFs This paperaims to insert as many tasks as possible into the schedulebefore completing the largest task Ri of the deadline Toaccomplish this objective a task with the maximum deadlineis selected as the baseline and the remaining tasks aresequentially inserted into the queue based on their processing

time119873[119894 119895] denotes the maximumnumber of j tasks selectedfrom the front i tasks and can be formulated as

119873[119894 119895] = max

119873[119894 minus 1 119895]

119873 [119894 minus 1 119895 minus 119905119894] + 1 119894119891 119895 le 119889119894119873[119894 minus 1 119895 minus 119905119894] 119894119891 119895 gt 119889119894

(7)

Equation (8) shows the initial condition of119873[119894 119895]

119873[1 119895] =

minusinfin 119894119891 119895 = 11990511 119894119891 119895 = 1199051 le 11988910 119894119891 119895 = 1199051 gt 1198891

(8)

Assume that there is no time gap betweenRi and119877119894+1 thatR1 starts from time 0 and that N = 1 If t1 le d1 then N = 0because the deadline is exceeded

Thewhole decision process is summarized inAlgorithm 1which starts with an empty schedule and inserts the availabletasks into this schedule in three different cases In case 1 if thecurrent time step j exceeds the deadline of Ri then N[i j] =N[iminus1 jminusti] In case (2) if there is not enough time to finishrequest Ri by the current time j then N[i j] = N[iminus1 j] Incase (3) if time j does not exceed the deadline of Ri and thereis enough time to finish request Ri by time j then N[i j] =max(N[iminus1 j] N[iminus1 jminusti ] + 1)

8 Wireless Communications and Mobile Computing

Input 119877 = 1198771 1198772 119877119899 ti di jOutput119873[119894 119895](1) Start(2) Set the largest task di as the baseline of the scheduler(3) for 119894 larr997888 1 to n initialize the maximum N at time 0 to be zero for each service

request in queue(4) 119873[119894 0] = 0(5) for 119895 larr997888 1 to 119889119899 determine the maximum N obtained by R1 at each time step(6) if ((j == t1) and (j lt= d1))(7) 119873[1 119895] = 1(8) else(9) 119873[1 119895] =119873[1 119895 minus 1](10) for 119894 larr997888 2 to n(11) for 119895 larr997888 1 to 119889119894(12) if (j gt di) case (1) the current time step j already exceeds 119877119894rsquos deadline(13) then119873[119894 119895] = 119873[119894 minus 1 119895 minus 119905119894](14) else if (j lt ti) case (2) there is not enough time to finish request Ri by the

current time j(15) then119873[119894 119895] = 119873[119894 minus 1 119895](16) else case (3) time j does not exceed 119877119894rsquos deadline and there is enough time

to finish request 119877119894 by time j(17) then119873[119894 119895] = max(119873[119894 minus 1 119895]119873[119894 minus 1 119895 minus 119905119894] + 1)(18) Schedule Completed(19) End

Algorithm 1 Greedy Available Fit algorithm

Virtualization technology can be applied on the CPE inmany ways such as by using the VM container or VM inte-gration of container However a traditional VM consumesmany system resources and cannot meet the requirementsfor light weight and service on-demand deployment In thiscase this research adopts container technology instead Acontainer handles only one service request at a time and stopsits operation after completing the delivery of a service TheVNF manager on the gateway is responsible for configuringand allocating each VNF The VNF template image that isrequired by different services is placed in the VNF resourcepool (with Docker Hub as the default other public or privateregistrations can also be specified) The edge gateway notonly controls the amount of gateway resources used byeach container but also manages several resources such asCPU and RAM to ensure that the container can obtain therequired resources without affecting the performance of theother executing containers on the edge gateway

Linux containers adopt a hierarchical structure to rapidlydeploy VNFs and to manage NFV flexible scheduling Theunderlying structure uses the file archiving mechanism ofDocker namely the advanced multilayered unification filesystem to incorporatemany different VNF imagesThe layersare stacked up andwhen someVNF service functions need tobe accessed the container retrieves the VNF images throughthe VNFmanager and uses these images directly To integratethe open resources of Docker Hub and to expand the privatewarehouse of the VNF service manager the system can storethe image files obtained from the public warehouse in aprivate warehouse for future use The judgment mechanismof the VNF service manager is described as follows

(1) If the image requested by the host is already in thelocal container then the VNF image can be taken outdirectly by the Docker daemon of the VNF servicemanager No complex configurations are required

(2) If the VNF image to be used by the host is not in thelocal container then the VNF manager connects tothe common warehouse Docker Hub on the Internetto find and store the desired service image file ina private warehouse via pushpull Afterward theacquired VNF image is controlled by the Dockerdaemon

By using the OS kernel for resource isolation the con-tainer technology does not need to rely on a virtual softwarelayer and does not require the VM to install a guest OSTherefore the capacity of the image file is much larger thanthat of the virtual machine The image file is small and canbe rapidly deployed through network transmission therebysaving network resources By using Linux container technol-ogy we can configure various VNF images to be executed onthe same edge gateway thereby replacing the virtual machineand achieving the goal of lightweight virtualization

5 Experiment

An experiment is conducted to evaluate the proposed taskscheduling algorithmand to test its performance in deployinglightweight VNF The experiment environment is shown inFigure 7 A simulation is initially performed to evaluate thetask scheduling performance of the algorithm by changingthe number of service requirements The service request is

Wireless Communications and Mobile Computing 9

Edge Network

Internet

Edge Gateway

DataCenter

Cloud Broker

Docker HubDocker Hub

Figure 7 The experiment environment

characterized as a Poisson process There we showed that theaverage number of service requests in the system is given by120582120583 where 120582 is the average arrival rate and 120583 is the averageservice rate The input parameter 120582120583 is 045 First comefirst service (FCFS) priority task scheduling [34] RR [37]and GBA [36] are then compared with the proposed GAFmechanism

The effect of the service requests on the average waitingtime average response time and task scheduling of differentscheduling mechanisms is evaluated in the simulation FCFSis based on the service requests that enter the edge gatewayqueuing system and schedules the tasks according to thesequence of these requests Priority task scheduling is anunfair scheduling algorithm where the service requests aresorted according to their priority and where the high-prioritytasks are performedfirstThose tasks having the same priorityare sorted by using the FIFO scheduling mechanism TheGBA sorts the tasks based on the completion time of theservice requests in queuing systemTheRR algorithm is basedon the conventional RR scheduling performed in the processscheduling RR scheduling fairly allocates the computingresources to those tasks with the same priority by using thetime slicing approach (time quantum)

Figure 8 shows the experiment result for average waitingtime where the x-axis refers to the number of service requestsand the y main axis refers to the average waiting time thatis the time a user spends to complete the service schedulingand to determine the resource allocation after receiving aservice request and a successful reply If the demand revertsto an errortimeout or if the edge gateway does not haveenough computing resources to support such demand thenthe sample is excluded from the calculation of the averagewaiting time As can be seen in the experimental resultthe RR algorithm obtains the longest average waiting timebecause of its application of time slice rules in order for eachtask to be processed within a fixed amount of time The timequantum affects the overall performance of the operation

0100200300400500600

10 20 30 40 50 60 70 80 90 100

Aver

age W

aitin

g Ti

me (

ms)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 8 The experiment result for average waiting time

Having a very long time quantum leads to a very long waitingtime while having a very short time quantum results intask schedule conversion a poor execution efficiency and anextended waiting time For FCFS given that the time spentdiffers across each task the waiting time for the next taskmust be determined based on the schedule of the previoustask Therefore if the last task schedule is too long thesystem cannot quickly process the subsequent task schedulesthereby affecting the overall task scheduling efficiency Giventhat its average waiting time is relatively shorter than thatof the FCFS mechanism priority task scheduling can satisfythe task requirements and flexibly perform the schedulingHowever prioritizing a high-priority program will delay allof the low-priority requests thereby creating an indefinitesituation in the low-priority program and extending thewaiting time GBA is based on the earliest completion time ofthe taskWhen the service demand is low the average waitingtime of the GBA algorithm is similar to that of priority taskscheduling However when the service demand increases theaverage waiting time of the GBAmechanism becomes longerthan that of priority task scheduling Compared with FCFS

10 Wireless Communications and Mobile Computing

050

100150200250300350

10 20 30 40 50 60 70 80 90 100

Ave

rage

Res

pons

e Tim

e (m

s)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 9 The experiment result for average response time

01020304050607080

10 20 30 40 50 60 70 80 90 100

Num

ber o

f Sch

edul

ed T

asks

Number of Service RequestsGAFFCFSPriority

GBARR

Figure 10 The experiment result for task scheduling efficiency

priority task scheduling RR and GBA the proposed GAFalgorithm selects the largest deadline task as the baseline andthen sequentially inserts the remaining tasks into the queuebased on their processing timeTherefore the GAF algorithmobtains the shortest average waiting time

Figure 9 shows the results for average response time Asexpected the RR algorithm outperforms all other approachesin terms of average response time because this algorithm isespecially designed for time sharing Using a fair strategy toallocate VNF resources ensures that each service request hasa fixed time If the service processing time exceeds the timequantum allocated by the system then the sample is excludedfrom the calculation of the average response time Given thatFCFS is processed in sequence the average response timewill be affected by the time spent by the previous 119877119894minus1 taskin the scheduler For example if the previous service requesthas a very long task schedule then the average responsetime increases Given that the proposed GAF algorithmsequentially inserts tasks into the queue its average responsetime is slightly longer than that of FCFS and GBA Althoughthe priority algorithm can sort the tasks according to thepriorities of the service requests the low-priority tasks areeasily delayed thereby increasing the average response time

Figure 10 shows the task scheduling efficiency of thecompared algorithms If the service request number is lessthan 60 then each algorithm can complete the task schedul-ing However when this number is exceeded the number

of tasks that can be scheduled varies across each algorithmGAF outperforms all the other algorithms as it ranks thelargest deadline task at the end and then sequentially sortsthe remaining tasks according to their processing time In thisway this algorithm ensures that most tasks will be completedwithin a minimum processing time Those service requeststhat cannot be scheduled will bypass the gateway and will bedirectly forwarded to the cloud data center When the servicerequest number is 80 GAF can complete 68 of the taskscheduling Meanwhile GBA performs the scheduling basedon the earliest task completion time Although GAF andGBAcan schedule the same number of tasks in the experimentalenvironment the latter has a longer average waiting timecompared with the former GAF also outperforms FCFCand priority task scheduling by 96 and 62 in terms oftask scheduling efficiency respectively because this approachranks the largest deadline task at the end and then sorts theremaining tasks according to their required processing timeGAF also outperforms theRRmechanismby 70because thelatter applies the time-sharing rule to minimize the numberof scheduled tasks Based on these results the GAF algorithmcan schedule more tasks in the edge gateway within a shortaverage waiting time and improve the operational efficiencyof edge computing services

In addition to task scheduling the VNF configuration isanother main factor that affects the operational efficiency ofthe edge computing service model This paper utilizes theDocker technology to achieve an agile deployment of VNFsSuch technology can provide a centralized management andflexible configuration of VNF services Depending on theservice requests of each user the deployment of differentVNFs can meet the application requirements of serviceson-demand The VNF image of Docker can be mainlydivided into four categories namely system images (egUbuntu CentOS and Debian) tool images (eg GolangFlask and Tomcat) service images (eg MySQL MongoDBand RabbitMQ) and application images (eg WordPressand DRUPAL) To test the practical deployment efficiency oflightweight VNF for the edge gateway a pull performancetest is performed for each type of Docker VNF image Theedge gateway is simulated by using a standalone PC with 34GHz dual-core CPU 16GBmemory and 5400 rpmHD500GThe experimental network bandwidth is set to 100 Mbps Tocompare the pull performance with the local edge gateway atest VM in the Google Cloud platform is used to simulate thecloud network broker Table 1 presents the result of the pullperformance test

As shown in Table 1 each VNF image size has differentphysical gateway and cloud network broker costs due to thelimitations in the network bandwidth A larger VNF imagetakes a longer time to be downloaded from Docker Hub Asexpected the cloud network broker has a better computingarchitecture than the local edge gateway because the locationof the edge gateway affects the performance of the VNFdeployment Although the edge computing architecture ofthe cloud network broker deploys VNF more efficiently thanthe local gateway the user service requestsmust still be sent tothe Internet for processing and this methodmakes it difficultto demonstrate the technical advantages of edge computing

Wireless Communications and Mobile Computing 11

Table 1 Pull performance test results for different VNFs

Image Types VNFs Image Size (MB) Physical GatewayCost (seconds)

Cloud NetworkBroker Cost(seconds)

System imagesubuntu 112 22095 7918centos 207 24319 8926debian 207 21635 5646

Tool imagesgolang 779 86793 1851flask 458 6222 1633

prometheus 112 2414 4737

Service images

httpd 177 20753 1031nginx 109 13216 4778bind 337 8438 15105squid 215 150031 9973mysql 374 24852 10449

mongodb 503 61347 17116

Application images wordpress 407 32866 13213drupal 452 8139 14405

11095

1593

0

20

40

60

80

100

120

140

160

180

Docker VM

Tim

e (s)

Figure 11 VNF boot time test for the VM and container

In the VNF boot time test we set the image to Ubuntu-16-04-x64 VM and Docker are used to perform the boottime test on the edge gateway We test the time of openingfive VMs and Docker containers VMware is used as the VMhypervisor As shown in Figure 11 the five VMs are openedin 1593 seconds while the Docker containers are openedin 11095 seconds In sum the VMs have a much longerdeployment time than Docker More time can be saved byusingDocker in the edge gateway that requires a large amountof VNFs

To test the average VNF boot time we boot a VM waitfor its activation shut it down and repeat the same steps for14 other VMs The test results are shown in Figure 12 Whenusing VMware the average time to boot the VMs is about3186 seconds However when using Docker these VMs arebooted in only 22 seconds In sum the Docker containerperforms approximately 15 times faster than the VM

The results of the VNF deployment test highlight thesignificant advantages of lightweight containers over tradi-tional virtualizationmethodsThese containers can be started

2219

3186

0

5

10

15

20

25

30

35

Docker VM

Tim

e (s)

Figure 12 Average VNF boot time test for the VM and container

within a few seconds because they adopt a common host OStherebymaking it unnecessary to execute the guestOS in eachcontainer The container also does not need to wait for theOS to start up thereby saving a few seconds Fast start is farfaster than traditional VMs Meanwhile given the design oflightweight VNF images and its highly efficient virtualizationthe application-centric virtualization technology Docker canautomatically obtain container images from the Docker Hubfor flexible deployment and management and can meet theapplication requirements of the edge network for the serviceon-demand of VNF deployment

6 Conclusions

This paper proposed a gateway-based edge computing servicemodel and a GAF task scheduling mechanism that allowsthe edge gateway to schedule more tasks within a short aver-age waiting time The resource estimation and lightweightVNF configuration technologies are used to improve theoperational efficiency of edge computing services to increase

12 Wireless Communications and Mobile Computing

resource utilization to achieve a rapid deployment of VNFand to meet service on-demand requests The combinationof resource estimation task scheduling and lightweight VNFconfiguration design provides an integrated solution that cansatisfy the service on-demand requests for 5G networks Thesimulation results showed that the proposedGAFmechanismoutperforms the other scheduling algorithmsMeanwhile thecomparison revealed that using the lightweight virtualizationtechnology in edge gateways ismore efficient and competitivethan using traditional VMs

In the future we plan to study the multiqueue taskscheduling problem for edge computing the possible coop-eration between edge gateway and cloud servers and the useof SD-WAN technology to achieve a seamless operation of thecloud-edge network

Data Availability

The data used to support the findings of this study areincluded within the article

Disclosure

The funder had no role in the study design data collectionand analysis decision to publish or preparation of thismanuscript

Conflicts of Interest

The authors declare no conflicts of interest regarding thepublication of this paper

Acknowledgments

The work described in this paper was supported in part bythe Ministry of Science and Technology of the Republic ofChina (Grants nos 104-2221-E-008-039 105-2221-E-008-071107-2623-E-008-002 and 107-2636-E-003-001)

References

[1] J Matias J Garay N Toledo J Unzilla and E Jacob ldquoTowardan SDN-enabled NFV architecturerdquo IEEE CommunicationsMagazine vol 53 no 4 pp 187ndash193 2015

[2] Q Duan N Ansari and M Toy ldquoSoftware-defined networkvirtualization An architectural framework for integrating SDNand NFV for service provisioning in future networksrdquo IEEENetwork vol 30 no 5 pp 10ndash16 2016

[3] TSI ldquoNetwork Functions Virtualization (NFV) architecturalframeworkrdquo[Online] Available httpswwwetsiorgdeliveretsi gsNFV001 099002010101 60gs NFV002v010101ppdf

[4] ONF ldquoSDN Architecturerdquo [Online] Available httpwwwopennetworkingorgimagesstoriesdownloadssdn-resourcestechnical-reportsTR SDN ARCH 10 06062014pdf

[5] A Basta A Blenk K Hoffmann H J Morper M Hoffmannand W Kellerer ldquoTowards a Cost Optimal Design for a 5GMobile Core Network Based on SDN and NFVrdquo IEEE Trans-actions on Network and Service Management vol 14 no 4 pp1061ndash1075 2017

[6] F Z Yousaf M Bredel S Schaller and F Schneider ldquoNFV andSDN-Key technology enablers for 5G networksrdquo IEEE Journal

on Selected Areas in Communications vol 35 no 11 pp 2468ndash2478 2017

[7] A C Baktir A Ozgovde and C Ersoy ldquoHow Can EdgeComputing Benefit FromSoftware-DefinedNetworking A Sur-vey Use Cases and Future Directionsrdquo EEE CommunicationsSurveys amp Tutorials vol 19 no 4 pp 2359ndash2391 2017

[8] Guangshun Li Jiping Wang Junhua Wu and Jianrong SongldquoData Processing Delay Optimization in Mobile Edge Comput-ingrdquoWireless Communications andMobile Computing vol 2018Article ID 6897523 9 pages 2018

[9] Zhimin Wang Qinglin Zhao Fangxin Xu Hongning Daiand Yujun Zhang ldquoDetection Performance of Packet Arrivalunder Downclocking for Mobile Edge Computingrdquo WirelessCommunications and Mobile Computing vol 2018 Article ID9641712 7 pages 2018

[10] W Yu F Liang X He et al ldquoA Survey on the Edge Computingfor the Internet of Thingsrdquo IEEE Access vol 6 pp 6900ndash69192018

[11] P Mach and Z Becvar ldquoMobile Edge Computing A Survey onArchitecture and Computation Offloadingrdquo IEEE Communica-tions Surveys amp Tutorials vol 19 no 3 pp 1628ndash1656 2017

[12] N Abbas Y Zhang A Taherkordi and T Skeie ldquoMobile EdgeComputing A Surveyrdquo IEEE Internet of Things Journal vol 5no 1 pp 450ndash465 2018

[13] Guangshun Li Jianrong Song Junhua Wu and Jiping WangldquoMethod of Resource Estimation Based on QoS in Edge Com-putingrdquo Wireless Communications and Mobile Computing vol2018 Article ID 7308913 9 pages 2018

[14] D Happ and A Wolisz ldquoTowards gateway to Cloud offloadingin IoT publishsubscribe systemsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 101ndash106 Valencia Spain May 2017

[15] G Premsankar M Di Francesco and T Taleb ldquoEdge Comput-ing for the Internet of Things A Case Studyrdquo IEEE Internet ofThings Journal vol 5 no 2 pp 1275ndash1284 2018

[16] L Zhao W Sun Y Shi and J Liu ldquoOptimal Placementof Cloudlets for Access Delay Minimization in SDN-BasedInternet of Things Networksrdquo IEEE Internet of Things Journalvol 5 no 2 pp 1334ndash1344 2018

[17] SWang X Zhang Y Zhang LWang J Yang andWWang ldquoASurvey onMobile Edge Networks Convergence of ComputingCaching and Communicationsrdquo IEEE Access vol 5 pp 6757ndash6779 2017

[18] Y Liu J E Fieldsend and G Min ldquoA Framework of FogComputing ArchitectureChallenges and Optimizationrdquo IEEEAccess vol 5 pp 25445ndash25454 2017

[19] ETSI Multi-access Edge Computing [Online] Availablehttpwwwetsiorgtechnologies-clusterstechnologiesmulti-access-edge-computing

[20] OpenFog Consortium [Online] Available httpswwwopen-fogconsortiumorg

[21] Y Mao C You J Zhang K Huang and K B Letaief ldquoA SurveyonMobile Edge ComputingThe Communication PerspectiverdquoIEEE Communications Surveys amp Tutorials vol 19 no 4 pp2322ndash2358 2017

[22] Cisco ldquoThe Internet of Things How the Next Evolution ofthe Internet Is Changing Everything white paperrdquo [Online]Available httpswwwciscocomcdamen usIoT IBSG0411FINALpdf

