Mobile Cloud-Based Big Healthcare Data Processing in Smart ... · Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities which is a resourceful local

SPECIAL SECTION ON ADVANCES OF MULTISENSORY SERVICES ANDTECHNOLOGIES FOR HEALTHCARE IN SMART CITIES

Received May 6, 2017, accepted May 13, 2017, date of publication May 25, 2017, date of current version July 17, 2017.

Digital Object Identifier 10.1109/ACCESS.2017.2707439

Mobile Cloud-Based Big Healthcare DataProcessing in Smart CitiesMD. MOFIJUL ISLAM1, MD. ABDUR RAZZAQUE1, (Senior Member, IEEE),MOHAMMAD MEHEDI HASSAN2, (Member, IEEE), WALAA NAGY ISMAIL2,AND BIAO SONG2, (Member, IEEE)1Department of Computer Science and Engineering, University of Dhaka, Dhaka 1000, Bangladesh2Information Systems Department, College of Computer and Information Sciences, King Saud University, Riyadh 11543, Saudi Arabia

Corresponding author: Md. Abdur Razzaque ([email protected])

This work was supported by the Deanship of Scientific Research at King Saud University through the Research Groupunder Grant RGP-281.

ABSTRACT In recent years, the Smart City concept has become popular for its promise to improve thequality of life of urban citizens. The concept involves multiple disciplines, such as Smart health care, Smarttransportation, and Smart community. Most services in Smart Cities, especially in the Smart healthcaredomain, require the real-time sharing, processing, and analyzing of Big Healthcare Data for intelligentdecision making. Therefore, a strong wireless and mobile communication infrastructure is necessary toconnect and access Smart healthcare services, people, and sensors seamlessly, anywhere at any time. In thisscenario, mobile cloud computing (MCC) can play a vital role by offloading Big Healthcare Data relatedtasks, such as sharing, processing, and analysis, frommobile applications to cloud resources, ensuring qualityof service demands of end users. Such resource migration, which is also termed virtual machine (VM)migration, is effective in the Smart healthcare scenario in Smart Cities. In this paper, we propose an antcolony optimization-based joint VM migration model for a heterogeneous, MCC-based Smart Healthcaresystem in Smart City environment. In this model, the user’s mobility and provisioned VM resources in thecloud address the VM migration problem. We also present a thorough performance evaluation to investigatethe effectiveness of our proposed model compared with the state-of-the-art approaches.

INDEX TERMS Smart health care, smart city, big data, quality of service (QoS), virtual machine migration,ant colony optimization.

I. INTRODUCTIONDue to recent advancements in Information and Commu-nication Technology, the Smart City concept has becomean excellent opportunity to improve the quality of everydayurban life activities [2]. By connecting Smart objects, people,and sensors various services can be provided, such as Smarthealthcare, Smart transportation, and Smart community [3].Most Smart City services, especially emerging Smart health-care services, demand anywhere, anytime real-time com-putation. Critical patient monitoring, telemedicine, patientdata collection, and personalized medical services [11],[17], [35], [40] are major applications in this domain.These healthcare services and applications generate copiousamounts of Big Healthcare Data in real-time, thus requiringcomputational resources to be made available nearby [9].However, without a strong wireless and mobile communi-cation infrastructure, it is difficult to connect and access

computational resources for processing, sharing, and analyz-ing of Big Healthcare Data with minimum latency [27], [43].

The emergence of mobile cloud computing (MCC) [10],[25], [34] facilitates reduction of task-execution time andreal-time communication latency for such Smart Health-care applications in the Smart City environment [1], [7],[12], [28], [37]. MCC effectively utilizes distributed cloudserver resources such as CPU, memory, network, andports to execute the mobile Smart healthcare applications.Using MCC, mobile devices (MD) can offload these appli-cations to a resourceful cloud server for faster execu-tion [6], [8], [13], [16]. Moreover, MCC allows real-timequery processing in Smart healthcare applications that is vitalfor patients life. However, the long distance between cloudservers and the MDs may increase the response time forinteractive applications, increasing the total execution time.To alleviate this problem, a cloudlet is proposed in [36],

VOLUME 5, 20172169-3536 2017 IEEE. Translations and content mining are permitted for academic research only.

Personal use is also permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

11887

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

which is a resourceful local cloud that brings remote cloudresources closer to the mobile user. By offloading tasks to thenearest cloudlet, the user can decrease total task executiontime.

Virtualization technology introduces a middle layerbetween the hardware and software layers in a cloudlet,allowing the hardware resources to be shared by meansof VM. Resources (e.g., CPU, memory, network band-width, etc.) in a cloudlet are provisioned to these VMs.Resource provisioning in cloud computing is a well-studiedarea [4], [30], [39], [41]. However, themobility inMCC intro-duces several challenges to maintain an acceptable Qualityof Service (QoS) when provisioning cloud resources. Mobileusers may move from one Access Point (AP) to another,increasing their distances between current locations and thecloudlet, where the tasks are provisioned. This increases thetask-execution time. To address this issue, we propose aVM migration technique for a heterogeneous MCC systemfollowing the user’s mobility pattern. That is, when a usermoves from one cloudlet to another cloudlet, the resource orVMmust bemigrated to the cloudlet that is nearest to the user.

Consider the following scenario: a blind user is executingan application that takes an image from his surroundings.Then, the application processes the image in the cloudlet andgives a response to the user’s local client. That is, the appli-cation continuously uploads some data and the cloud serverprocesses this data to provide responses back to the user. Now,if the blind user moves away from the current cloudlet, thenhe or she will experience a delayed response from the mobileapplication executing