[23] IDC ldquoFutureScape Worldwide Internet of Things 2017 Pre-dictionsrdquo November 2016 [Online] Available httpswwwidccomgetdocjspcontainerId=US41910716

Wireless Communications and Mobile Computing 5

Virtual Network Functions

VNF

Hardware Resource

Com

putin

g Re

sour

ce

Service RequestVNF

VNF

VNF

R

R

R

R

Figure 4 Network function scheduling problem in CPE

3 Problem Definition

NFV is currently the most valued solution to the prob-lems associated with the operation cost and efficiency ofnetwork services Accordingly many ICT companies havebegun to examine virtualization technology Despite thehigh expectations toward this technology satisfying servicerequests effectively allocating virtual computing resourcesand providing service on-demand application still challengethe deployment of NFV and play key roles in the futuredevelopment of 5GIoT services NFV also allows the estab-lishment of multiple independent heterogeneous virtualnetworks based on common underlying network resourcesthereby enabling service providers to provide customizedservices according to the demands of users Virtual networkembedding is a process of mapping the virtual networkto the underlying network (substrate network) through themapping algorithm and according to the current resourcesituation of the infrastructure providerThis process is anNP-Hard problem and has been investigated in NFV resourceallocation research [38 39] Previous studies [40 41] havealso systematically discussed the problem of virtual networkmapping and provided good references for research

This paper aims to optimize task scheduling and resourceallocation by using the proposed edge computing servicemodel Scheduling tasks in edge computing aremore complexthan that in cloud computing An edge computing operationis typically spread over the device of the client the edgegateways and occasionally a broker of the cloud networkTherefore deciding where to schedule computational tasksremains a key problem in edge computing Given that thegateway-based IoT architecture is currently the mainstreamthe task scheduling mechanism discussed in this paperfocuses on the edge gateway Task scheduling and resourceallocation are the main problems to be solved in this gatewaygiven the limited amount of available computing resourcesLightweight NFV technology plays an indispensable role inthe rapid deployment of the edge gateway The relationshipbetween task scheduling and resource allocation for hostservice requests is plotted in Figure 4 As each service requestarrives at a different time the required VNF service andprocessing time also differ The edge gateway can adjust thetask scheduling and resource configuration according to the

differences among the service requests As the basic idea forthe service on-demand model only one VNF is dedicated toa single service request This service model problem can beanalyzed by using queuing theory which regards the requestand react processing as a waiting-line system the input of awaiting-line system as the service request the service counteras the gateway scheduling function and the output as therequested VNF resource Although many queuing modelsmay be used in operations management [42] this paperprimarily focuses on the task scheduling approach that allowsa certain edge gateway to processmultipleVNFs fromaqueueone after another Given its limited capacity the edge gatewaycan only schedule a limited number of service requests whilethe subsequent service requests need to be forwarded to thecloud for processing With these considerations the researchproblem is formulated as follows Assume that the servicerequest is Poissonrsquos ratio within one edge gateway and thatthe service time is an exponential distribution The servicerequest set can be denoted as 119877 = 1198771 1198772 1198773 119877119899 wheren denotes the number of service requests in the systemEquations (1) and (2) are used to determine the probabilityof n service requests in the system In these equations Ndenotes the maximum number of service requests that canbe scheduled in the gateway 120582 is the average number ofincoming user requests in one unit of time 120583 is the serviceefficiency (the ability of the service counter) 120588 = 120582120583 is theratio that the request can be met within one unit of time P0denotes the initial condition L denotes the total number ofservice requests arriving in the systemwithin a planned timeand 119871119902 denotes the total number of server requests queued inthe system L and 119871119902 can be computed by using (3) and (4)respectively Equation (5) computes the waiting and servicetime of a service in the system while (6) computes119882119902 or theaverage waiting time for a user request

1198750 =

1 minus 1205881 minus 120588119873+1 120588 = 11119873 + 1 120588 = 1

(1)

119875119899 =

1205881198991198750 = 1205881198991 minus 1205881 minus 120588119873+1 120588 = 1

1198750 =1119873 + 1 120588 = 1

(2)

6 Wireless Communications and Mobile Computing

Binaries Libraries

Host OS

Hardware Resource

VNF

Docker Engine

ResourceEstimation Scheduler VNF

Configuration

VN

F Manager

Container images

R

R

R

R

R

Edge Computing Node VNFs Resource PoolService Request

VNF

VNF

PublicRegistry

PrivateRegistry

SDN Controller

Figure 5 Service on-demand edge computing model

119871 =

1205881 minus 120588 minus

(119873 + 1) 120588119873+1

1 minus 120588119873+1 120588 = 11198732 120588 = 1

(3)

119871119902 = L minus (1 minus 1198750) (4)

119882 = 119871120582 (1 minus 119875119899)

(5)

119882119902 =119871119902120582

(6)

The results reveal that the waiting time for the userrequest can be reduced in two ways namely by reducing thenumber of service requests queued in the edge gateway andby increasing the processing speed of task scheduling in theedge computing gateway A large number of tasks must becompleted within a short period to achieve an efficient edgecomputing

4 Design of the Edge ComputingService Model

How to construct an elastic and cost-effective edge com-puting service model improve management efficiency meetuser service requests achieve centralized management anddevelop flexible configuration service models are the currentresearch trends in the development of a 5G SDNNFV net-work This paper proposes a gateway-based edge computingservicemodel (Figure 5) to improve the operational efficiencyof the edge computing node to accelerate the processingof user service requests and to increase the utilizationefficiency of a limited number of computing resources In thismodel when different user requests enter the edge gatewaythis gateway determines whether the requested services canbe processed or not If the edge gateway itself lacks thecomputing capacity or resources then the controller forwardsthe service request to the cloud to reduce the data processinglatency

The proposed edge computing service model can bedivided into resource estimation scheduler and lightweightVNF configuration Figure 6 presents the flowchart of theoperations in the proposed edge computing service modeland these three parts are further described in the followingsections

Resource estimation checkswhether the edge gateway hasa sufficient amount of computing resources to provide edgecomputing services For the user request set R the resourceallocation must abide by the following rules which are alsoused by the edge gateway

(1) For a single service request Ri any resource of Vi(eg CPU memory and disk) is less than the totalresources of P forall119877119862119894ltPC forall119877119872119894ltPM and forall119877119863119894ltPD119877119862119894 119877119872119894 and 119877119863119894 denote the CPU memory and diskspace requests sent to Ri respectively

(2) The sum of computing resources that are allocated tothe VNF in the gateway is less than the total numberof resources of the physical machine P 119881119862119894 lt PC 119881119872119894lt PM 119881119863119894 lt PD 119881119862119894 119881119872119894 and 119881119863119894 denote the sum ofCPU memory and disk space resources allocated toVi respectively

As shown in Figure 6 after receiving the user servicerequest the edge gateway checks whether a sufficient amountof computing resources is available to satisfy such requestThe following situations may be encountered in this case

(1) If the edge gateway has a sufficient amount ofresources then the user service request is processedthrough the scheduler and queued in the systembased on the result of the task scheduling algorithm

(2) If the edge gateway does not have a sufficient amountof available resources then the user service request isdirectly transferred to the cloud instead

Task scheduling aims to increase the operational effi-ciency of the edge gateway Given that one service requestdiffers from another task scheduling examines how the

Wireless Communications and Mobile Computing 7

Forward Ri to Cloud DC

Start

Download from private storage

RCilt(PCiminusVCi)

RMilt(PMiminusVMi)

RDilt(PDiminusVDi)

No

Yes

Yes

Yes

Resource Estimation for Ri

VNF allocation

Download from public storage

End

No

No

No

Yes

Task Scheduling Docker Image is exist

Riconfigured to

the scheduler No

PCi = (PCiminusRCi)PMi = (PMiminusRMi)PDi =(PDiminusRDi)

Service on-demand VNF Matching

Yes

Figure 6 Edge computing service model flowchart

edge gateway can meet the requirements of different servicerequests and accomplish task scheduling in the system As itsprimary purpose scheduling attempts to reduce the amountof time spent on dealing with the most demanding servicerequirements asmuch as possible For this purpose this paperconstructs the Greedy Available Fit (GAF) task schedulingmechanism to enhance the operational efficiency of edgecomputing services

Assume that each service request Ri configures a VNFvirtualization service resource Vi ti denotes the processingtime of the ith service request di denotes its deadline andj denotes its completion time in the system The Ri in thesystem must be completed at schedule and before the basictime limit Otherwise this service request must be forwardeddirectly to the cloud As each service request arrives at adifferent time the time required for the operation processingand the deadline time also differ In this case deadline isselected as a priority parameter for the task scheduling Thedeadline can be used to define the processing priority forservices that is the deadline for those services that requirereal-time processing may be set according to their processingand precedence requirements on different VNFs This paperaims to insert as many tasks as possible into the schedulebefore completing the largest task Ri of the deadline Toaccomplish this objective a task with the maximum deadlineis selected as the baseline and the remaining tasks aresequentially inserted into the queue based on their processing

time119873[119894 119895] denotes the maximumnumber of j tasks selectedfrom the front i tasks and can be formulated as

119873[119894 119895] = max

119873[119894 minus 1 119895]

119873 [119894 minus 1 119895 minus 119905119894] + 1 119894119891 119895 le 119889119894119873[119894 minus 1 119895 minus 119905119894] 119894119891 119895 gt 119889119894

(7)

Equation (8) shows the initial condition of119873[119894 119895]

119873[1 119895] =

minusinfin 119894119891 119895 = 11990511 119894119891 119895 = 1199051 le 11988910 119894119891 119895 = 1199051 gt 1198891

(8)

Assume that there is no time gap betweenRi and119877119894+1 thatR1 starts from time 0 and that N = 1 If t1 le d1 then N = 0because the deadline is exceeded

Thewhole decision process is summarized inAlgorithm 1which starts with an empty schedule and inserts the availabletasks into this schedule in three different cases In case 1 if thecurrent time step j exceeds the deadline of Ri then N[i j] =N[iminus1 jminusti] In case (2) if there is not enough time to finishrequest Ri by the current time j then N[i j] = N[iminus1 j] Incase (3) if time j does not exceed the deadline of Ri and thereis enough time to finish request Ri by time j then N[i j] =max(N[iminus1 j] N[iminus1 jminusti ] + 1)

8 Wireless Communications and Mobile Computing

Input 119877 = 1198771 1198772 119877119899 ti di jOutput119873[119894 119895](1) Start(2) Set the largest task di as the baseline of the scheduler(3) for 119894 larr997888 1 to n initialize the maximum N at time 0 to be zero for each service

request in queue(4) 119873[119894 0] = 0(5) for 119895 larr997888 1 to 119889119899 determine the maximum N obtained by R1 at each time step(6) if ((j == t1) and (j lt= d1))(7) 119873[1 119895] = 1(8) else(9) 119873[1 119895] =119873[1 119895 minus 1](10) for 119894 larr997888 2 to n(11) for 119895 larr997888 1 to 119889119894(12) if (j gt di) case (1) the current time step j already exceeds 119877119894rsquos deadline(13) then119873[119894 119895] = 119873[119894 minus 1 119895 minus 119905119894](14) else if (j lt ti) case (2) there is not enough time to finish request Ri by the

current time j(15) then119873[119894 119895] = 119873[119894 minus 1 119895](16) else case (3) time j does not exceed 119877119894rsquos deadline and there is enough time

to finish request 119877119894 by time j(17) then119873[119894 119895] = max(119873[119894 minus 1 119895]119873[119894 minus 1 119895 minus 119905119894] + 1)(18) Schedule Completed(19) End

Algorithm 1 Greedy Available Fit algorithm

Virtualization technology can be applied on the CPE inmany ways such as by using the VM container or VM inte-gration of container However a traditional VM consumesmany system resources and cannot meet the requirementsfor light weight and service on-demand deployment In thiscase this research adopts container technology instead Acontainer handles only one service request at a time and stopsits operation after completing the delivery of a service TheVNF manager on the gateway is responsible for configuringand allocating each VNF The VNF template image that isrequired by different services is placed in the VNF resourcepool (with Docker Hub as the default other public or privateregistrations can also be specified) The edge gateway notonly controls the amount of gateway resources used byeach container but also manages several resources such asCPU and RAM to ensure that the container can obtain therequired resources without affecting the performance of theother executing containers on the edge gateway

Linux containers adopt a hierarchical structure to rapidlydeploy VNFs and to manage NFV flexible scheduling Theunderlying structure uses the file archiving mechanism ofDocker namely the advanced multilayered unification filesystem to incorporatemany different VNF imagesThe layersare stacked up andwhen someVNF service functions need tobe accessed the container retrieves the VNF images throughthe VNFmanager and uses these images directly To integratethe open resources of Docker Hub and to expand the privatewarehouse of the VNF service manager the system can storethe image files obtained from the public warehouse in aprivate warehouse for future use The judgment mechanismof the VNF service manager is described as follows

(1) If the image requested by the host is already in thelocal container then the VNF image can be taken outdirectly by the Docker daemon of the VNF servicemanager No complex configurations are required

(2) If the VNF image to be used by the host is not in thelocal container then the VNF manager connects tothe common warehouse Docker Hub on the Internetto find and store the desired service image file ina private warehouse via pushpull Afterward theacquired VNF image is controlled by the Dockerdaemon

By using the OS kernel for resource isolation the con-tainer technology does not need to rely on a virtual softwarelayer and does not require the VM to install a guest OSTherefore the capacity of the image file is much larger thanthat of the virtual machine The image file is small and canbe rapidly deployed through network transmission therebysaving network resources By using Linux container technol-ogy we can configure various VNF images to be executed onthe same edge gateway thereby replacing the virtual machineand achieving the goal of lightweight virtualization

5 Experiment

An experiment is conducted to evaluate the proposed taskscheduling algorithmand to test its performance in deployinglightweight VNF The experiment environment is shown inFigure 7 A simulation is initially performed to evaluate thetask scheduling performance of the algorithm by changingthe number of service requirements The service request is

Wireless Communications and Mobile Computing 9

Edge Network

Internet

Edge Gateway

DataCenter

Cloud Broker

Docker HubDocker Hub

Figure 7 The experiment environment

characterized as a Poisson process There we showed that theaverage number of service requests in the system is given by120582120583 where 120582 is the average arrival rate and 120583 is the averageservice rate The input parameter 120582120583 is 045 First comefirst service (FCFS) priority task scheduling [34] RR [37]and GBA [36] are then compared with the proposed GAFmechanism

The effect of the service requests on the average waitingtime average response time and task scheduling of differentscheduling mechanisms is evaluated in the simulation FCFSis based on the service requests that enter the edge gatewayqueuing system and schedules the tasks according to thesequence of these requests Priority task scheduling is anunfair scheduling algorithm where the service requests aresorted according to their priority and where the high-prioritytasks are performedfirstThose tasks having the same priorityare sorted by using the FIFO scheduling mechanism TheGBA sorts the tasks based on the completion time of theservice requests in queuing systemTheRR algorithm is basedon the conventional RR scheduling performed in the processscheduling RR scheduling fairly allocates the computingresources to those tasks with the same priority by using thetime slicing approach (time quantum)

Figure 8 shows the experiment result for average waitingtime where the x-axis refers to the number of service requestsand the y main axis refers to the average waiting time thatis the time a user spends to complete the service schedulingand to determine the resource allocation after receiving aservice request and a successful reply If the demand revertsto an errortimeout or if the edge gateway does not haveenough computing resources to support such demand thenthe sample is excluded from the calculation of the averagewaiting time As can be seen in the experimental resultthe RR algorithm obtains the longest average waiting timebecause of its application of time slice rules in order for eachtask to be processed within a fixed amount of time The timequantum affects the overall performance of the operation

0100200300400500600

10 20 30 40 50 60 70 80 90 100

Aver

age W

aitin

g Ti

me (

ms)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 8 The experiment result for average waiting time

Having a very long time quantum leads to a very long waitingtime while having a very short time quantum results intask schedule conversion a poor execution efficiency and anextended waiting time For FCFS given that the time spentdiffers across each task the waiting time for the next taskmust be determined based on the schedule of the previoustask Therefore if the last task schedule is too long thesystem cannot quickly process the subsequent task schedulesthereby affecting the overall task scheduling efficiency Giventhat its average waiting time is relatively shorter than thatof the FCFS mechanism priority task scheduling can satisfythe task requirements and flexibly perform the schedulingHowever prioritizing a high-priority program will delay allof the low-priority requests thereby creating an indefinitesituation in the low-priority program and extending thewaiting time GBA is based on the earliest completion time ofthe taskWhen the service demand is low the average waitingtime of the GBA algorithm is similar to that of priority taskscheduling However when the service demand increases theaverage waiting time of the GBAmechanism becomes longerthan that of priority task scheduling Compared with FCFS

10 Wireless Communications and Mobile Computing

050

100150200250300350

10 20 30 40 50 60 70 80 90 100

Ave

rage

Res

pons

e Tim

e (m

s)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 9 The experiment result for average response time

01020304050607080

10 20 30 40 50 60 70 80 90 100

Num

ber o

f Sch

edul

ed T

asks

Number of Service RequestsGAFFCFSPriority

GBARR

Figure 10 The experiment result for task scheduling efficiency

priority task scheduling RR and GBA the proposed GAFalgorithm selects the largest deadline task as the baseline andthen sequentially inserts the remaining tasks into the queuebased on their processing timeTherefore the GAF algorithmobtains the shortest average waiting time

Figure 9 shows the results for average response time Asexpected the RR algorithm outperforms all other approachesin terms of average response time because this algorithm isespecially designed for time sharing Using a fair strategy toallocate VNF resources ensures that each service request hasa fixed time If the service processing time exceeds the timequantum allocated by the system then the sample is excludedfrom the calculation of the average response time Given thatFCFS is processed in sequence the average response timewill be affected by the time spent by the previous 119877119894minus1 taskin the scheduler For example if the previous service requesthas a very long task schedule then the average responsetime increases Given that the proposed GAF algorithmsequentially inserts tasks into the queue its average responsetime is slightly longer than that of FCFS and GBA Althoughthe priority algorithm can sort the tasks according to thepriorities of the service requests the low-priority tasks areeasily delayed thereby increasing the average response time

Figure 10 shows the task scheduling efficiency of thecompared algorithms If the service request number is lessthan 60 then each algorithm can complete the task schedul-ing However when this number is exceeded the number

of tasks that can be scheduled varies across each algorithmGAF outperforms all the other algorithms as it ranks thelargest deadline task at the end and then sequentially sortsthe remaining tasks according to their processing time In thisway this algorithm ensures that most tasks will be completedwithin a minimum processing time Those service requeststhat cannot be scheduled will bypass the gateway and will bedirectly forwarded to the cloud data center When the servicerequest number is 80 GAF can complete 68 of the taskscheduling Meanwhile GBA performs the scheduling basedon the earliest task completion time Although GAF andGBAcan schedule the same number of tasks in the experimentalenvironment the latter has a longer average waiting timecompared with the former GAF also outperforms FCFCand priority task scheduling by 96 and 62 in terms oftask scheduling efficiency respectively because this approachranks the largest deadline task at the end and then sorts theremaining tasks according to their required processing timeGAF also outperforms theRRmechanismby 70because thelatter applies the time-sharing rule to minimize the numberof scheduled tasks Based on these results the GAF algorithmcan schedule more tasks in the edge gateway within a shortaverage waiting time and improve the operational efficiencyof edge computing services

In addition to task scheduling the VNF configuration isanother main factor that affects the operational efficiency ofthe edge computing service model This paper utilizes theDocker technology to achieve an agile deployment of VNFsSuch technology can provide a centralized management andflexible configuration of VNF services Depending on theservice requests of each user the deployment of differentVNFs can meet the application requirements of serviceson-demand The VNF image of Docker can be mainlydivided into four categories namely system images (egUbuntu CentOS and Debian) tool images (eg GolangFlask and Tomcat) service images (eg MySQL MongoDBand RabbitMQ) and application images (eg WordPressand DRUPAL) To test the practical deployment efficiency oflightweight VNF for the edge gateway a pull performancetest is performed for each type of Docker VNF image Theedge gateway is simulated by using a standalone PC with 34GHz dual-core CPU 16GBmemory and 5400 rpmHD500GThe experimental network bandwidth is set to 100 Mbps Tocompare the pull performance with the local edge gateway atest VM in the Google Cloud platform is used to simulate thecloud network broker Table 1 presents the result of the pullperformance test

As shown in Table 1 each VNF image size has differentphysical gateway and cloud network broker costs due to thelimitations in the network bandwidth A larger VNF imagetakes a longer time to be downloaded from Docker Hub Asexpected the cloud network broker has a better computingarchitecture than the local edge gateway because the locationof the edge gateway affects the performance of the VNFdeployment Although the edge computing architecture ofthe cloud network broker deploys VNF more efficiently thanthe local gateway the user service requestsmust still be sent tothe Internet for processing and this methodmakes it difficultto demonstrate the technical advantages of edge computing

Wireless Communications and Mobile Computing 11

Table 1 Pull performance test results for different VNFs

Image Types VNFs Image Size (MB) Physical GatewayCost (seconds)

Cloud NetworkBroker Cost(seconds)

System imagesubuntu 112 22095 7918centos 207 24319 8926debian 207 21635 5646

Tool imagesgolang 779 86793 1851flask 458 6222 1633

prometheus 112 2414 4737

Service images

httpd 177 20753 1031nginx 109 13216 4778bind 337 8438 15105squid 215 150031 9973mysql 374 24852 10449

mongodb 503 61347 17116

Application images wordpress 407 32866 13213drupal 452 8139 14405

11095

1593

0

20

40

60

80

100

120

140

160

180

Docker VM

Tim

e (s)

Figure 11 VNF boot time test for the VM and container

In the VNF boot time test we set the image to Ubuntu-16-04-x64 VM and Docker are used to perform the boottime test on the edge gateway We test the time of openingfive VMs and Docker containers VMware is used as the VMhypervisor As shown in Figure 11 the five VMs are openedin 1593 seconds while the Docker containers are openedin 11095 seconds In sum the VMs have a much longerdeployment time than Docker More time can be saved byusingDocker in the edge gateway that requires a large amountof VNFs

To test the average VNF boot time we boot a VM waitfor its activation shut it down and repeat the same steps for14 other VMs The test results are shown in Figure 12 Whenusing VMware the average time to boot the VMs is about3186 seconds However when using Docker these VMs arebooted in only 22 seconds In sum the Docker containerperforms approximately 15 times faster than the VM

The results of the VNF deployment test highlight thesignificant advantages of lightweight containers over tradi-tional virtualizationmethodsThese containers can be started

2219

3186

0

5

10

15

20

25

30

35

Docker VM

Tim

e (s)

Figure 12 Average VNF boot time test for the VM and container

within a few seconds because they adopt a common host OStherebymaking it unnecessary to execute the guestOS in eachcontainer The container also does not need to wait for theOS to start up thereby saving a few seconds Fast start is farfaster than traditional VMs Meanwhile given the design oflightweight VNF images and its highly efficient virtualizationthe application-centric virtualization technology Docker canautomatically obtain container images from the Docker Hubfor flexible deployment and management and can meet theapplication requirements of the edge network for the serviceon-demand of VNF deployment

6 Conclusions

This paper proposed a gateway-based edge computing servicemodel and a GAF task scheduling mechanism that allowsthe edge gateway to schedule more tasks within a short aver-age waiting time The resource estimation and lightweightVNF configuration technologies are used to improve theoperational efficiency of edge computing services to increase

12 Wireless Communications and Mobile Computing

resource utilization to achieve a rapid deployment of VNFand to meet service on-demand requests The combinationof resource estimation task scheduling and lightweight VNFconfiguration design provides an integrated solution that cansatisfy the service on-demand requests for 5G networks Thesimulation results showed that the proposedGAFmechanismoutperforms the other scheduling algorithmsMeanwhile thecomparison revealed that using the lightweight virtualizationtechnology in edge gateways ismore efficient and competitivethan using traditional VMs

In the future we plan to study the multiqueue taskscheduling problem for edge computing the possible coop-eration between edge gateway and cloud servers and the useof SD-WAN technology to achieve a seamless operation of thecloud-edge network

Data Availability

The data used to support the findings of this study areincluded within the article

Disclosure

The funder had no role in the study design data collectionand analysis decision to publish or preparation of thismanuscript

Conflicts of Interest

The authors declare no conflicts of interest regarding thepublication of this paper

Acknowledgments

The work described in this paper was supported in part bythe Ministry of Science and Technology of the Republic ofChina (Grants nos 104-2221-E-008-039 105-2221-E-008-071107-2623-E-008-002 and 107-2636-E-003-001)

References

[1] J Matias J Garay N Toledo J Unzilla and E Jacob ldquoTowardan SDN-enabled NFV architecturerdquo IEEE CommunicationsMagazine vol 53 no 4 pp 187ndash193 2015

[2] Q Duan N Ansari and M Toy ldquoSoftware-defined networkvirtualization An architectural framework for integrating SDNand NFV for service provisioning in future networksrdquo IEEENetwork vol 30 no 5 pp 10ndash16 2016

[3] TSI ldquoNetwork Functions Virtualization (NFV) architecturalframeworkrdquo[Online] Available httpswwwetsiorgdeliveretsi gsNFV001 099002010101 60gs NFV002v010101ppdf

[4] ONF ldquoSDN Architecturerdquo [Online] Available httpwwwopennetworkingorgimagesstoriesdownloadssdn-resourcestechnical-reportsTR SDN ARCH 10 06062014pdf

[5] A Basta A Blenk K Hoffmann H J Morper M Hoffmannand W Kellerer ldquoTowards a Cost Optimal Design for a 5GMobile Core Network Based on SDN and NFVrdquo IEEE Trans-actions on Network and Service Management vol 14 no 4 pp1061ndash1075 2017

[6] F Z Yousaf M Bredel S Schaller and F Schneider ldquoNFV andSDN-Key technology enablers for 5G networksrdquo IEEE Journal

on Selected Areas in Communications vol 35 no 11 pp 2468ndash2478 2017

[7] A C Baktir A Ozgovde and C Ersoy ldquoHow Can EdgeComputing Benefit FromSoftware-DefinedNetworking A Sur-vey Use Cases and Future Directionsrdquo EEE CommunicationsSurveys amp Tutorials vol 19 no 4 pp 2359ndash2391 2017

[8] Guangshun Li Jiping Wang Junhua Wu and Jianrong SongldquoData Processing Delay Optimization in Mobile Edge Comput-ingrdquoWireless Communications andMobile Computing vol 2018Article ID 6897523 9 pages 2018

[9] Zhimin Wang Qinglin Zhao Fangxin Xu Hongning Daiand Yujun Zhang ldquoDetection Performance of Packet Arrivalunder Downclocking for Mobile Edge Computingrdquo WirelessCommunications and Mobile Computing vol 2018 Article ID9641712 7 pages 2018

[10] W Yu F Liang X He et al ldquoA Survey on the Edge Computingfor the Internet of Thingsrdquo IEEE Access vol 6 pp 6900ndash69192018

[11] P Mach and Z Becvar ldquoMobile Edge Computing A Survey onArchitecture and Computation Offloadingrdquo IEEE Communica-tions Surveys amp Tutorials vol 19 no 3 pp 1628ndash1656 2017

[12] N Abbas Y Zhang A Taherkordi and T Skeie ldquoMobile EdgeComputing A Surveyrdquo IEEE Internet of Things Journal vol 5no 1 pp 450ndash465 2018

[13] Guangshun Li Jianrong Song Junhua Wu and Jiping WangldquoMethod of Resource Estimation Based on QoS in Edge Com-putingrdquo Wireless Communications and Mobile Computing vol2018 Article ID 7308913 9 pages 2018

[14] D Happ and A Wolisz ldquoTowards gateway to Cloud offloadingin IoT publishsubscribe systemsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 101ndash106 Valencia Spain May 2017

[15] G Premsankar M Di Francesco and T Taleb ldquoEdge Comput-ing for the Internet of Things A Case Studyrdquo IEEE Internet ofThings Journal vol 5 no 2 pp 1275ndash1284 2018

[16] L Zhao W Sun Y Shi and J Liu ldquoOptimal Placementof Cloudlets for Access Delay Minimization in SDN-BasedInternet of Things Networksrdquo IEEE Internet of Things Journalvol 5 no 2 pp 1334ndash1344 2018

[17] SWang X Zhang Y Zhang LWang J Yang andWWang ldquoASurvey onMobile Edge Networks Convergence of ComputingCaching and Communicationsrdquo IEEE Access vol 5 pp 6757ndash6779 2017

[18] Y Liu J E Fieldsend and G Min ldquoA Framework of FogComputing ArchitectureChallenges and Optimizationrdquo IEEEAccess vol 5 pp 25445ndash25454 2017

[19] ETSI Multi-access Edge Computing [Online] Availablehttpwwwetsiorgtechnologies-clusterstechnologiesmulti-access-edge-computing

[20] OpenFog Consortium [Online] Available httpswwwopen-fogconsortiumorg

[21] Y Mao C You J Zhang K Huang and K B Letaief ldquoA SurveyonMobile Edge ComputingThe Communication PerspectiverdquoIEEE Communications Surveys amp Tutorials vol 19 no 4 pp2322ndash2358 2017

[22] Cisco ldquoThe Internet of Things How the Next Evolution ofthe Internet Is Changing Everything white paperrdquo [Online]Available httpswwwciscocomcdamen usIoT IBSG0411FINALpdf

[23] IDC ldquoFutureScape Worldwide Internet of Things 2017 Pre-dictionsrdquo November 2016 [Online] Available httpswwwidccomgetdocjspcontainerId=US41910716

6 Wireless Communications and Mobile Computing

Binaries Libraries

Host OS

Hardware Resource

VNF

Docker Engine

ResourceEstimation Scheduler VNF

Configuration

VN

F Manager

Container images

R

R

R

R

R

Edge Computing Node VNFs Resource PoolService Request

VNF

VNF

PublicRegistry

PrivateRegistry

SDN Controller

Figure 5 Service on-demand edge computing model

119871 =

1205881 minus 120588 minus

(119873 + 1) 120588119873+1

1 minus 120588119873+1 120588 = 11198732 120588 = 1

(3)

119871119902 = L minus (1 minus 1198750) (4)

119882 = 119871120582 (1 minus 119875119899)

(5)

119882119902 =119871119902120582

(6)

The results reveal that the waiting time for the userrequest can be reduced in two ways namely by reducing thenumber of service requests queued in the edge gateway andby increasing the processing speed of task scheduling in theedge computing gateway A large number of tasks must becompleted within a short period to achieve an efficient edgecomputing

4 Design of the Edge ComputingService Model

How to construct an elastic and cost-effective edge com-puting service model improve management efficiency meetuser service requests achieve centralized management anddevelop flexible configuration service models are the currentresearch trends in the development of a 5G SDNNFV net-work This paper proposes a gateway-based edge computingservicemodel (Figure 5) to improve the operational efficiencyof the edge computing node to accelerate the processingof user service requests and to increase the utilizationefficiency of a limited number of computing resources In thismodel when different user requests enter the edge gatewaythis gateway determines whether the requested services canbe processed or not If the edge gateway itself lacks thecomputing capacity or resources then the controller forwardsthe service request to the cloud to reduce the data processinglatency

The proposed edge computing service model can bedivided into resource estimation scheduler and lightweightVNF configuration Figure 6 presents the flowchart of theoperations in the proposed edge computing service modeland these three parts are further described in the followingsections

Resource estimation checkswhether the edge gateway hasa sufficient amount of computing resources to provide edgecomputing services For the user request set R the resourceallocation must abide by the following rules which are alsoused by the edge gateway

(1) For a single service request Ri any resource of Vi(eg CPU memory and disk) is less than the totalresources of P forall119877119862119894ltPC forall119877119872119894ltPM and forall119877119863119894ltPD119877119862119894 119877119872119894 and 119877119863119894 denote the CPU memory and diskspace requests sent to Ri respectively

(2) The sum of computing resources that are allocated tothe VNF in the gateway is less than the total numberof resources of the physical machine P 119881119862119894 lt PC 119881119872119894lt PM 119881119863119894 lt PD 119881119862119894 119881119872119894 and 119881119863119894 denote the sum ofCPU memory and disk space resources allocated toVi respectively

As shown in Figure 6 after receiving the user servicerequest the edge gateway checks whether a sufficient amountof computing resources is available to satisfy such requestThe following situations may be encountered in this case

(1) If the edge gateway has a sufficient amount ofresources then the user service request is processedthrough the scheduler and queued in the systembased on the result of the task scheduling algorithm

(2) If the edge gateway does not have a sufficient amountof available resources then the user service request isdirectly transferred to the cloud instead

Task scheduling aims to increase the operational effi-ciency of the edge gateway Given that one service requestdiffers from another task scheduling examines how the

Wireless Communications and Mobile Computing 7

Forward Ri to Cloud DC

Start

Download from private storage

RCilt(PCiminusVCi)

RMilt(PMiminusVMi)

RDilt(PDiminusVDi)

No

Yes

Yes

Yes

Resource Estimation for Ri

VNF allocation

Download from public storage

End

No

No

No

Yes

Task Scheduling Docker Image is exist

Riconfigured to

the scheduler No

PCi = (PCiminusRCi)PMi = (PMiminusRMi)PDi =(PDiminusRDi)

Service on-demand VNF Matching

Yes

Figure 6 Edge computing service model flowchart

edge gateway can meet the requirements of different servicerequests and accomplish task scheduling in the system As itsprimary purpose scheduling attempts to reduce the amountof time spent on dealing with the most demanding servicerequirements asmuch as possible For this purpose this paperconstructs the Greedy Available Fit (GAF) task schedulingmechanism to enhance the operational efficiency of edgecomputing services

Assume that each service request Ri configures a VNFvirtualization service resource Vi ti denotes the processingtime of the ith service request di denotes its deadline andj denotes its completion time in the system The Ri in thesystem must be completed at schedule and before the basictime limit Otherwise this service request must be forwardeddirectly to the cloud As each service request arrives at adifferent time the time required for the operation processingand the deadline time also differ In this case deadline isselected as a priority parameter for the task scheduling Thedeadline can be used to define the processing priority forservices that is the deadline for those services that requirereal-time processing may be set according to their processingand precedence requirements on different VNFs This paperaims to insert as many tasks as possible into the schedulebefore completing the largest task Ri of the deadline Toaccomplish this objective a task with the maximum deadlineis selected as the baseline and the remaining tasks aresequentially inserted into the queue based on their processing

time119873[119894 119895] denotes the maximumnumber of j tasks selectedfrom the front i tasks and can be formulated as

119873[119894 119895] = max

119873[119894 minus 1 119895]

119873 [119894 minus 1 119895 minus 119905119894] + 1 119894119891 119895 le 119889119894119873[119894 minus 1 119895 minus 119905119894] 119894119891 119895 gt 119889119894

(7)

Equation (8) shows the initial condition of119873[119894 119895]

119873[1 119895] =

minusinfin 119894119891 119895 = 11990511 119894119891 119895 = 1199051 le 11988910 119894119891 119895 = 1199051 gt 1198891

(8)

Assume that there is no time gap betweenRi and119877119894+1 thatR1 starts from time 0 and that N = 1 If t1 le d1 then N = 0because the deadline is exceeded

Thewhole decision process is summarized inAlgorithm 1which starts with an empty schedule and inserts the availabletasks into this schedule in three different cases In case 1 if thecurrent time step j exceeds the deadline of Ri then N[i j] =N[iminus1 jminusti] In case (2) if there is not enough time to finishrequest Ri by the current time j then N[i j] = N[iminus1 j] Incase (3) if time j does not exceed the deadline of Ri and thereis enough time to finish request Ri by time j then N[i j] =max(N[iminus1 j] N[iminus1 jminusti ] + 1)

8 Wireless Communications and Mobile Computing

Input 119877 = 1198771 1198772 119877119899 ti di jOutput119873[119894 119895](1) Start(2) Set the largest task di as the baseline of the scheduler(3) for 119894 larr997888 1 to n initialize the maximum N at time 0 to be zero for each service

request in queue(4) 119873[119894 0] = 0(5) for 119895 larr997888 1 to 119889119899 determine the maximum N obtained by R1 at each time step(6) if ((j == t1) and (j lt= d1))(7) 119873[1 119895] = 1(8) else(9) 119873[1 119895] =119873[1 119895 minus 1](10) for 119894 larr997888 2 to n(11) for 119895 larr997888 1 to 119889119894(12) if (j gt di) case (1) the current time step j already exceeds 119877119894rsquos deadline(13) then119873[119894 119895] = 119873[119894 minus 1 119895 minus 119905119894](14) else if (j lt ti) case (2) there is not enough time to finish request Ri by the

current time j(15) then119873[119894 119895] = 119873[119894 minus 1 119895](16) else case (3) time j does not exceed 119877119894rsquos deadline and there is enough time

to finish request 119877119894 by time j(17) then119873[119894 119895] = max(119873[119894 minus 1 119895]119873[119894 minus 1 119895 minus 119905119894] + 1)(18) Schedule Completed(19) End

Algorithm 1 Greedy Available Fit algorithm

Virtualization technology can be applied on the CPE inmany ways such as by using the VM container or VM inte-gration of container However a traditional VM consumesmany system resources and cannot meet the requirementsfor light weight and service on-demand deployment In thiscase this research adopts container technology instead Acontainer handles only one service request at a time and stopsits operation after completing the delivery of a service TheVNF manager on the gateway is responsible for configuringand allocating each VNF The VNF template image that isrequired by different services is placed in the VNF resourcepool (with Docker Hub as the default other public or privateregistrations can also be specified) The edge gateway notonly controls the amount of gateway resources used byeach container but also manages several resources such asCPU and RAM to ensure that the container can obtain therequired resources without affecting the performance of theother executing containers on the edge gateway

Linux containers adopt a hierarchical structure to rapidlydeploy VNFs and to manage NFV flexible scheduling Theunderlying structure uses the file archiving mechanism ofDocker namely the advanced multilayered unification filesystem to incorporatemany different VNF imagesThe layersare stacked up andwhen someVNF service functions need tobe accessed the container retrieves the VNF images throughthe VNFmanager and uses these images directly To integratethe open resources of Docker Hub and to expand the privatewarehouse of the VNF service manager the system can storethe image files obtained from the public warehouse in aprivate warehouse for future use The judgment mechanismof the VNF service manager is described as follows

(1) If the image requested by the host is already in thelocal container then the VNF image can be taken outdirectly by the Docker daemon of the VNF servicemanager No complex configurations are required

(2) If the VNF image to be used by the host is not in thelocal container then the VNF manager connects tothe common warehouse Docker Hub on the Internetto find and store the desired service image file ina private warehouse via pushpull Afterward theacquired VNF image is controlled by the Dockerdaemon

By using the OS kernel for resource isolation the con-tainer technology does not need to rely on a virtual softwarelayer and does not require the VM to install a guest OSTherefore the capacity of the image file is much larger thanthat of the virtual machine The image file is small and canbe rapidly deployed through network transmission therebysaving network resources By using Linux container technol-ogy we can configure various VNF images to be executed onthe same edge gateway thereby replacing the virtual machineand achieving the goal of lightweight virtualization

5 Experiment

An experiment is conducted to evaluate the proposed taskscheduling algorithmand to test its performance in deployinglightweight VNF The experiment environment is shown inFigure 7 A simulation is initially performed to evaluate thetask scheduling performance of the algorithm by changingthe number of service requirements The service request is

Wireless Communications and Mobile Computing 9

Edge Network

Internet

Edge Gateway

DataCenter

Cloud Broker

Docker HubDocker Hub

Figure 7 The experiment environment

characterized as a Poisson process There we showed that theaverage number of service requests in the system is given by120582120583 where 120582 is the average arrival rate and 120583 is the averageservice rate The input parameter 120582120583 is 045 First comefirst service (FCFS) priority task scheduling [34] RR [37]and GBA [36] are then compared with the proposed GAFmechanism

The effect of the service requests on the average waitingtime average response time and task scheduling of differentscheduling mechanisms is evaluated in the simulation FCFSis based on the service requests that enter the edge gatewayqueuing system and schedules the tasks according to thesequence of these requests Priority task scheduling is anunfair scheduling algorithm where the service requests aresorted according to their priority and where the high-prioritytasks are performedfirstThose tasks having the same priorityare sorted by using the FIFO scheduling mechanism TheGBA sorts the tasks based on the completion time of theservice requests in queuing systemTheRR algorithm is basedon the conventional RR scheduling performed in the processscheduling RR scheduling fairly allocates the computingresources to those tasks with the same priority by using thetime slicing approach (time quantum)

Figure 8 shows the experiment result for average waitingtime where the x-axis refers to the number of service requestsand the y main axis refers to the average waiting time thatis the time a user spends to complete the service schedulingand to determine the resource allocation after receiving aservice request and a successful reply If the demand revertsto an errortimeout or if the edge gateway does not haveenough computing resources to support such demand thenthe sample is excluded from the calculation of the averagewaiting time As can be seen in the experimental resultthe RR algorithm obtains the longest average waiting timebecause of its application of time slice rules in order for eachtask to be processed within a fixed amount of time The timequantum affects the overall performance of the operation

0100200300400500600

10 20 30 40 50 60 70 80 90 100

Aver

age W

aitin

g Ti

me (

ms)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 8 The experiment result for average waiting time

Having a very long time quantum leads to a very long waitingtime while having a very short time quantum results intask schedule conversion a poor execution efficiency and anextended waiting time For FCFS given that the time spentdiffers across each task the waiting time for the next taskmust be determined based on the schedule of the previoustask Therefore if the last task schedule is too long thesystem cannot quickly process the subsequent task schedulesthereby affecting the overall task scheduling efficiency Giventhat its average waiting time is relatively shorter than thatof the FCFS mechanism priority task scheduling can satisfythe task requirements and flexibly perform the schedulingHowever prioritizing a high-priority program will delay allof the low-priority requests thereby creating an indefinitesituation in the low-priority program and extending thewaiting time GBA is based on the earliest completion time ofthe taskWhen the service demand is low the average waitingtime of the GBA algorithm is similar to that of priority taskscheduling However when the service demand increases theaverage waiting time of the GBAmechanism becomes longerthan that of priority task scheduling Compared with FCFS

10 Wireless Communications and Mobile Computing

050

100150200250300350

10 20 30 40 50 60 70 80 90 100

Ave

rage

Res

pons

e Tim

e (m

s)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 9 The experiment result for average response time

01020304050607080

10 20 30 40 50 60 70 80 90 100

Num

ber o

f Sch

edul

ed T

asks

Number of Service RequestsGAFFCFSPriority

GBARR

Figure 10 The experiment result for task scheduling efficiency

priority task scheduling RR and GBA the proposed GAFalgorithm selects the largest deadline task as the baseline andthen sequentially inserts the remaining tasks into the queuebased on their processing timeTherefore the GAF algorithmobtains the shortest average waiting time

Figure 9 shows the results for average response time Asexpected the RR algorithm outperforms all other approachesin terms of average response time because this algorithm isespecially designed for time sharing Using a fair strategy toallocate VNF resources ensures that each service request hasa fixed time If the service processing time exceeds the timequantum allocated by the system then the sample is excludedfrom the calculation of the average response time Given thatFCFS is processed in sequence the average response timewill be affected by the time spent by the previous 119877119894minus1 taskin the scheduler For example if the previous service requesthas a very long task schedule then the average responsetime increases Given that the proposed GAF algorithmsequentially inserts tasks into the queue its average responsetime is slightly longer than that of FCFS and GBA Althoughthe priority algorithm can sort the tasks according to thepriorities of the service requests the low-priority tasks areeasily delayed thereby increasing the average response time

Figure 10 shows the task scheduling efficiency of thecompared algorithms If the service request number is lessthan 60 then each algorithm can complete the task schedul-ing However when this number is exceeded the number

of tasks that can be scheduled varies across each algorithmGAF outperforms all the other algorithms as it ranks thelargest deadline task at the end and then sequentially sortsthe remaining tasks according to their processing time In thisway this algorithm ensures that most tasks will be completedwithin a minimum processing time Those service requeststhat cannot be scheduled will bypass the gateway and will bedirectly forwarded to the cloud data center When the servicerequest number is 80 GAF can complete 68 of the taskscheduling Meanwhile GBA performs the scheduling basedon the earliest task completion time Although GAF andGBAcan schedule the same number of tasks in the experimentalenvironment the latter has a longer average waiting timecompared with the former GAF also outperforms FCFCand priority task scheduling by 96 and 62 in terms oftask scheduling efficiency respectively because this approachranks the largest deadline task at the end and then sorts theremaining tasks according to their required processing timeGAF also outperforms theRRmechanismby 70because thelatter applies the time-sharing rule to minimize the numberof scheduled tasks Based on these results the GAF algorithmcan schedule more tasks in the edge gateway within a shortaverage waiting time and improve the operational efficiencyof edge computing services

In addition to task scheduling the VNF configuration isanother main factor that affects the operational efficiency ofthe edge computing service model This paper utilizes theDocker technology to achieve an agile deployment of VNFsSuch technology can provide a centralized management andflexible configuration of VNF services Depending on theservice requests of each user the deployment of differentVNFs can meet the application requirements of serviceson-demand The VNF image of Docker can be mainlydivided into four categories namely system images (egUbuntu CentOS and Debian) tool images (eg GolangFlask and Tomcat) service images (eg MySQL MongoDBand RabbitMQ) and application images (eg WordPressand DRUPAL) To test the practical deployment efficiency oflightweight VNF for the edge gateway a pull performancetest is performed for each type of Docker VNF image Theedge gateway is simulated by using a standalone PC with 34GHz dual-core CPU 16GBmemory and 5400 rpmHD500GThe experimental network bandwidth is set to 100 Mbps Tocompare the pull performance with the local edge gateway atest VM in the Google Cloud platform is used to simulate thecloud network broker Table 1 presents the result of the pullperformance test

As shown in Table 1 each VNF image size has differentphysical gateway and cloud network broker costs due to thelimitations in the network bandwidth A larger VNF imagetakes a longer time to be downloaded from Docker Hub Asexpected the cloud network broker has a better computingarchitecture than the local edge gateway because the locationof the edge gateway affects the performance of the VNFdeployment Although the edge computing architecture ofthe cloud network broker deploys VNF more efficiently thanthe local gateway the user service requestsmust still be sent tothe Internet for processing and this methodmakes it difficultto demonstrate the technical advantages of edge computing

Wireless Communications and Mobile Computing 11

Table 1 Pull performance test results for different VNFs

Image Types VNFs Image Size (MB) Physical GatewayCost (seconds)

Cloud NetworkBroker Cost(seconds)

System imagesubuntu 112 22095 7918centos 207 24319 8926debian 207 21635 5646

Tool imagesgolang 779 86793 1851flask 458 6222 1633

prometheus 112 2414 4737

Service images

httpd 177 20753 1031nginx 109 13216 4778bind 337 8438 15105squid 215 150031 9973mysql 374 24852 10449

mongodb 503 61347 17116

Application images wordpress 407 32866 13213drupal 452 8139 14405

11095

1593

0

20

40

60

80

100

120

140

160

180

Docker VM

Tim

e (s)

Figure 11 VNF boot time test for the VM and container

In the VNF boot time test we set the image to Ubuntu-16-04-x64 VM and Docker are used to perform the boottime test on the edge gateway We test the time of openingfive VMs and Docker containers VMware is used as the VMhypervisor As shown in Figure 11 the five VMs are openedin 1593 seconds while the Docker containers are openedin 11095 seconds In sum the VMs have a much longerdeployment time than Docker More time can be saved byusingDocker in the edge gateway that requires a large amountof VNFs

To test the average VNF boot time we boot a VM waitfor its activation shut it down and repeat the same steps for14 other VMs The test results are shown in Figure 12 Whenusing VMware the average time to boot the VMs is about3186 seconds However when using Docker these VMs arebooted in only 22 seconds In sum the Docker containerperforms approximately 15 times faster than the VM

The results of the VNF deployment test highlight thesignificant advantages of lightweight containers over tradi-tional virtualizationmethodsThese containers can be started

2219

3186

0

5

10

15

20

25

30

35

Docker VM

Tim

e (s)

Figure 12 Average VNF boot time test for the VM and container

within a few seconds because they adopt a common host OStherebymaking it unnecessary to execute the guestOS in eachcontainer The container also does not need to wait for theOS to start up thereby saving a few seconds Fast start is farfaster than traditional VMs Meanwhile given the design oflightweight VNF images and its highly efficient virtualizationthe application-centric virtualization technology Docker canautomatically obtain container images from the Docker Hubfor flexible deployment and management and can meet theapplication requirements of the edge network for the serviceon-demand of VNF deployment

6 Conclusions

This paper proposed a gateway-based edge computing servicemodel and a GAF task scheduling mechanism that allowsthe edge gateway to schedule more tasks within a short aver-age waiting time The resource estimation and lightweightVNF configuration technologies are used to improve theoperational efficiency of edge computing services to increase

12 Wireless Communications and Mobile Computing

resource utilization to achieve a rapid deployment of VNFand to meet service on-demand requests The combinationof resource estimation task scheduling and lightweight VNFconfiguration design provides an integrated solution that cansatisfy the service on-demand requests for 5G networks Thesimulation results showed that the proposedGAFmechanismoutperforms the other scheduling algorithmsMeanwhile thecomparison revealed that using the lightweight virtualizationtechnology in edge gateways ismore efficient and competitivethan using traditional VMs

In the future we plan to study the multiqueue taskscheduling problem for edge computing the possible coop-eration between edge gateway and cloud servers and the useof SD-WAN technology to achieve a seamless operation of thecloud-edge network

Data Availability

The data used to support the findings of this study areincluded within the article

Disclosure

The funder had no role in the study design data collectionand analysis decision to publish or preparation of thismanuscript

Conflicts of Interest

The authors declare no conflicts of interest regarding thepublication of this paper

Acknowledgments

The work described in this paper was supported in part bythe Ministry of Science and Technology of the Republic ofChina (Grants nos 104-2221-E-008-039 105-2221-E-008-071107-2623-E-008-002 and 107-2636-E-003-001)

References

[1] J Matias J Garay N Toledo J Unzilla and E Jacob ldquoTowardan SDN-enabled NFV architecturerdquo IEEE CommunicationsMagazine vol 53 no 4 pp 187ndash193 2015

[2] Q Duan N Ansari and M Toy ldquoSoftware-defined networkvirtualization An architectural framework for integrating SDNand NFV for service provisioning in future networksrdquo IEEENetwork vol 30 no 5 pp 10ndash16 2016

[3] TSI ldquoNetwork Functions Virtualization (NFV) architecturalframeworkrdquo[Online] Available httpswwwetsiorgdeliveretsi gsNFV001 099002010101 60gs NFV002v010101ppdf

[4] ONF ldquoSDN Architecturerdquo [Online] Available httpwwwopennetworkingorgimagesstoriesdownloadssdn-resourcestechnical-reportsTR SDN ARCH 10 06062014pdf

[5] A Basta A Blenk K Hoffmann H J Morper M Hoffmannand W Kellerer ldquoTowards a Cost Optimal Design for a 5GMobile Core Network Based on SDN and NFVrdquo IEEE Trans-actions on Network and Service Management vol 14 no 4 pp1061ndash1075 2017

[6] F Z Yousaf M Bredel S Schaller and F Schneider ldquoNFV andSDN-Key technology enablers for 5G networksrdquo IEEE Journal

on Selected Areas in Communications vol 35 no 11 pp 2468ndash2478 2017

[7] A C Baktir A Ozgovde and C Ersoy ldquoHow Can EdgeComputing Benefit FromSoftware-DefinedNetworking A Sur-vey Use Cases and Future Directionsrdquo EEE CommunicationsSurveys amp Tutorials vol 19 no 4 pp 2359ndash2391 2017

[8] Guangshun Li Jiping Wang Junhua Wu and Jianrong SongldquoData Processing Delay Optimization in Mobile Edge Comput-ingrdquoWireless Communications andMobile Computing vol 2018Article ID 6897523 9 pages 2018

[9] Zhimin Wang Qinglin Zhao Fangxin Xu Hongning Daiand Yujun Zhang ldquoDetection Performance of Packet Arrivalunder Downclocking for Mobile Edge Computingrdquo WirelessCommunications and Mobile Computing vol 2018 Article ID9641712 7 pages 2018

[10] W Yu F Liang X He et al ldquoA Survey on the Edge Computingfor the Internet of Thingsrdquo IEEE Access vol 6 pp 6900ndash69192018

[11] P Mach and Z Becvar ldquoMobile Edge Computing A Survey onArchitecture and Computation Offloadingrdquo IEEE Communica-tions Surveys amp Tutorials vol 19 no 3 pp 1628ndash1656 2017

[12] N Abbas Y Zhang A Taherkordi and T Skeie ldquoMobile EdgeComputing A Surveyrdquo IEEE Internet of Things Journal vol 5no 1 pp 450ndash465 2018

[13] Guangshun Li Jianrong Song Junhua Wu and Jiping WangldquoMethod of Resource Estimation Based on QoS in Edge Com-putingrdquo Wireless Communications and Mobile Computing vol2018 Article ID 7308913 9 pages 2018

[14] D Happ and A Wolisz ldquoTowards gateway to Cloud offloadingin IoT publishsubscribe systemsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 101ndash106 Valencia Spain May 2017

[15] G Premsankar M Di Francesco and T Taleb ldquoEdge Comput-ing for the Internet of Things A Case Studyrdquo IEEE Internet ofThings Journal vol 5 no 2 pp 1275ndash1284 2018

[16] L Zhao W Sun Y Shi and J Liu ldquoOptimal Placementof Cloudlets for Access Delay Minimization in SDN-BasedInternet of Things Networksrdquo IEEE Internet of Things Journalvol 5 no 2 pp 1334ndash1344 2018

[17] SWang X Zhang Y Zhang LWang J Yang andWWang ldquoASurvey onMobile Edge Networks Convergence of ComputingCaching and Communicationsrdquo IEEE Access vol 5 pp 6757ndash6779 2017

[18] Y Liu J E Fieldsend and G Min ldquoA Framework of FogComputing ArchitectureChallenges and Optimizationrdquo IEEEAccess vol 5 pp 25445ndash25454 2017

[19] ETSI Multi-access Edge Computing [Online] Availablehttpwwwetsiorgtechnologies-clusterstechnologiesmulti-access-edge-computing

[20] OpenFog Consortium [Online] Available httpswwwopen-fogconsortiumorg

[21] Y Mao C You J Zhang K Huang and K B Letaief ldquoA SurveyonMobile Edge ComputingThe Communication PerspectiverdquoIEEE Communications Surveys amp Tutorials vol 19 no 4 pp2322ndash2358 2017

[22] Cisco ldquoThe Internet of Things How the Next Evolution ofthe Internet Is Changing Everything white paperrdquo [Online]Available httpswwwciscocomcdamen usIoT IBSG0411FINALpdf

[23] IDC ldquoFutureScape Worldwide Internet of Things 2017 Pre-dictionsrdquo November 2016 [Online] Available httpswwwidccomgetdocjspcontainerId=US41910716

Wireless Communications and Mobile Computing 7

Forward Ri to Cloud DC

Start

Download from private storage

RCilt(PCiminusVCi)

RMilt(PMiminusVMi)

RDilt(PDiminusVDi)

No

Yes

Yes

Yes

Resource Estimation for Ri

VNF allocation

Download from public storage

End

No

No

No

Yes

Task Scheduling Docker Image is exist

Riconfigured to

the scheduler No

PCi = (PCiminusRCi)PMi = (PMiminusRMi)PDi =(PDiminusRDi)

Service on-demand VNF Matching

Yes

Figure 6 Edge computing service model flowchart

edge gateway can meet the requirements of different servicerequests and accomplish task scheduling in the system As itsprimary purpose scheduling attempts to reduce the amountof time spent on dealing with the most demanding servicerequirements asmuch as possible For this purpose this paperconstructs the Greedy Available Fit (GAF) task schedulingmechanism to enhance the operational efficiency of edgecomputing services

Assume that each service request Ri configures a VNFvirtualization service resource Vi ti denotes the processingtime of the ith service request di denotes its deadline andj denotes its completion time in the system The Ri in thesystem must be completed at schedule and before the basictime limit Otherwise this service request must be forwardeddirectly to the cloud As each service request arrives at adifferent time the time required for the operation processingand the deadline time also differ In this case deadline isselected as a priority parameter for the task scheduling Thedeadline can be used to define the processing priority forservices that is the deadline for those services that requirereal-time processing may be set according to their processingand precedence requirements on different VNFs This paperaims to insert as many tasks as possible into the schedulebefore completing the largest task Ri of the deadline Toaccomplish this objective a task with the maximum deadlineis selected as the baseline and the remaining tasks aresequentially inserted into the queue based on their processing

time119873[119894 119895] denotes the maximumnumber of j tasks selectedfrom the front i tasks and can be formulated as

119873[119894 119895] = max

119873[119894 minus 1 119895]

119873 [119894 minus 1 119895 minus 119905119894] + 1 119894119891 119895 le 119889119894119873[119894 minus 1 119895 minus 119905119894] 119894119891 119895 gt 119889119894

(7)

Equation (8) shows the initial condition of119873[119894 119895]

119873[1 119895] =

minusinfin 119894119891 119895 = 11990511 119894119891 119895 = 1199051 le 11988910 119894119891 119895 = 1199051 gt 1198891

(8)

Assume that there is no time gap betweenRi and119877119894+1 thatR1 starts from time 0 and that N = 1 If t1 le d1 then N = 0because the deadline is exceeded

Thewhole decision process is summarized inAlgorithm 1which starts with an empty schedule and inserts the availabletasks into this schedule in three different cases In case 1 if thecurrent time step j exceeds the deadline of Ri then N[i j] =N[iminus1 jminusti] In case (2) if there is not enough time to finishrequest Ri by the current time j then N[i j] = N[iminus1 j] Incase (3) if time j does not exceed the deadline of Ri and thereis enough time to finish request Ri by time j then N[i j] =max(N[iminus1 j] N[iminus1 jminusti ] + 1)

8 Wireless Communications and Mobile Computing

Input 119877 = 1198771 1198772 119877119899 ti di jOutput119873[119894 119895](1) Start(2) Set the largest task di as the baseline of the scheduler(3) for 119894 larr997888 1 to n initialize the maximum N at time 0 to be zero for each service

request in queue(4) 119873[119894 0] = 0(5) for 119895 larr997888 1 to 119889119899 determine the maximum N obtained by R1 at each time step(6) if ((j == t1) and (j lt= d1))(7) 119873[1 119895] = 1(8) else(9) 119873[1 119895] =119873[1 119895 minus 1](10) for 119894 larr997888 2 to n(11) for 119895 larr997888 1 to 119889119894(12) if (j gt di) case (1) the current time step j already exceeds 119877119894rsquos deadline(13) then119873[119894 119895] = 119873[119894 minus 1 119895 minus 119905119894](14) else if (j lt ti) case (2) there is not enough time to finish request Ri by the

current time j(15) then119873[119894 119895] = 119873[119894 minus 1 119895](16) else case (3) time j does not exceed 119877119894rsquos deadline and there is enough time

to finish request 119877119894 by time j(17) then119873[119894 119895] = max(119873[119894 minus 1 119895]119873[119894 minus 1 119895 minus 119905119894] + 1)(18) Schedule Completed(19) End

Algorithm 1 Greedy Available Fit algorithm

Virtualization technology can be applied on the CPE inmany ways such as by using the VM container or VM inte-gration of container However a traditional VM consumesmany system resources and cannot meet the requirementsfor light weight and service on-demand deployment In thiscase this research adopts container technology instead Acontainer handles only one service request at a time and stopsits operation after completing the delivery of a service TheVNF manager on the gateway is responsible for configuringand allocating each VNF The VNF template image that isrequired by different services is placed in the VNF resourcepool (with Docker Hub as the default other public or privateregistrations can also be specified) The edge gateway notonly controls the amount of gateway resources used byeach container but also manages several resources such asCPU and RAM to ensure that the container can obtain therequired resources without affecting the performance of theother executing containers on the edge gateway

Linux containers adopt a hierarchical structure to rapidlydeploy VNFs and to manage NFV flexible scheduling Theunderlying structure uses the file archiving mechanism ofDocker namely the advanced multilayered unification filesystem to incorporatemany different VNF imagesThe layersare stacked up andwhen someVNF service functions need tobe accessed the container retrieves the VNF images throughthe VNFmanager and uses these images directly To integratethe open resources of Docker Hub and to expand the privatewarehouse of the VNF service manager the system can storethe image files obtained from the public warehouse in aprivate warehouse for future use The judgment mechanismof the VNF service manager is described as follows

(1) If the image requested by the host is already in thelocal container then the VNF image can be taken outdirectly by the Docker daemon of the VNF servicemanager No complex configurations are required

(2) If the VNF image to be used by the host is not in thelocal container then the VNF manager connects tothe common warehouse Docker Hub on the Internetto find and store the desired service image file ina private warehouse via pushpull Afterward theacquired VNF image is controlled by the Dockerdaemon

By using the OS kernel for resource isolation the con-tainer technology does not need to rely on a virtual softwarelayer and does not require the VM to install a guest OSTherefore the capacity of the image file is much larger thanthat of the virtual machine The image file is small and canbe rapidly deployed through network transmission therebysaving network resources By using Linux container technol-ogy we can configure various VNF images to be executed onthe same edge gateway thereby replacing the virtual machineand achieving the goal of lightweight virtualization

5 Experiment

An experiment is conducted to evaluate the proposed taskscheduling algorithmand to test its performance in deployinglightweight VNF The experiment environment is shown inFigure 7 A simulation is initially performed to evaluate thetask scheduling performance of the algorithm by changingthe number of service requirements The service request is

Wireless Communications and Mobile Computing 9

Edge Network

Internet

Edge Gateway

DataCenter

Cloud Broker

Docker HubDocker Hub

Figure 7 The experiment environment

characterized as a Poisson process There we showed that theaverage number of service requests in the system is given by120582120583 where 120582 is the average arrival rate and 120583 is the averageservice rate The input parameter 120582120583 is 045 First comefirst service (FCFS) priority task scheduling [34] RR [37]and GBA [36] are then compared with the proposed GAFmechanism

The effect of the service requests on the average waitingtime average response time and task scheduling of differentscheduling mechanisms is evaluated in the simulation FCFSis based on the service requests that enter the edge gatewayqueuing system and schedules the tasks according to thesequence of these requests Priority task scheduling is anunfair scheduling algorithm where the service requests aresorted according to their priority and where the high-prioritytasks are performedfirstThose tasks having the same priorityare sorted by using the FIFO scheduling mechanism TheGBA sorts the tasks based on the completion time of theservice requests in queuing systemTheRR algorithm is basedon the conventional RR scheduling performed in the processscheduling RR scheduling fairly allocates the computingresources to those tasks with the same priority by using thetime slicing approach (time quantum)

Figure 8 shows the experiment result for average waitingtime where the x-axis refers to the number of service requestsand the y main axis refers to the average waiting time thatis the time a user spends to complete the service schedulingand to determine the resource allocation after receiving aservice request and a successful reply If the demand revertsto an errortimeout or if the edge gateway does not haveenough computing resources to support such demand thenthe sample is excluded from the calculation of the averagewaiting time As can be seen in the experimental resultthe RR algorithm obtains the longest average waiting timebecause of its application of time slice rules in order for eachtask to be processed within a fixed amount of time The timequantum affects the overall performance of the operation

0100200300400500600

10 20 30 40 50 60 70 80 90 100

Aver

age W

aitin

g Ti

me (

ms)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 8 The experiment result for average waiting time

Having a very long time quantum leads to a very long waitingtime while having a very short time quantum results intask schedule conversion a poor execution efficiency and anextended waiting time For FCFS given that the time spentdiffers across each task the waiting time for the next taskmust be determined based on the schedule of the previoustask Therefore if the last task schedule is too long thesystem cannot quickly process the subsequent task schedulesthereby affecting the overall task scheduling efficiency Giventhat its average waiting time is relatively shorter than thatof the FCFS mechanism priority task scheduling can satisfythe task requirements and flexibly perform the schedulingHowever prioritizing a high-priority program will delay allof the low-priority requests thereby creating an indefinitesituation in the low-priority program and extending thewaiting time GBA is based on the earliest completion time ofthe taskWhen the service demand is low the average waitingtime of the GBA algorithm is similar to that of priority taskscheduling However when the service demand increases theaverage waiting time of the GBAmechanism becomes longerthan that of priority task scheduling Compared with FCFS

10 Wireless Communications and Mobile Computing

050

100150200250300350

10 20 30 40 50 60 70 80 90 100

Ave

rage

Res

pons

e Tim

e (m

s)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 9 The experiment result for average response time

01020304050607080

10 20 30 40 50 60 70 80 90 100

Num

ber o

f Sch

edul

ed T

asks

Number of Service RequestsGAFFCFSPriority

GBARR

Figure 10 The experiment result for task scheduling efficiency

priority task scheduling RR and GBA the proposed GAFalgorithm selects the largest deadline task as the baseline andthen sequentially inserts the remaining tasks into the queuebased on their processing timeTherefore the GAF algorithmobtains the shortest average waiting time

Figure 9 shows the results for average response time Asexpected the RR algorithm outperforms all other approachesin terms of average response time because this algorithm isespecially designed for time sharing Using a fair strategy toallocate VNF resources ensures that each service request hasa fixed time If the service processing time exceeds the timequantum allocated by the system then the sample is excludedfrom the calculation of the average response time Given thatFCFS is processed in sequence the average response timewill be affected by the time spent by the previous 119877119894minus1 taskin the scheduler For example if the previous service requesthas a very long task schedule then the average responsetime increases Given that the proposed GAF algorithmsequentially inserts tasks into the queue its average responsetime is slightly longer than that of FCFS and GBA Althoughthe priority algorithm can sort the tasks according to thepriorities of the service requests the low-priority tasks areeasily delayed thereby increasing the average response time

Figure 10 shows the task scheduling efficiency of thecompared algorithms If the service request number is lessthan 60 then each algorithm can complete the task schedul-ing However when this number is exceeded the number

of tasks that can be scheduled varies across each algorithmGAF outperforms all the other algorithms as it ranks thelargest deadline task at the end and then sequentially sortsthe remaining tasks according to their processing time In thisway this algorithm ensures that most tasks will be completedwithin a minimum processing time Those service requeststhat cannot be scheduled will bypass the gateway and will bedirectly forwarded to the cloud data center When the servicerequest number is 80 GAF can complete 68 of the taskscheduling Meanwhile GBA performs the scheduling basedon the earliest task completion time Although GAF andGBAcan schedule the same number of tasks in the experimentalenvironment the latter has a longer average waiting timecompared with the former GAF also outperforms FCFCand priority task scheduling by 96 and 62 in terms oftask scheduling efficiency respectively because this approachranks the largest deadline task at the end and then sorts theremaining tasks according to their required processing timeGAF also outperforms theRRmechanismby 70because thelatter applies the time-sharing rule to minimize the numberof scheduled tasks Based on these results the GAF algorithmcan schedule more tasks in the edge gateway within a shortaverage waiting time and improve the operational efficiencyof edge computing services

In addition to task scheduling the VNF configuration isanother main factor that affects the operational efficiency ofthe edge computing service model This paper utilizes theDocker technology to achieve an agile deployment of VNFsSuch technology can provide a centralized management andflexible configuration of VNF services Depending on theservice requests of each user the deployment of differentVNFs can meet the application requirements of serviceson-demand The VNF image of Docker can be mainlydivided into four categories namely system images (egUbuntu CentOS and Debian) tool images (eg GolangFlask and Tomcat) service images (eg MySQL MongoDBand RabbitMQ) and application images (eg WordPressand DRUPAL) To test the practical deployment efficiency oflightweight VNF for the edge gateway a pull performancetest is performed for each type of Docker VNF image Theedge gateway is simulated by using a standalone PC with 34GHz dual-core CPU 16GBmemory and 5400 rpmHD500GThe experimental network bandwidth is set to 100 Mbps Tocompare the pull performance with the local edge gateway atest VM in the Google Cloud platform is used to simulate thecloud network broker Table 1 presents the result of the pullperformance test

As shown in Table 1 each VNF image size has differentphysical gateway and cloud network broker costs due to thelimitations in the network bandwidth A larger VNF imagetakes a longer time to be downloaded from Docker Hub Asexpected the cloud network broker has a better computingarchitecture than the local edge gateway because the locationof the edge gateway affects the performance of the VNFdeployment Although the edge computing architecture ofthe cloud network broker deploys VNF more efficiently thanthe local gateway the user service requestsmust still be sent tothe Internet for processing and this methodmakes it difficultto demonstrate the technical advantages of edge computing

Wireless Communications and Mobile Computing 11

Table 1 Pull performance test results for different VNFs

Image Types VNFs Image Size (MB) Physical GatewayCost (seconds)

Cloud NetworkBroker Cost(seconds)

System imagesubuntu 112 22095 7918centos 207 24319 8926debian 207 21635 5646

Tool imagesgolang 779 86793 1851flask 458 6222 1633

prometheus 112 2414 4737

Service images

httpd 177 20753 1031nginx 109 13216 4778bind 337 8438 15105squid 215 150031 9973mysql 374 24852 10449

mongodb 503 61347 17116

Application images wordpress 407 32866 13213drupal 452 8139 14405

11095

1593

0

20

40

60

80

100

120

140

160

180

Docker VM

Tim

e (s)

Figure 11 VNF boot time test for the VM and container

In the VNF boot time test we set the image to Ubuntu-16-04-x64 VM and Docker are used to perform the boottime test on the edge gateway We test the time of openingfive VMs and Docker containers VMware is used as the VMhypervisor As shown in Figure 11 the five VMs are openedin 1593 seconds while the Docker containers are openedin 11095 seconds In sum the VMs have a much longerdeployment time than Docker More time can be saved byusingDocker in the edge gateway that requires a large amountof VNFs

To test the average VNF boot time we boot a VM waitfor its activation shut it down and repeat the same steps for14 other VMs The test results are shown in Figure 12 Whenusing VMware the average time to boot the VMs is about3186 seconds However when using Docker these VMs arebooted in only 22 seconds In sum the Docker containerperforms approximately 15 times faster than the VM

The results of the VNF deployment test highlight thesignificant advantages of lightweight containers over tradi-tional virtualizationmethodsThese containers can be started

2219

3186

0

5

10

15

20

25

30

35

Docker VM

Tim

e (s)

Figure 12 Average VNF boot time test for the VM and container

within a few seconds because they adopt a common host OStherebymaking it unnecessary to execute the guestOS in eachcontainer The container also does not need to wait for theOS to start up thereby saving a few seconds Fast start is farfaster than traditional VMs Meanwhile given the design oflightweight VNF images and its highly efficient virtualizationthe application-centric virtualization technology Docker canautomatically obtain container images from the Docker Hubfor flexible deployment and management and can meet theapplication requirements of the edge network for the serviceon-demand of VNF deployment

6 Conclusions

This paper proposed a gateway-based edge computing servicemodel and a GAF task scheduling mechanism that allowsthe edge gateway to schedule more tasks within a short aver-age waiting time The resource estimation and lightweightVNF configuration technologies are used to improve theoperational efficiency of edge computing services to increase

12 Wireless Communications and Mobile Computing

resource utilization to achieve a rapid deployment of VNFand to meet service on-demand requests The combinationof resource estimation task scheduling and lightweight VNFconfiguration design provides an integrated solution that cansatisfy the service on-demand requests for 5G networks Thesimulation results showed that the proposedGAFmechanismoutperforms the other scheduling algorithmsMeanwhile thecomparison revealed that using the lightweight virtualizationtechnology in edge gateways ismore efficient and competitivethan using traditional VMs

In the future we plan to study the multiqueue taskscheduling problem for edge computing the possible coop-eration between edge gateway and cloud servers and the useof SD-WAN technology to achieve a seamless operation of thecloud-edge network

Data Availability

The data used to support the findings of this study areincluded within the article

Disclosure

The funder had no role in the study design data collectionand analysis decision to publish or preparation of thismanuscript

Conflicts of Interest

The authors declare no conflicts of interest regarding thepublication of this paper

Acknowledgments

The work described in this paper was supported in part bythe Ministry of Science and Technology of the Republic ofChina (Grants nos 104-2221-E-008-039 105-2221-E-008-071107-2623-E-008-002 and 107-2636-E-003-001)

References

[1] J Matias J Garay N Toledo J Unzilla and E Jacob ldquoTowardan SDN-enabled NFV architecturerdquo IEEE CommunicationsMagazine vol 53 no 4 pp 187ndash193 2015

[2] Q Duan N Ansari and M Toy ldquoSoftware-defined networkvirtualization An architectural framework for integrating SDNand NFV for service provisioning in future networksrdquo IEEENetwork vol 30 no 5 pp 10ndash16 2016

[3] TSI ldquoNetwork Functions Virtualization (NFV) architecturalframeworkrdquo[Online] Available httpswwwetsiorgdeliveretsi gsNFV001 099002010101 60gs NFV002v010101ppdf

[4] ONF ldquoSDN Architecturerdquo [Online] Available httpwwwopennetworkingorgimagesstoriesdownloadssdn-resourcestechnical-reportsTR SDN ARCH 10 06062014pdf

[5] A Basta A Blenk K Hoffmann H J Morper M Hoffmannand W Kellerer ldquoTowards a Cost Optimal Design for a 5GMobile Core Network Based on SDN and NFVrdquo IEEE Trans-actions on Network and Service Management vol 14 no 4 pp1061ndash1075 2017

[6] F Z Yousaf M Bredel S Schaller and F Schneider ldquoNFV andSDN-Key technology enablers for 5G networksrdquo IEEE Journal

on Selected Areas in Communications vol 35 no 11 pp 2468ndash2478 2017

[7] A C Baktir A Ozgovde and C Ersoy ldquoHow Can EdgeComputing Benefit FromSoftware-DefinedNetworking A Sur-vey Use Cases and Future Directionsrdquo EEE CommunicationsSurveys amp Tutorials vol 19 no 4 pp 2359ndash2391 2017

[8] Guangshun Li Jiping Wang Junhua Wu and Jianrong SongldquoData Processing Delay Optimization in Mobile Edge Comput-ingrdquoWireless Communications andMobile Computing vol 2018Article ID 6897523 9 pages 2018

[9] Zhimin Wang Qinglin Zhao Fangxin Xu Hongning Daiand Yujun Zhang ldquoDetection Performance of Packet Arrivalunder Downclocking for Mobile Edge Computingrdquo WirelessCommunications and Mobile Computing vol 2018 Article ID9641712 7 pages 2018

[10] W Yu F Liang X He et al ldquoA Survey on the Edge Computingfor the Internet of Thingsrdquo IEEE Access vol 6 pp 6900ndash69192018

[11] P Mach and Z Becvar ldquoMobile Edge Computing A Survey onArchitecture and Computation Offloadingrdquo IEEE Communica-tions Surveys amp Tutorials vol 19 no 3 pp 1628ndash1656 2017

[12] N Abbas Y Zhang A Taherkordi and T Skeie ldquoMobile EdgeComputing A Surveyrdquo IEEE Internet of Things Journal vol 5no 1 pp 450ndash465 2018

[13] Guangshun Li Jianrong Song Junhua Wu and Jiping WangldquoMethod of Resource Estimation Based on QoS in Edge Com-putingrdquo Wireless Communications and Mobile Computing vol2018 Article ID 7308913 9 pages 2018

[14] D Happ and A Wolisz ldquoTowards gateway to Cloud offloadingin IoT publishsubscribe systemsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 101ndash106 Valencia Spain May 2017

[15] G Premsankar M Di Francesco and T Taleb ldquoEdge Comput-ing for the Internet of Things A Case Studyrdquo IEEE Internet ofThings Journal vol 5 no 2 pp 1275ndash1284 2018

[16] L Zhao W Sun Y Shi and J Liu ldquoOptimal Placementof Cloudlets for Access Delay Minimization in SDN-BasedInternet of Things Networksrdquo IEEE Internet of Things Journalvol 5 no 2 pp 1334ndash1344 2018

[17] SWang X Zhang Y Zhang LWang J Yang andWWang ldquoASurvey onMobile Edge Networks Convergence of ComputingCaching and Communicationsrdquo IEEE Access vol 5 pp 6757ndash6779 2017

[18] Y Liu J E Fieldsend and G Min ldquoA Framework of FogComputing ArchitectureChallenges and Optimizationrdquo IEEEAccess vol 5 pp 25445ndash25454 2017

[19] ETSI Multi-access Edge Computing [Online] Availablehttpwwwetsiorgtechnologies-clusterstechnologiesmulti-access-edge-computing

[20] OpenFog Consortium [Online] Available httpswwwopen-fogconsortiumorg

[21] Y Mao C You J Zhang K Huang and K B Letaief ldquoA SurveyonMobile Edge ComputingThe Communication PerspectiverdquoIEEE Communications Surveys amp Tutorials vol 19 no 4 pp2322ndash2358 2017

[22] Cisco ldquoThe Internet of Things How the Next Evolution ofthe Internet Is Changing Everything white paperrdquo [Online]Available httpswwwciscocomcdamen usIoT IBSG0411FINALpdf

[23] IDC ldquoFutureScape Worldwide Internet of Things 2017 Pre-dictionsrdquo November 2016 [Online] Available httpswwwidccomgetdocjspcontainerId=US41910716

8 Wireless Communications and Mobile Computing

Input 119877 = 1198771 1198772 119877119899 ti di jOutput119873[119894 119895](1) Start(2) Set the largest task di as the baseline of the scheduler(3) for 119894 larr997888 1 to n initialize the maximum N at time 0 to be zero for each service

request in queue(4) 119873[119894 0] = 0(5) for 119895 larr997888 1 to 119889119899 determine the maximum N obtained by R1 at each time step(6) if ((j == t1) and (j lt= d1))(7) 119873[1 119895] = 1(8) else(9) 119873[1 119895] =119873[1 119895 minus 1](10) for 119894 larr997888 2 to n(11) for 119895 larr997888 1 to 119889119894(12) if (j gt di) case (1) the current time step j already exceeds 119877119894rsquos deadline(13) then119873[119894 119895] = 119873[119894 minus 1 119895 minus 119905119894](14) else if (j lt ti) case (2) there is not enough time to finish request Ri by the

current time j(15) then119873[119894 119895] = 119873[119894 minus 1 119895](16) else case (3) time j does not exceed 119877119894rsquos deadline and there is enough time

to finish request 119877119894 by time j(17) then119873[119894 119895] = max(119873[119894 minus 1 119895]119873[119894 minus 1 119895 minus 119905119894] + 1)(18) Schedule Completed(19) End

Algorithm 1 Greedy Available Fit algorithm

Virtualization technology can be applied on the CPE inmany ways such as by using the VM container or VM inte-gration of container However a traditional VM consumesmany system resources and cannot meet the requirementsfor light weight and service on-demand deployment In thiscase this research adopts container technology instead Acontainer handles only one service request at a time and stopsits operation after completing the delivery of a service TheVNF manager on the gateway is responsible for configuringand allocating each VNF The VNF template image that isrequired by different services is placed in the VNF resourcepool (with Docker Hub as the default other public or privateregistrations can also be specified) The edge gateway notonly controls the amount of gateway resources used byeach container but also manages several resources such asCPU and RAM to ensure that the container can obtain therequired resources without affecting the performance of theother executing containers on the edge gateway

Linux containers adopt a hierarchical structure to rapidlydeploy VNFs and to manage NFV flexible scheduling Theunderlying structure uses the file archiving mechanism ofDocker namely the advanced multilayered unification filesystem to incorporatemany different VNF imagesThe layersare stacked up andwhen someVNF service functions need tobe accessed the container retrieves the VNF images throughthe VNFmanager and uses these images directly To integratethe open resources of Docker Hub and to expand the privatewarehouse of the VNF service manager the system can storethe image files obtained from the public warehouse in aprivate warehouse for future use The judgment mechanismof the VNF service manager is described as follows

(1) If the image requested by the host is already in thelocal container then the VNF image can be taken outdirectly by the Docker daemon of the VNF servicemanager No complex configurations are required

(2) If the VNF image to be used by the host is not in thelocal container then the VNF manager connects tothe common warehouse Docker Hub on the Internetto find and store the desired service image file ina private warehouse via pushpull Afterward theacquired VNF image is controlled by the Dockerdaemon

By using the OS kernel for resource isolation the con-tainer technology does not need to rely on a virtual softwarelayer and does not require the VM to install a guest OSTherefore the capacity of the image file is much larger thanthat of the virtual machine The image file is small and canbe rapidly deployed through network transmission therebysaving network resources By using Linux container technol-ogy we can configure various VNF images to be executed onthe same edge gateway thereby replacing the virtual machineand achieving the goal of lightweight virtualization

5 Experiment

An experiment is conducted to evaluate the proposed taskscheduling algorithmand to test its performance in deployinglightweight VNF The experiment environment is shown inFigure 7 A simulation is initially performed to evaluate thetask scheduling performance of the algorithm by changingthe number of service requirements The service request is

Wireless Communications and Mobile Computing 9

Edge Network

Internet

Edge Gateway

DataCenter

Cloud Broker

Docker HubDocker Hub

Figure 7 The experiment environment

characterized as a Poisson process There we showed that theaverage number of service requests in the system is given by120582120583 where 120582 is the average arrival rate and 120583 is the averageservice rate The input parameter 120582120583 is 045 First comefirst service (FCFS) priority task scheduling [34] RR [37]and GBA [36] are then compared with the proposed GAFmechanism

The effect of the service requests on the average waitingtime average response time and task scheduling of differentscheduling mechanisms is evaluated in the simulation FCFSis based on the service requests that enter the edge gatewayqueuing system and schedules the tasks according to thesequence of these requests Priority task scheduling is anunfair scheduling algorithm where the service requests aresorted according to their priority and where the high-prioritytasks are performedfirstThose tasks having the same priorityare sorted by using the FIFO scheduling mechanism TheGBA sorts the tasks based on the completion time of theservice requests in queuing systemTheRR algorithm is basedon the conventional RR scheduling performed in the processscheduling RR scheduling fairly allocates the computingresources to those tasks with the same priority by using thetime slicing approach (time quantum)

Figure 8 shows the experiment result for average waitingtime where the x-axis refers to the number of service requestsand the y main axis refers to the average waiting time thatis the time a user spends to complete the service schedulingand to determine the resource allocation after receiving aservice request and a successful reply If the demand revertsto an errortimeout or if the edge gateway does not haveenough computing resources to support such demand thenthe sample is excluded from the calculation of the averagewaiting time As can be seen in the experimental resultthe RR algorithm obtains the longest average waiting timebecause of its application of time slice rules in order for eachtask to be processed within a fixed amount of time The timequantum affects the overall performance of the operation

0100200300400500600

10 20 30 40 50 60 70 80 90 100

Aver

age W

aitin

g Ti

me (

ms)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 8 The experiment result for average waiting time

Having a very long time quantum leads to a very long waitingtime while having a very short time quantum results intask schedule conversion a poor execution efficiency and anextended waiting time For FCFS given that the time spentdiffers across each task the waiting time for the next taskmust be determined based on the schedule of the previoustask Therefore if the last task schedule is too long thesystem cannot quickly process the subsequent task schedulesthereby affecting the overall task scheduling efficiency Giventhat its average waiting time is relatively shorter than thatof the FCFS mechanism priority task scheduling can satisfythe task requirements and flexibly perform the schedulingHowever prioritizing a high-priority program will delay allof the low-priority requests thereby creating an indefinitesituation in the low-priority program and extending thewaiting time GBA is based on the earliest completion time ofthe taskWhen the service demand is low the average waitingtime of the GBA algorithm is similar to that of priority taskscheduling However when the service demand increases theaverage waiting time of the GBAmechanism becomes longerthan that of priority task scheduling Compared with FCFS

10 Wireless Communications and Mobile Computing

050

100150200250300350

10 20 30 40 50 60 70 80 90 100

Ave

rage

Res

pons

e Tim

e (m

s)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 9 The experiment result for average response time

01020304050607080

10 20 30 40 50 60 70 80 90 100

Num

ber o

f Sch

edul

ed T

asks

Number of Service RequestsGAFFCFSPriority

GBARR

Figure 10 The experiment result for task scheduling efficiency

priority task scheduling RR and GBA the proposed GAFalgorithm selects the largest deadline task as the baseline andthen sequentially inserts the remaining tasks into the queuebased on their processing timeTherefore the GAF algorithmobtains the shortest average waiting time

Figure 9 shows the results for average response time Asexpected the RR algorithm outperforms all other approachesin terms of average response time because this algorithm isespecially designed for time sharing Using a fair strategy toallocate VNF resources ensures that each service request hasa fixed time If the service processing time exceeds the timequantum allocated by the system then the sample is excludedfrom the calculation of the average response time Given thatFCFS is processed in sequence the average response timewill be affected by the time spent by the previous 119877119894minus1 taskin the scheduler For example if the previous service requesthas a very long task schedule then the average responsetime increases Given that the proposed GAF algorithmsequentially inserts tasks into the queue its average responsetime is slightly longer than that of FCFS and GBA Althoughthe priority algorithm can sort the tasks according to thepriorities of the service requests the low-priority tasks areeasily delayed thereby increasing the average response time

Figure 10 shows the task scheduling efficiency of thecompared algorithms If the service request number is lessthan 60 then each algorithm can complete the task schedul-ing However when this number is exceeded the number

of tasks that can be scheduled varies across each algorithmGAF outperforms all the other algorithms as it ranks thelargest deadline task at the end and then sequentially sortsthe remaining tasks according to their processing time In thisway this algorithm ensures that most tasks will be completedwithin a minimum processing time Those service requeststhat cannot be scheduled will bypass the gateway and will bedirectly forwarded to the cloud data center When the servicerequest number is 80 GAF can complete 68 of the taskscheduling Meanwhile GBA performs the scheduling basedon the earliest task completion time Although GAF andGBAcan schedule the same number of tasks in the experimentalenvironment the latter has a longer average waiting timecompared with the former GAF also outperforms FCFCand priority task scheduling by 96 and 62 in terms oftask scheduling efficiency respectively because this approachranks the largest deadline task at the end and then sorts theremaining tasks according to their required processing timeGAF also outperforms theRRmechanismby 70because thelatter applies the time-sharing rule to minimize the numberof scheduled tasks Based on these results the GAF algorithmcan schedule more tasks in the edge gateway within a shortaverage waiting time and improve the operational efficiencyof edge computing services

In addition to task scheduling the VNF configuration isanother main factor that affects the operational efficiency ofthe edge computing service model This paper utilizes theDocker technology to achieve an agile deployment of VNFsSuch technology can provide a centralized management andflexible configuration of VNF services Depending on theservice requests of each user the deployment of differentVNFs can meet the application requirements of serviceson-demand The VNF image of Docker can be mainlydivided into four categories namely system images (egUbuntu CentOS and Debian) tool images (eg GolangFlask and Tomcat) service images (eg MySQL MongoDBand RabbitMQ) and application images (eg WordPressand DRUPAL) To test the practical deployment efficiency oflightweight VNF for the edge gateway a pull performancetest is performed for each type of Docker VNF image Theedge gateway is simulated by using a standalone PC with 34GHz dual-core CPU 16GBmemory and 5400 rpmHD500GThe experimental network bandwidth is set to 100 Mbps Tocompare the pull performance with the local edge gateway atest VM in the Google Cloud platform is used to simulate thecloud network broker Table 1 presents the result of the pullperformance test

As shown in Table 1 each VNF image size has differentphysical gateway and cloud network broker costs due to thelimitations in the network bandwidth A larger VNF imagetakes a longer time to be downloaded from Docker Hub Asexpected the cloud network broker has a better computingarchitecture than the local edge gateway because the locationof the edge gateway affects the performance of the VNFdeployment Although the edge computing architecture ofthe cloud network broker deploys VNF more efficiently thanthe local gateway the user service requestsmust still be sent tothe Internet for processing and this methodmakes it difficultto demonstrate the technical advantages of edge computing

Wireless Communications and Mobile Computing 11

Table 1 Pull performance test results for different VNFs

Image Types VNFs Image Size (MB) Physical GatewayCost (seconds)

Cloud NetworkBroker Cost(seconds)

System imagesubuntu 112 22095 7918centos 207 24319 8926debian 207 21635 5646

Tool imagesgolang 779 86793 1851flask 458 6222 1633

prometheus 112 2414 4737

Service images

httpd 177 20753 1031nginx 109 13216 4778bind 337 8438 15105squid 215 150031 9973mysql 374 24852 10449

mongodb 503 61347 17116

Application images wordpress 407 32866 13213drupal 452 8139 14405

11095

1593

0

20

40

60

80

100

120

140

160

180

Docker VM

Tim

e (s)

Figure 11 VNF boot time test for the VM and container

In the VNF boot time test we set the image to Ubuntu-16-04-x64 VM and Docker are used to perform the boottime test on the edge gateway We test the time of openingfive VMs and Docker containers VMware is used as the VMhypervisor As shown in Figure 11 the five VMs are openedin 1593 seconds while the Docker containers are openedin 11095 seconds In sum the VMs have a much longerdeployment time than Docker More time can be saved byusingDocker in the edge gateway that requires a large amountof VNFs

To test the average VNF boot time we boot a VM waitfor its activation shut it down and repeat the same steps for14 other VMs The test results are shown in Figure 12 Whenusing VMware the average time to boot the VMs is about3186 seconds However when using Docker these VMs arebooted in only 22 seconds In sum the Docker containerperforms approximately 15 times faster than the VM

The results of the VNF deployment test highlight thesignificant advantages of lightweight containers over tradi-tional virtualizationmethodsThese containers can be started

2219

3186

0

5

10

15

20

25

30

35

Docker VM

Tim

e (s)

Figure 12 Average VNF boot time test for the VM and container

within a few seconds because they adopt a common host OStherebymaking it unnecessary to execute the guestOS in eachcontainer The container also does not need to wait for theOS to start up thereby saving a few seconds Fast start is farfaster than traditional VMs Meanwhile given the design oflightweight VNF images and its highly efficient virtualizationthe application-centric virtualization technology Docker canautomatically obtain container images from the Docker Hubfor flexible deployment and management and can meet theapplication requirements of the edge network for the serviceon-demand of VNF deployment

6 Conclusions

This paper proposed a gateway-based edge computing servicemodel and a GAF task scheduling mechanism that allowsthe edge gateway to schedule more tasks within a short aver-age waiting time The resource estimation and lightweightVNF configuration technologies are used to improve theoperational efficiency of edge computing services to increase

12 Wireless Communications and Mobile Computing

resource utilization to achieve a rapid deployment of VNFand to meet service on-demand requests The combinationof resource estimation task scheduling and lightweight VNFconfiguration design provides an integrated solution that cansatisfy the service on-demand requests for 5G networks Thesimulation results showed that the proposedGAFmechanismoutperforms the other scheduling algorithmsMeanwhile thecomparison revealed that using the lightweight virtualizationtechnology in edge gateways ismore efficient and competitivethan using traditional VMs

In the future we plan to study the multiqueue taskscheduling problem for edge computing the possible coop-eration between edge gateway and cloud servers and the useof SD-WAN technology to achieve a seamless operation of thecloud-edge network

Data Availability

The data used to support the findings of this study areincluded within the article

Disclosure

The funder had no role in the study design data collectionand analysis decision to publish or preparation of thismanuscript

Conflicts of Interest

The authors declare no conflicts of interest regarding thepublication of this paper

Acknowledgments

The work described in this paper was supported in part bythe Ministry of Science and Technology of the Republic ofChina (Grants nos 104-2221-E-008-039 105-2221-E-008-071107-2623-E-008-002 and 107-2636-E-003-001)

References

[1] J Matias J Garay N Toledo J Unzilla and E Jacob ldquoTowardan SDN-enabled NFV architecturerdquo IEEE CommunicationsMagazine vol 53 no 4 pp 187ndash193 2015

[2] Q Duan N Ansari and M Toy ldquoSoftware-defined networkvirtualization An architectural framework for integrating SDNand NFV for service provisioning in future networksrdquo IEEENetwork vol 30 no 5 pp 10ndash16 2016

[3] TSI ldquoNetwork Functions Virtualization (NFV) architecturalframeworkrdquo[Online] Available httpswwwetsiorgdeliveretsi gsNFV001 099002010101 60gs NFV002v010101ppdf

[4] ONF ldquoSDN Architecturerdquo [Online] Available httpwwwopennetworkingorgimagesstoriesdownloadssdn-resourcestechnical-reportsTR SDN ARCH 10 06062014pdf

[5] A Basta A Blenk K Hoffmann H J Morper M Hoffmannand W Kellerer ldquoTowards a Cost Optimal Design for a 5GMobile Core Network Based on SDN and NFVrdquo IEEE Trans-actions on Network and Service Management vol 14 no 4 pp1061ndash1075 2017

[6] F Z Yousaf M Bredel S Schaller and F Schneider ldquoNFV andSDN-Key technology enablers for 5G networksrdquo IEEE Journal

on Selected Areas in Communications vol 35 no 11 pp 2468ndash2478 2017

[7] A C Baktir A Ozgovde and C Ersoy ldquoHow Can EdgeComputing Benefit FromSoftware-DefinedNetworking A Sur-vey Use Cases and Future Directionsrdquo EEE CommunicationsSurveys amp Tutorials vol 19 no 4 pp 2359ndash2391 2017

[8] Guangshun Li Jiping Wang Junhua Wu and Jianrong SongldquoData Processing Delay Optimization in Mobile Edge Comput-ingrdquoWireless Communications andMobile Computing vol 2018Article ID 6897523 9 pages 2018

[9] Zhimin Wang Qinglin Zhao Fangxin Xu Hongning Daiand Yujun Zhang ldquoDetection Performance of Packet Arrivalunder Downclocking for Mobile Edge Computingrdquo WirelessCommunications and Mobile Computing vol 2018 Article ID9641712 7 pages 2018

[10] W Yu F Liang X He et al ldquoA Survey on the Edge Computingfor the Internet of Thingsrdquo IEEE Access vol 6 pp 6900ndash69192018

[11] P Mach and Z Becvar ldquoMobile Edge Computing A Survey onArchitecture and Computation Offloadingrdquo IEEE Communica-tions Surveys amp Tutorials vol 19 no 3 pp 1628ndash1656 2017

[12] N Abbas Y Zhang A Taherkordi and T Skeie ldquoMobile EdgeComputing A Surveyrdquo IEEE Internet of Things Journal vol 5no 1 pp 450ndash465 2018

[13] Guangshun Li Jianrong Song Junhua Wu and Jiping WangldquoMethod of Resource Estimation Based on QoS in Edge Com-putingrdquo Wireless Communications and Mobile Computing vol2018 Article ID 7308913 9 pages 2018

[14] D Happ and A Wolisz ldquoTowards gateway to Cloud offloadingin IoT publishsubscribe systemsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 101ndash106 Valencia Spain May 2017

[15] G Premsankar M Di Francesco and T Taleb ldquoEdge Comput-ing for the Internet of Things A Case Studyrdquo IEEE Internet ofThings Journal vol 5 no 2 pp 1275ndash1284 2018

[16] L Zhao W Sun Y Shi and J Liu ldquoOptimal Placementof Cloudlets for Access Delay Minimization in SDN-BasedInternet of Things Networksrdquo IEEE Internet of Things Journalvol 5 no 2 pp 1334ndash1344 2018

[17] SWang X Zhang Y Zhang LWang J Yang andWWang ldquoASurvey onMobile Edge Networks Convergence of ComputingCaching and Communicationsrdquo IEEE Access vol 5 pp 6757ndash6779 2017

[18] Y Liu J E Fieldsend and G Min ldquoA Framework of FogComputing ArchitectureChallenges and Optimizationrdquo IEEEAccess vol 5 pp 25445ndash25454 2017

[19] ETSI Multi-access Edge Computing [Online] Availablehttpwwwetsiorgtechnologies-clusterstechnologiesmulti-access-edge-computing

[20] OpenFog Consortium [Online] Available httpswwwopen-fogconsortiumorg

[21] Y Mao C You J Zhang K Huang and K B Letaief ldquoA SurveyonMobile Edge ComputingThe Communication PerspectiverdquoIEEE Communications Surveys amp Tutorials vol 19 no 4 pp2322ndash2358 2017

[22] Cisco ldquoThe Internet of Things How the Next Evolution ofthe Internet Is Changing Everything white paperrdquo [Online]Available httpswwwciscocomcdamen usIoT IBSG0411FINALpdf

[23] IDC ldquoFutureScape Worldwide Internet of Things 2017 Pre-dictionsrdquo November 2016 [Online] Available httpswwwidccomgetdocjspcontainerId=US41910716

Wireless Communications and Mobile Computing 9

Edge Network

Internet

Edge Gateway

DataCenter

Cloud Broker

Docker HubDocker Hub

Figure 7 The experiment environment

characterized as a Poisson process There we showed that theaverage number of service requests in the system is given by120582120583 where 120582 is the average arrival rate and 120583 is the averageservice rate The input parameter 120582120583 is 045 First comefirst service (FCFS) priority task scheduling [34] RR [37]and GBA [36] are then compared with the proposed GAFmechanism

The effect of the service requests on the average waitingtime average response time and task scheduling of differentscheduling mechanisms is evaluated in the simulation FCFSis based on the service requests that enter the edge gatewayqueuing system and schedules the tasks according to thesequence of these requests Priority task scheduling is anunfair scheduling algorithm where the service requests aresorted according to their priority and where the high-prioritytasks are performedfirstThose tasks having the same priorityare sorted by using the FIFO scheduling mechanism TheGBA sorts the tasks based on the completion time of theservice requests in queuing systemTheRR algorithm is basedon the conventional RR scheduling performed in the processscheduling RR scheduling fairly allocates the computingresources to those tasks with the same priority by using thetime slicing approach (time quantum)

Figure 8 shows the experiment result for average waitingtime where the x-axis refers to the number of service requestsand the y main axis refers to the average waiting time thatis the time a user spends to complete the service schedulingand to determine the resource allocation after receiving aservice request and a successful reply If the demand revertsto an errortimeout or if the edge gateway does not haveenough computing resources to support such demand thenthe sample is excluded from the calculation of the averagewaiting time As can be seen in the experimental resultthe RR algorithm obtains the longest average waiting timebecause of its application of time slice rules in order for eachtask to be processed within a fixed amount of time The timequantum affects the overall performance of the operation

0100200300400500600

10 20 30 40 50 60 70 80 90 100

Aver

age W

aitin

g Ti

me (

ms)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 8 The experiment result for average waiting time

Having a very long time quantum leads to a very long waitingtime while having a very short time quantum results intask schedule conversion a poor execution efficiency and anextended waiting time For FCFS given that the time spentdiffers across each task the waiting time for the next taskmust be determined based on the schedule of the previoustask Therefore if the last task schedule is too long thesystem cannot quickly process the subsequent task schedulesthereby affecting the overall task scheduling efficiency Giventhat its average waiting time is relatively shorter than thatof the FCFS mechanism priority task scheduling can satisfythe task requirements and flexibly perform the schedulingHowever prioritizing a high-priority program will delay allof the low-priority requests thereby creating an indefinitesituation in the low-priority program and extending thewaiting time GBA is based on the earliest completion time ofthe taskWhen the service demand is low the average waitingtime of the GBA algorithm is similar to that of priority taskscheduling However when the service demand increases theaverage waiting time of the GBAmechanism becomes longerthan that of priority task scheduling Compared with FCFS

10 Wireless Communications and Mobile Computing

050

100150200250300350

10 20 30 40 50 60 70 80 90 100

Ave

rage

Res

pons

e Tim

e (m

s)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 9 The experiment result for average response time

01020304050607080

10 20 30 40 50 60 70 80 90 100

Num

ber o

f Sch

edul

ed T

asks

Number of Service RequestsGAFFCFSPriority

GBARR

Figure 10 The experiment result for task scheduling efficiency

priority task scheduling RR and GBA the proposed GAFalgorithm selects the largest deadline task as the baseline andthen sequentially inserts the remaining tasks into the queuebased on their processing timeTherefore the GAF algorithmobtains the shortest average waiting time

Figure 9 shows the results for average response time Asexpected the RR algorithm outperforms all other approachesin terms of average response time because this algorithm isespecially designed for time sharing Using a fair strategy toallocate VNF resources ensures that each service request hasa fixed time If the service processing time exceeds the timequantum allocated by the system then the sample is excludedfrom the calculation of the average response time Given thatFCFS is processed in sequence the average response timewill be affected by the time spent by the previous 119877119894minus1 taskin the scheduler For example if the previous service requesthas a very long task schedule then the average responsetime increases Given that the proposed GAF algorithmsequentially inserts tasks into the queue its average responsetime is slightly longer than that of FCFS and GBA Althoughthe priority algorithm can sort the tasks according to thepriorities of the service requests the low-priority tasks areeasily delayed thereby increasing the average response time

Figure 10 shows the task scheduling efficiency of thecompared algorithms If the service request number is lessthan 60 then each algorithm can complete the task schedul-ing However when this number is exceeded the number

of tasks that can be scheduled varies across each algorithmGAF outperforms all the other algorithms as it ranks thelargest deadline task at the end and then sequentially sortsthe remaining tasks according to their processing time In thisway this algorithm ensures that most tasks will be completedwithin a minimum processing time Those service requeststhat cannot be scheduled will bypass the gateway and will bedirectly forwarded to the cloud data center When the servicerequest number is 80 GAF can complete 68 of the taskscheduling Meanwhile GBA performs the scheduling basedon the earliest task completion time Although GAF andGBAcan schedule the same number of tasks in the experimentalenvironment the latter has a longer average waiting timecompared with the former GAF also outperforms FCFCand priority task scheduling by 96 and 62 in terms oftask scheduling efficiency respectively because this approachranks the largest deadline task at the end and then sorts theremaining tasks according to their required processing timeGAF also outperforms theRRmechanismby 70because thelatter applies the time-sharing rule to minimize the numberof scheduled tasks Based on these results the GAF algorithmcan schedule more tasks in the edge gateway within a shortaverage waiting time and improve the operational efficiencyof edge computing services

In addition to task scheduling the VNF configuration isanother main factor that affects the operational efficiency ofthe edge computing service model This paper utilizes theDocker technology to achieve an agile deployment of VNFsSuch technology can provide a centralized management andflexible configuration of VNF services Depending on theservice requests of each user the deployment of differentVNFs can meet the application requirements of serviceson-demand The VNF image of Docker can be mainlydivided into four categories namely system images (egUbuntu CentOS and Debian) tool images (eg GolangFlask and Tomcat) service images (eg MySQL MongoDBand RabbitMQ) and application images (eg WordPressand DRUPAL) To test the practical deployment efficiency oflightweight VNF for the edge gateway a pull performancetest is performed for each type of Docker VNF image Theedge gateway is simulated by using a standalone PC with 34GHz dual-core CPU 16GBmemory and 5400 rpmHD500GThe experimental network bandwidth is set to 100 Mbps Tocompare the pull performance with the local edge gateway atest VM in the Google Cloud platform is used to simulate thecloud network broker Table 1 presents the result of the pullperformance test

As shown in Table 1 each VNF image size has differentphysical gateway and cloud network broker costs due to thelimitations in the network bandwidth A larger VNF imagetakes a longer time to be downloaded from Docker Hub Asexpected the cloud network broker has a better computingarchitecture than the local edge gateway because the locationof the edge gateway affects the performance of the VNFdeployment Although the edge computing architecture ofthe cloud network broker deploys VNF more efficiently thanthe local gateway the user service requestsmust still be sent tothe Internet for processing and this methodmakes it difficultto demonstrate the technical advantages of edge computing

Wireless Communications and Mobile Computing 11

Table 1 Pull performance test results for different VNFs

Image Types VNFs Image Size (MB) Physical GatewayCost (seconds)

Cloud NetworkBroker Cost(seconds)

System imagesubuntu 112 22095 7918centos 207 24319 8926debian 207 21635 5646

Tool imagesgolang 779 86793 1851flask 458 6222 1633

prometheus 112 2414 4737

Service images

httpd 177 20753 1031nginx 109 13216 4778bind 337 8438 15105squid 215 150031 9973mysql 374 24852 10449

mongodb 503 61347 17116

Application images wordpress 407 32866 13213drupal 452 8139 14405

11095

1593

0

20

40

60

80

100

120

140

160

180

Docker VM

Tim

e (s)

Figure 11 VNF boot time test for the VM and container

In the VNF boot time test we set the image to Ubuntu-16-04-x64 VM and Docker are used to perform the boottime test on the edge gateway We test the time of openingfive VMs and Docker containers VMware is used as the VMhypervisor As shown in Figure 11 the five VMs are openedin 1593 seconds while the Docker containers are openedin 11095 seconds In sum the VMs have a much longerdeployment time than Docker More time can be saved byusingDocker in the edge gateway that requires a large amountof VNFs

To test the average VNF boot time we boot a VM waitfor its activation shut it down and repeat the same steps for14 other VMs The test results are shown in Figure 12 Whenusing VMware the average time to boot the VMs is about3186 seconds However when using Docker these VMs arebooted in only 22 seconds In sum the Docker containerperforms approximately 15 times faster than the VM

The results of the VNF deployment test highlight thesignificant advantages of lightweight containers over tradi-tional virtualizationmethodsThese containers can be started

2219

3186

0

5

10

15

20

25

30

35

Docker VM

Tim

e (s)

Figure 12 Average VNF boot time test for the VM and container

within a few seconds because they adopt a common host OStherebymaking it unnecessary to execute the guestOS in eachcontainer The container also does not need to wait for theOS to start up thereby saving a few seconds Fast start is farfaster than traditional VMs Meanwhile given the design oflightweight VNF images and its highly efficient virtualizationthe application-centric virtualization technology Docker canautomatically obtain container images from the Docker Hubfor flexible deployment and management and can meet theapplication requirements of the edge network for the serviceon-demand of VNF deployment

6 Conclusions

This paper proposed a gateway-based edge computing servicemodel and a GAF task scheduling mechanism that allowsthe edge gateway to schedule more tasks within a short aver-age waiting time The resource estimation and lightweightVNF configuration technologies are used to improve theoperational efficiency of edge computing services to increase

12 Wireless Communications and Mobile Computing

resource utilization to achieve a rapid deployment of VNFand to meet service on-demand requests The combinationof resource estimation task scheduling and lightweight VNFconfiguration design provides an integrated solution that cansatisfy the service on-demand requests for 5G networks Thesimulation results showed that the proposedGAFmechanismoutperforms the other scheduling algorithmsMeanwhile thecomparison revealed that using the lightweight virtualizationtechnology in edge gateways ismore efficient and competitivethan using traditional VMs

In the future we plan to study the multiqueue taskscheduling problem for edge computing the possible coop-eration between edge gateway and cloud servers and the useof SD-WAN technology to achieve a seamless operation of thecloud-edge network

Data Availability

The data used to support the findings of this study areincluded within the article

Disclosure

The funder had no role in the study design data collectionand analysis decision to publish or preparation of thismanuscript

Conflicts of Interest

The authors declare no conflicts of interest regarding thepublication of this paper

Acknowledgments

The work described in this paper was supported in part bythe Ministry of Science and Technology of the Republic ofChina (Grants nos 104-2221-E-008-039 105-2221-E-008-071107-2623-E-008-002 and 107-2636-E-003-001)

References

[1] J Matias J Garay N Toledo J Unzilla and E Jacob ldquoTowardan SDN-enabled NFV architecturerdquo IEEE CommunicationsMagazine vol 53 no 4 pp 187ndash193 2015

[2] Q Duan N Ansari and M Toy ldquoSoftware-defined networkvirtualization An architectural framework for integrating SDNand NFV for service provisioning in future networksrdquo IEEENetwork vol 30 no 5 pp 10ndash16 2016

[3] TSI ldquoNetwork Functions Virtualization (NFV) architecturalframeworkrdquo[Online] Available httpswwwetsiorgdeliveretsi gsNFV001 099002010101 60gs NFV002v010101ppdf

[4] ONF ldquoSDN Architecturerdquo [Online] Available httpwwwopennetworkingorgimagesstoriesdownloadssdn-resourcestechnical-reportsTR SDN ARCH 10 06062014pdf

[5] A Basta A Blenk K Hoffmann H J Morper M Hoffmannand W Kellerer ldquoTowards a Cost Optimal Design for a 5GMobile Core Network Based on SDN and NFVrdquo IEEE Trans-actions on Network and Service Management vol 14 no 4 pp1061ndash1075 2017

[6] F Z Yousaf M Bredel S Schaller and F Schneider ldquoNFV andSDN-Key technology enablers for 5G networksrdquo IEEE Journal

on Selected Areas in Communications vol 35 no 11 pp 2468ndash2478 2017

[7] A C Baktir A Ozgovde and C Ersoy ldquoHow Can EdgeComputing Benefit FromSoftware-DefinedNetworking A Sur-vey Use Cases and Future Directionsrdquo EEE CommunicationsSurveys amp Tutorials vol 19 no 4 pp 2359ndash2391 2017

[8] Guangshun Li Jiping Wang Junhua Wu and Jianrong SongldquoData Processing Delay Optimization in Mobile Edge Comput-ingrdquoWireless Communications andMobile Computing vol 2018Article ID 6897523 9 pages 2018

[9] Zhimin Wang Qinglin Zhao Fangxin Xu Hongning Daiand Yujun Zhang ldquoDetection Performance of Packet Arrivalunder Downclocking for Mobile Edge Computingrdquo WirelessCommunications and Mobile Computing vol 2018 Article ID9641712 7 pages 2018

[10] W Yu F Liang X He et al ldquoA Survey on the Edge Computingfor the Internet of Thingsrdquo IEEE Access vol 6 pp 6900ndash69192018

[11] P Mach and Z Becvar ldquoMobile Edge Computing A Survey onArchitecture and Computation Offloadingrdquo IEEE Communica-tions Surveys amp Tutorials vol 19 no 3 pp 1628ndash1656 2017

[12] N Abbas Y Zhang A Taherkordi and T Skeie ldquoMobile EdgeComputing A Surveyrdquo IEEE Internet of Things Journal vol 5no 1 pp 450ndash465 2018

[13] Guangshun Li Jianrong Song Junhua Wu and Jiping WangldquoMethod of Resource Estimation Based on QoS in Edge Com-putingrdquo Wireless Communications and Mobile Computing vol2018 Article ID 7308913 9 pages 2018

[14] D Happ and A Wolisz ldquoTowards gateway to Cloud offloadingin IoT publishsubscribe systemsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 101ndash106 Valencia Spain May 2017

[15] G Premsankar M Di Francesco and T Taleb ldquoEdge Comput-ing for the Internet of Things A Case Studyrdquo IEEE Internet ofThings Journal vol 5 no 2 pp 1275ndash1284 2018

[16] L Zhao W Sun Y Shi and J Liu ldquoOptimal Placementof Cloudlets for Access Delay Minimization in SDN-BasedInternet of Things Networksrdquo IEEE Internet of Things Journalvol 5 no 2 pp 1334ndash1344 2018

[17] SWang X Zhang Y Zhang LWang J Yang andWWang ldquoASurvey onMobile Edge Networks Convergence of ComputingCaching and Communicationsrdquo IEEE Access vol 5 pp 6757ndash6779 2017

[18] Y Liu J E Fieldsend and G Min ldquoA Framework of FogComputing ArchitectureChallenges and Optimizationrdquo IEEEAccess vol 5 pp 25445ndash25454 2017

[19] ETSI Multi-access Edge Computing [Online] Availablehttpwwwetsiorgtechnologies-clusterstechnologiesmulti-access-edge-computing

[20] OpenFog Consortium [Online] Available httpswwwopen-fogconsortiumorg

[21] Y Mao C You J Zhang K Huang and K B Letaief ldquoA SurveyonMobile Edge ComputingThe Communication PerspectiverdquoIEEE Communications Surveys amp Tutorials vol 19 no 4 pp2322ndash2358 2017

[22] Cisco ldquoThe Internet of Things How the Next Evolution ofthe Internet Is Changing Everything white paperrdquo [Online]Available httpswwwciscocomcdamen usIoT IBSG0411FINALpdf

[23] IDC ldquoFutureScape Worldwide Internet of Things 2017 Pre-dictionsrdquo November 2016 [Online] Available httpswwwidccomgetdocjspcontainerId=US41910716

10 Wireless Communications and Mobile Computing

050

100150200250300350

10 20 30 40 50 60 70 80 90 100

Ave

rage

Res

pons

e Tim

e (m

s)

Number of Service Requests

GAFFCFSPriority

GBARR

Figure 9 The experiment result for average response time

01020304050607080

10 20 30 40 50 60 70 80 90 100

Num

ber o

f Sch

edul

ed T

asks

Number of Service RequestsGAFFCFSPriority

GBARR

Figure 10 The experiment result for task scheduling efficiency

priority task scheduling RR and GBA the proposed GAFalgorithm selects the largest deadline task as the baseline andthen sequentially inserts the remaining tasks into the queuebased on their processing timeTherefore the GAF algorithmobtains the shortest average waiting time

Figure 9 shows the results for average response time Asexpected the RR algorithm outperforms all other approachesin terms of average response time because this algorithm isespecially designed for time sharing Using a fair strategy toallocate VNF resources ensures that each service request hasa fixed time If the service processing time exceeds the timequantum allocated by the system then the sample is excludedfrom the calculation of the average response time Given thatFCFS is processed in sequence the average response timewill be affected by the time spent by the previous 119877119894minus1 taskin the scheduler For example if the previous service requesthas a very long task schedule then the average responsetime increases Given that the proposed GAF algorithmsequentially inserts tasks into the queue its average responsetime is slightly longer than that of FCFS and GBA Althoughthe priority algorithm can sort the tasks according to thepriorities of the service requests the low-priority tasks areeasily delayed thereby increasing the average response time

Figure 10 shows the task scheduling efficiency of thecompared algorithms If the service request number is lessthan 60 then each algorithm can complete the task schedul-ing However when this number is exceeded the number

of tasks that can be scheduled varies across each algorithmGAF outperforms all the other algorithms as it ranks thelargest deadline task at the end and then sequentially sortsthe remaining tasks according to their processing time In thisway this algorithm ensures that most tasks will be completedwithin a minimum processing time Those service requeststhat cannot be scheduled will bypass the gateway and will bedirectly forwarded to the cloud data center When the servicerequest number is 80 GAF can complete 68 of the taskscheduling Meanwhile GBA performs the scheduling basedon the earliest task completion time Although GAF andGBAcan schedule the same number of tasks in the experimentalenvironment the latter has a longer average waiting timecompared with the former GAF also outperforms FCFCand priority task scheduling by 96 and 62 in terms oftask scheduling efficiency respectively because this approachranks the largest deadline task at the end and then sorts theremaining tasks according to their required processing timeGAF also outperforms theRRmechanismby 70because thelatter applies the time-sharing rule to minimize the numberof scheduled tasks Based on these results the GAF algorithmcan schedule more tasks in the edge gateway within a shortaverage waiting time and improve the operational efficiencyof edge computing services

In addition to task scheduling the VNF configuration isanother main factor that affects the operational efficiency ofthe edge computing service model This paper utilizes theDocker technology to achieve an agile deployment of VNFsSuch technology can provide a centralized management andflexible configuration of VNF services Depending on theservice requests of each user the deployment of differentVNFs can meet the application requirements of serviceson-demand The VNF image of Docker can be mainlydivided into four categories namely system images (egUbuntu CentOS and Debian) tool images (eg GolangFlask and Tomcat) service images (eg MySQL MongoDBand RabbitMQ) and application images (eg WordPressand DRUPAL) To test the practical deployment efficiency oflightweight VNF for the edge gateway a pull performancetest is performed for each type of Docker VNF image Theedge gateway is simulated by using a standalone PC with 34GHz dual-core CPU 16GBmemory and 5400 rpmHD500GThe experimental network bandwidth is set to 100 Mbps Tocompare the pull performance with the local edge gateway atest VM in the Google Cloud platform is used to simulate thecloud network broker Table 1 presents the result of the pullperformance test

As shown in Table 1 each VNF image size has differentphysical gateway and cloud network broker costs due to thelimitations in the network bandwidth A larger VNF imagetakes a longer time to be downloaded from Docker Hub Asexpected the cloud network broker has a better computingarchitecture than the local edge gateway because the locationof the edge gateway affects the performance of the VNFdeployment Although the edge computing architecture ofthe cloud network broker deploys VNF more efficiently thanthe local gateway the user service requestsmust still be sent tothe Internet for processing and this methodmakes it difficultto demonstrate the technical advantages of edge computing

Wireless Communications and Mobile Computing 11

Table 1 Pull performance test results for different VNFs

Image Types VNFs Image Size (MB) Physical GatewayCost (seconds)

Cloud NetworkBroker Cost(seconds)

System imagesubuntu 112 22095 7918centos 207 24319 8926debian 207 21635 5646

Tool imagesgolang 779 86793 1851flask 458 6222 1633

prometheus 112 2414 4737

Service images

httpd 177 20753 1031nginx 109 13216 4778bind 337 8438 15105squid 215 150031 9973mysql 374 24852 10449

mongodb 503 61347 17116

Application images wordpress 407 32866 13213drupal 452 8139 14405

11095

1593

0

20

40

60

80

100

120

140

160

180

Docker VM

Tim

e (s)

Figure 11 VNF boot time test for the VM and container

In the VNF boot time test we set the image to Ubuntu-16-04-x64 VM and Docker are used to perform the boottime test on the edge gateway We test the time of openingfive VMs and Docker containers VMware is used as the VMhypervisor As shown in Figure 11 the five VMs are openedin 1593 seconds while the Docker containers are openedin 11095 seconds In sum the VMs have a much longerdeployment time than Docker More time can be saved byusingDocker in the edge gateway that requires a large amountof VNFs

To test the average VNF boot time we boot a VM waitfor its activation shut it down and repeat the same steps for14 other VMs The test results are shown in Figure 12 Whenusing VMware the average time to boot the VMs is about3186 seconds However when using Docker these VMs arebooted in only 22 seconds In sum the Docker containerperforms approximately 15 times faster than the VM

The results of the VNF deployment test highlight thesignificant advantages of lightweight containers over tradi-tional virtualizationmethodsThese containers can be started

2219

3186

0

5

10

15

20

25

30

35

Docker VM

Tim

e (s)

Figure 12 Average VNF boot time test for the VM and container

within a few seconds because they adopt a common host OStherebymaking it unnecessary to execute the guestOS in eachcontainer The container also does not need to wait for theOS to start up thereby saving a few seconds Fast start is farfaster than traditional VMs Meanwhile given the design oflightweight VNF images and its highly efficient virtualizationthe application-centric virtualization technology Docker canautomatically obtain container images from the Docker Hubfor flexible deployment and management and can meet theapplication requirements of the edge network for the serviceon-demand of VNF deployment

6 Conclusions

This paper proposed a gateway-based edge computing servicemodel and a GAF task scheduling mechanism that allowsthe edge gateway to schedule more tasks within a short aver-age waiting time The resource estimation and lightweightVNF configuration technologies are used to improve theoperational efficiency of edge computing services to increase

12 Wireless Communications and Mobile Computing

resource utilization to achieve a rapid deployment of VNFand to meet service on-demand requests The combinationof resource estimation task scheduling and lightweight VNFconfiguration design provides an integrated solution that cansatisfy the service on-demand requests for 5G networks Thesimulation results showed that the proposedGAFmechanismoutperforms the other scheduling algorithmsMeanwhile thecomparison revealed that using the lightweight virtualizationtechnology in edge gateways ismore efficient and competitivethan using traditional VMs

In the future we plan to study the multiqueue taskscheduling problem for edge computing the possible coop-eration between edge gateway and cloud servers and the useof SD-WAN technology to achieve a seamless operation of thecloud-edge network

Data Availability

The data used to support the findings of this study areincluded within the article

Disclosure

The funder had no role in the study design data collectionand analysis decision to publish or preparation of thismanuscript

Conflicts of Interest

The authors declare no conflicts of interest regarding thepublication of this paper

Acknowledgments

The work described in this paper was supported in part bythe Ministry of Science and Technology of the Republic ofChina (Grants nos 104-2221-E-008-039 105-2221-E-008-071107-2623-E-008-002 and 107-2636-E-003-001)

References

[1] J Matias J Garay N Toledo J Unzilla and E Jacob ldquoTowardan SDN-enabled NFV architecturerdquo IEEE CommunicationsMagazine vol 53 no 4 pp 187ndash193 2015

[2] Q Duan N Ansari and M Toy ldquoSoftware-defined networkvirtualization An architectural framework for integrating SDNand NFV for service provisioning in future networksrdquo IEEENetwork vol 30 no 5 pp 10ndash16 2016

[3] TSI ldquoNetwork Functions Virtualization (NFV) architecturalframeworkrdquo[Online] Available httpswwwetsiorgdeliveretsi gsNFV001 099002010101 60gs NFV002v010101ppdf

[4] ONF ldquoSDN Architecturerdquo [Online] Available httpwwwopennetworkingorgimagesstoriesdownloadssdn-resourcestechnical-reportsTR SDN ARCH 10 06062014pdf

[5] A Basta A Blenk K Hoffmann H J Morper M Hoffmannand W Kellerer ldquoTowards a Cost Optimal Design for a 5GMobile Core Network Based on SDN and NFVrdquo IEEE Trans-actions on Network and Service Management vol 14 no 4 pp1061ndash1075 2017

[6] F Z Yousaf M Bredel S Schaller and F Schneider ldquoNFV andSDN-Key technology enablers for 5G networksrdquo IEEE Journal

on Selected Areas in Communications vol 35 no 11 pp 2468ndash2478 2017

[7] A C Baktir A Ozgovde and C Ersoy ldquoHow Can EdgeComputing Benefit FromSoftware-DefinedNetworking A Sur-vey Use Cases and Future Directionsrdquo EEE CommunicationsSurveys amp Tutorials vol 19 no 4 pp 2359ndash2391 2017

[8] Guangshun Li Jiping Wang Junhua Wu and Jianrong SongldquoData Processing Delay Optimization in Mobile Edge Comput-ingrdquoWireless Communications andMobile Computing vol 2018Article ID 6897523 9 pages 2018

[9] Zhimin Wang Qinglin Zhao Fangxin Xu Hongning Daiand Yujun Zhang ldquoDetection Performance of Packet Arrivalunder Downclocking for Mobile Edge Computingrdquo WirelessCommunications and Mobile Computing vol 2018 Article ID9641712 7 pages 2018

[10] W Yu F Liang X He et al ldquoA Survey on the Edge Computingfor the Internet of Thingsrdquo IEEE Access vol 6 pp 6900ndash69192018

[11] P Mach and Z Becvar ldquoMobile Edge Computing A Survey onArchitecture and Computation Offloadingrdquo IEEE Communica-tions Surveys amp Tutorials vol 19 no 3 pp 1628ndash1656 2017

[12] N Abbas Y Zhang A Taherkordi and T Skeie ldquoMobile EdgeComputing A Surveyrdquo IEEE Internet of Things Journal vol 5no 1 pp 450ndash465 2018

[13] Guangshun Li Jianrong Song Junhua Wu and Jiping WangldquoMethod of Resource Estimation Based on QoS in Edge Com-putingrdquo Wireless Communications and Mobile Computing vol2018 Article ID 7308913 9 pages 2018

[14] D Happ and A Wolisz ldquoTowards gateway to Cloud offloadingin IoT publishsubscribe systemsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 101ndash106 Valencia Spain May 2017

[15] G Premsankar M Di Francesco and T Taleb ldquoEdge Comput-ing for the Internet of Things A Case Studyrdquo IEEE Internet ofThings Journal vol 5 no 2 pp 1275ndash1284 2018

[16] L Zhao W Sun Y Shi and J Liu ldquoOptimal Placementof Cloudlets for Access Delay Minimization in SDN-BasedInternet of Things Networksrdquo IEEE Internet of Things Journalvol 5 no 2 pp 1334ndash1344 2018

[17] SWang X Zhang Y Zhang LWang J Yang andWWang ldquoASurvey onMobile Edge Networks Convergence of ComputingCaching and Communicationsrdquo IEEE Access vol 5 pp 6757ndash6779 2017

[18] Y Liu J E Fieldsend and G Min ldquoA Framework of FogComputing ArchitectureChallenges and Optimizationrdquo IEEEAccess vol 5 pp 25445ndash25454 2017

[19] ETSI Multi-access Edge Computing [Online] Availablehttpwwwetsiorgtechnologies-clusterstechnologiesmulti-access-edge-computing

[20] OpenFog Consortium [Online] Available httpswwwopen-fogconsortiumorg

[21] Y Mao C You J Zhang K Huang and K B Letaief ldquoA SurveyonMobile Edge ComputingThe Communication PerspectiverdquoIEEE Communications Surveys amp Tutorials vol 19 no 4 pp2322ndash2358 2017

[22] Cisco ldquoThe Internet of Things How the Next Evolution ofthe Internet Is Changing Everything white paperrdquo [Online]Available httpswwwciscocomcdamen usIoT IBSG0411FINALpdf

[23] IDC ldquoFutureScape Worldwide Internet of Things 2017 Pre-dictionsrdquo November 2016 [Online] Available httpswwwidccomgetdocjspcontainerId=US41910716

Wireless Communications and Mobile Computing 11

Table 1 Pull performance test results for different VNFs

Image Types VNFs Image Size (MB) Physical GatewayCost (seconds)

Cloud NetworkBroker Cost(seconds)

System imagesubuntu 112 22095 7918centos 207 24319 8926debian 207 21635 5646

Tool imagesgolang 779 86793 1851flask 458 6222 1633

prometheus 112 2414 4737

Service images

httpd 177 20753 1031nginx 109 13216 4778bind 337 8438 15105squid 215 150031 9973mysql 374 24852 10449

mongodb 503 61347 17116

Application images wordpress 407 32866 13213drupal 452 8139 14405

11095

1593

0

20

40

60

80

100

120

140

160

180

Docker VM

Tim

e (s)

Figure 11 VNF boot time test for the VM and container

In the VNF boot time test we set the image to Ubuntu-16-04-x64 VM and Docker are used to perform the boottime test on the edge gateway We test the time of openingfive VMs and Docker containers VMware is used as the VMhypervisor As shown in Figure 11 the five VMs are openedin 1593 seconds while the Docker containers are openedin 11095 seconds In sum the VMs have a much longerdeployment time than Docker More time can be saved byusingDocker in the edge gateway that requires a large amountof VNFs

To test the average VNF boot time we boot a VM waitfor its activation shut it down and repeat the same steps for14 other VMs The test results are shown in Figure 12 Whenusing VMware the average time to boot the VMs is about3186 seconds However when using Docker these VMs arebooted in only 22 seconds In sum the Docker containerperforms approximately 15 times faster than the VM

The results of the VNF deployment test highlight thesignificant advantages of lightweight containers over tradi-tional virtualizationmethodsThese containers can be started

2219

3186

0

5

10

15

20

25

30

35

Docker VM

Tim

e (s)

Figure 12 Average VNF boot time test for the VM and container

within a few seconds because they adopt a common host OStherebymaking it unnecessary to execute the guestOS in eachcontainer The container also does not need to wait for theOS to start up thereby saving a few seconds Fast start is farfaster than traditional VMs Meanwhile given the design oflightweight VNF images and its highly efficient virtualizationthe application-centric virtualization technology Docker canautomatically obtain container images from the Docker Hubfor flexible deployment and management and can meet theapplication requirements of the edge network for the serviceon-demand of VNF deployment

6 Conclusions

This paper proposed a gateway-based edge computing servicemodel and a GAF task scheduling mechanism that allowsthe edge gateway to schedule more tasks within a short aver-age waiting time The resource estimation and lightweightVNF configuration technologies are used to improve theoperational efficiency of edge computing services to increase

12 Wireless Communications and Mobile Computing

resource utilization to achieve a rapid deployment of VNFand to meet service on-demand requests The combinationof resource estimation task scheduling and lightweight VNFconfiguration design provides an integrated solution that cansatisfy the service on-demand requests for 5G networks Thesimulation results showed that the proposedGAFmechanismoutperforms the other scheduling algorithmsMeanwhile thecomparison revealed that using the lightweight virtualizationtechnology in edge gateways ismore efficient and competitivethan using traditional VMs

In the future we plan to study the multiqueue taskscheduling problem for edge computing the possible coop-eration between edge gateway and cloud servers and the useof SD-WAN technology to achieve a seamless operation of thecloud-edge network

Data Availability

The data used to support the findings of this study areincluded within the article

Disclosure

The funder had no role in the study design data collectionand analysis decision to publish or preparation of thismanuscript

Conflicts of Interest

The authors declare no conflicts of interest regarding thepublication of this paper

Acknowledgments

The work described in this paper was supported in part bythe Ministry of Science and Technology of the Republic ofChina (Grants nos 104-2221-E-008-039 105-2221-E-008-071107-2623-E-008-002 and 107-2636-E-003-001)

References

[1] J Matias J Garay N Toledo J Unzilla and E Jacob ldquoTowardan SDN-enabled NFV architecturerdquo IEEE CommunicationsMagazine vol 53 no 4 pp 187ndash193 2015

[2] Q Duan N Ansari and M Toy ldquoSoftware-defined networkvirtualization An architectural framework for integrating SDNand NFV for service provisioning in future networksrdquo IEEENetwork vol 30 no 5 pp 10ndash16 2016

[3] TSI ldquoNetwork Functions Virtualization (NFV) architecturalframeworkrdquo[Online] Available httpswwwetsiorgdeliveretsi gsNFV001 099002010101 60gs NFV002v010101ppdf

[4] ONF ldquoSDN Architecturerdquo [Online] Available httpwwwopennetworkingorgimagesstoriesdownloadssdn-resourcestechnical-reportsTR SDN ARCH 10 06062014pdf

[5] A Basta A Blenk K Hoffmann H J Morper M Hoffmannand W Kellerer ldquoTowards a Cost Optimal Design for a 5GMobile Core Network Based on SDN and NFVrdquo IEEE Trans-actions on Network and Service Management vol 14 no 4 pp1061ndash1075 2017

[6] F Z Yousaf M Bredel S Schaller and F Schneider ldquoNFV andSDN-Key technology enablers for 5G networksrdquo IEEE Journal

on Selected Areas in Communications vol 35 no 11 pp 2468ndash2478 2017

[7] A C Baktir A Ozgovde and C Ersoy ldquoHow Can EdgeComputing Benefit FromSoftware-DefinedNetworking A Sur-vey Use Cases and Future Directionsrdquo EEE CommunicationsSurveys amp Tutorials vol 19 no 4 pp 2359ndash2391 2017

[8] Guangshun Li Jiping Wang Junhua Wu and Jianrong SongldquoData Processing Delay Optimization in Mobile Edge Comput-ingrdquoWireless Communications andMobile Computing vol 2018Article ID 6897523 9 pages 2018

[9] Zhimin Wang Qinglin Zhao Fangxin Xu Hongning Daiand Yujun Zhang ldquoDetection Performance of Packet Arrivalunder Downclocking for Mobile Edge Computingrdquo WirelessCommunications and Mobile Computing vol 2018 Article ID9641712 7 pages 2018

[10] W Yu F Liang X He et al ldquoA Survey on the Edge Computingfor the Internet of Thingsrdquo IEEE Access vol 6 pp 6900ndash69192018

[11] P Mach and Z Becvar ldquoMobile Edge Computing A Survey onArchitecture and Computation Offloadingrdquo IEEE Communica-tions Surveys amp Tutorials vol 19 no 3 pp 1628ndash1656 2017

[12] N Abbas Y Zhang A Taherkordi and T Skeie ldquoMobile EdgeComputing A Surveyrdquo IEEE Internet of Things Journal vol 5no 1 pp 450ndash465 2018

[13] Guangshun Li Jianrong Song Junhua Wu and Jiping WangldquoMethod of Resource Estimation Based on QoS in Edge Com-putingrdquo Wireless Communications and Mobile Computing vol2018 Article ID 7308913 9 pages 2018

[14] D Happ and A Wolisz ldquoTowards gateway to Cloud offloadingin IoT publishsubscribe systemsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 101ndash106 Valencia Spain May 2017

[15] G Premsankar M Di Francesco and T Taleb ldquoEdge Comput-ing for the Internet of Things A Case Studyrdquo IEEE Internet ofThings Journal vol 5 no 2 pp 1275ndash1284 2018

[16] L Zhao W Sun Y Shi and J Liu ldquoOptimal Placementof Cloudlets for Access Delay Minimization in SDN-BasedInternet of Things Networksrdquo IEEE Internet of Things Journalvol 5 no 2 pp 1334ndash1344 2018

[17] SWang X Zhang Y Zhang LWang J Yang andWWang ldquoASurvey onMobile Edge Networks Convergence of ComputingCaching and Communicationsrdquo IEEE Access vol 5 pp 6757ndash6779 2017

[18] Y Liu J E Fieldsend and G Min ldquoA Framework of FogComputing ArchitectureChallenges and Optimizationrdquo IEEEAccess vol 5 pp 25445ndash25454 2017

[19] ETSI Multi-access Edge Computing [Online] Availablehttpwwwetsiorgtechnologies-clusterstechnologiesmulti-access-edge-computing

[20] OpenFog Consortium [Online] Available httpswwwopen-fogconsortiumorg

[21] Y Mao C You J Zhang K Huang and K B Letaief ldquoA SurveyonMobile Edge ComputingThe Communication PerspectiverdquoIEEE Communications Surveys amp Tutorials vol 19 no 4 pp2322ndash2358 2017

[22] Cisco ldquoThe Internet of Things How the Next Evolution ofthe Internet Is Changing Everything white paperrdquo [Online]Available httpswwwciscocomcdamen usIoT IBSG0411FINALpdf

[23] IDC ldquoFutureScape Worldwide Internet of Things 2017 Pre-dictionsrdquo November 2016 [Online] Available httpswwwidccomgetdocjspcontainerId=US41910716

12 Wireless Communications and Mobile Computing

resource utilization to achieve a rapid deployment of VNFand to meet service on-demand requests The combinationof resource estimation task scheduling and lightweight VNFconfiguration design provides an integrated solution that cansatisfy the service on-demand requests for 5G networks Thesimulation results showed that the proposedGAFmechanismoutperforms the other scheduling algorithmsMeanwhile thecomparison revealed that using the lightweight virtualizationtechnology in edge gateways ismore efficient and competitivethan using traditional VMs

In the future we plan to study the multiqueue taskscheduling problem for edge computing the possible coop-eration between edge gateway and cloud servers and the useof SD-WAN technology to achieve a seamless operation of thecloud-edge network

Data Availability

The data used to support the findings of this study areincluded within the article

Disclosure

The funder had no role in the study design data collectionand analysis decision to publish or preparation of thismanuscript

Conflicts of Interest

The authors declare no conflicts of interest regarding thepublication of this paper

Acknowledgments

The work described in this paper was supported in part bythe Ministry of Science and Technology of the Republic ofChina (Grants nos 104-2221-E-008-039 105-2221-E-008-071107-2623-E-008-002 and 107-2636-E-003-001)

References

[1] J Matias J Garay N Toledo J Unzilla and E Jacob ldquoTowardan SDN-enabled NFV architecturerdquo IEEE CommunicationsMagazine vol 53 no 4 pp 187ndash193 2015

[2] Q Duan N Ansari and M Toy ldquoSoftware-defined networkvirtualization An architectural framework for integrating SDNand NFV for service provisioning in future networksrdquo IEEENetwork vol 30 no 5 pp 10ndash16 2016

[3] TSI ldquoNetwork Functions Virtualization (NFV) architecturalframeworkrdquo[Online] Available httpswwwetsiorgdeliveretsi gsNFV001 099002010101 60gs NFV002v010101ppdf

[4] ONF ldquoSDN Architecturerdquo [Online] Available httpwwwopennetworkingorgimagesstoriesdownloadssdn-resourcestechnical-reportsTR SDN ARCH 10 06062014pdf

[5] A Basta A Blenk K Hoffmann H J Morper M Hoffmannand W Kellerer ldquoTowards a Cost Optimal Design for a 5GMobile Core Network Based on SDN and NFVrdquo IEEE Trans-actions on Network and Service Management vol 14 no 4 pp1061ndash1075 2017

[6] F Z Yousaf M Bredel S Schaller and F Schneider ldquoNFV andSDN-Key technology enablers for 5G networksrdquo IEEE Journal

on Selected Areas in Communications vol 35 no 11 pp 2468ndash2478 2017

[7] A C Baktir A Ozgovde and C Ersoy ldquoHow Can EdgeComputing Benefit FromSoftware-DefinedNetworking A Sur-vey Use Cases and Future Directionsrdquo EEE CommunicationsSurveys amp Tutorials vol 19 no 4 pp 2359ndash2391 2017

[8] Guangshun Li Jiping Wang Junhua Wu and Jianrong SongldquoData Processing Delay Optimization in Mobile Edge Comput-ingrdquoWireless Communications andMobile Computing vol 2018Article ID 6897523 9 pages 2018

[9] Zhimin Wang Qinglin Zhao Fangxin Xu Hongning Daiand Yujun Zhang ldquoDetection Performance of Packet Arrivalunder Downclocking for Mobile Edge Computingrdquo WirelessCommunications and Mobile Computing vol 2018 Article ID9641712 7 pages 2018

[10] W Yu F Liang X He et al ldquoA Survey on the Edge Computingfor the Internet of Thingsrdquo IEEE Access vol 6 pp 6900ndash69192018

[11] P Mach and Z Becvar ldquoMobile Edge Computing A Survey onArchitecture and Computation Offloadingrdquo IEEE Communica-tions Surveys amp Tutorials vol 19 no 3 pp 1628ndash1656 2017

[12] N Abbas Y Zhang A Taherkordi and T Skeie ldquoMobile EdgeComputing A Surveyrdquo IEEE Internet of Things Journal vol 5no 1 pp 450ndash465 2018

[13] Guangshun Li Jianrong Song Junhua Wu and Jiping WangldquoMethod of Resource Estimation Based on QoS in Edge Com-putingrdquo Wireless Communications and Mobile Computing vol2018 Article ID 7308913 9 pages 2018

[14] D Happ and A Wolisz ldquoTowards gateway to Cloud offloadingin IoT publishsubscribe systemsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 101ndash106 Valencia Spain May 2017

[15] G Premsankar M Di Francesco and T Taleb ldquoEdge Comput-ing for the Internet of Things A Case Studyrdquo IEEE Internet ofThings Journal vol 5 no 2 pp 1275ndash1284 2018

[16] L Zhao W Sun Y Shi and J Liu ldquoOptimal Placementof Cloudlets for Access Delay Minimization in SDN-BasedInternet of Things Networksrdquo IEEE Internet of Things Journalvol 5 no 2 pp 1334ndash1344 2018

[17] SWang X Zhang Y Zhang LWang J Yang andWWang ldquoASurvey onMobile Edge Networks Convergence of ComputingCaching and Communicationsrdquo IEEE Access vol 5 pp 6757ndash6779 2017

[18] Y Liu J E Fieldsend and G Min ldquoA Framework of FogComputing ArchitectureChallenges and Optimizationrdquo IEEEAccess vol 5 pp 25445ndash25454 2017

[19] ETSI Multi-access Edge Computing [Online] Availablehttpwwwetsiorgtechnologies-clusterstechnologiesmulti-access-edge-computing

[20] OpenFog Consortium [Online] Available httpswwwopen-fogconsortiumorg

[21] Y Mao C You J Zhang K Huang and K B Letaief ldquoA SurveyonMobile Edge ComputingThe Communication PerspectiverdquoIEEE Communications Surveys amp Tutorials vol 19 no 4 pp2322ndash2358 2017

[22] Cisco ldquoThe Internet of Things How the Next Evolution ofthe Internet Is Changing Everything white paperrdquo [Online]Available httpswwwciscocomcdamen usIoT IBSG0411FINALpdf

[23] IDC ldquoFutureScape Worldwide Internet of Things 2017 Pre-dictionsrdquo November 2016 [Online] Available httpswwwidccomgetdocjspcontainerId=US41910716

Wireless Communications and Mobile Computing 13

[24] T Salah M J Zemerly C Y Yeun M Al-Qutayri and Y Al-Hammadi ldquoPerformance comparison between container-basedand VM-based servicesrdquo in Proceedings of the 20th Conferenceon Innovations in Clouds Internet and Networks ICIN 2017 pp185ndash190 Paris France March 2017

[25] Docker Hub[Online] Available httpshubdockercom[26] C-W Tseng M-S Tsai Y-T Yang and L-D Chou ldquoA Rapid

Auto-Scaling Mechanism in Cloud Environmentrdquo in Proceed-ings of theThe 13th International Conference on Grid Cloud andCluster Computing Las Vegas Nev USA 2017

[27] B I Ismail E Mostajeran Goortani M B Ab Karim etal ldquoEvaluation of Docker as Edge computing platformrdquo inProceedings of the 2015 IEEE Conference on Open Systems(ICOS) pp 130ndash135 Melaka Malaysia August 2015

[28] R Morabito and N Beijar ldquoEnabling Data Processing at theNetwork Edge through Lightweight Virtualization Technolo-giesrdquo in Proceedings of the 2016 IEEE International Conferenceon Sensing Communication andNetworking SECONWorkshops2016 UK June 2016

[29] T Lin B Park H Bannazadeh and A Leon-Garcia ldquoDemoAbstract End-to-end orchestration across sdi smart edgesrdquo inProceedings of the 1st IEEEACM Symposium on Edge Comput-ing SEC 2016 pp 127-128 USA October 2016

[30] A Amjad F Rabby S Sadia M Patwary and E Benkhe-lifa ldquoCognitive Edge Computing based resource allocationframework for Internet of Thingsrdquo in Proceedings of the 2ndInternational Conference on Fog and Mobile Edge ComputingFMEC 2017 pp 194ndash200 Spain May 2017

[31] C Fan H Deng F Wang S Wei W Dai and B Liang ldquoAsurvey on task schedulingmethod in heterogeneous computingsystemrdquo in Proceedings of the 8th International Conference onIntelligentNetworks and Intelligent Systems ICINIS 2015 pp 90ndash93 Tianjin China November 2015

[32] P Akilandeswari and H Srimathi ldquoSurvey and analysis on taskscheduling in cloud environmentrdquo Indian Journal of Science andTechnology vol 9 no 37 2016

[33] E Meriam and N Tabbane ldquoA survey on cloud computingscheduling algorithmsrdquo in Proceedings of the 2016 Global Sum-mit on Computer and Information Technology GSCIT 2016 pp42ndash47 Sousse Tunisia July 2016

[34] H Wang J Gong Y Zhuang H Shen and J LachldquoHealthEdge Task scheduling for edge computing with healthemergency andhuman behavior consideration in smart homesrdquoin Proceedings of the 2017 IEEE International Conference on BigData (Big Data) pp 1213ndash1222 Boston MA USA December2017

[35] J Liu Y Mao J Zhang and K B Letaief ldquoDelay-optimal com-putation task scheduling for mobile-edge computing systemsrdquoin Proceedings of the 2016 IEEE International Symposium onInformation Theory (ISIT) pp 1451ndash1455 Barcelona Spain July2016

[36] RMijumbi J Serrat J L Gorricho N Bouten F De Turck andS Davy ldquoDesign and evaluation of algorithms for mapping andscheduling of virtual network functionsrdquo in Proceedings of the1st IEEE Conference on Network Softwarization (NetSoft rsquo15) pp1ndash9 University College London London UK April 2015

[37] P Samal and P Mishra ldquoAnalysis of variants in Round RobinAlgorithms for load balancing in Cloud Computingrdquo Interna-tional Journal of Computer Science and Information Technolo-gies vol 4 no 3 pp 416ndash419 2013

[38] M Chowdhury M R Rahman and R Boutaba ldquoViNEYardvirtual network embedding algorithms with coordinated node

and linkmappingrdquo IEEEACMTransactions on Networking vol20 no 1 pp 206ndash219 2012

[39] M G Rabbani R P Esteves M Podlesny G Simon L ZGranville and R Boutaba ldquoOn tackling virtual data centerembedding problemrdquo in Proceedings of the 2013 IFIPIEEEInternational Symposium on Integrated Network ManagementIM 2013 pp 177ndash184 Belgium May 2013

[40] A Fischer J F Botero M T Beck H De Meer and XHesselbach ldquoVirtual network embedding a surveyrdquo IEEE Com-munications Surveys amp Tutorials vol 15 no 4 pp 1888ndash19062013

[41] J Gil Herrera and J F Botero ldquoResource Allocation in NFVA Comprehensive Surveyrdquo IEEE Transactions on Network andService Management vol 13 no 3 pp 518ndash532 2016

[42] D Gross J F Shortle and J M Thompson Fundamentals ofQueueing Theory John Wiley amp Sons New York NY USA 4thedition 2008

International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Active and Passive Electronic Components

VLSI Design

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Shock and Vibration

Hindawiwwwhindawicom Volume 2018

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Journal ofEngineeringVolume 2018

SensorsJournal of

Hindawiwwwhindawicom Volume 2018

International Journal of

RotatingMachinery

Hindawiwwwhindawicom Volume 2018

Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Navigation and Observation

International Journal of

Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom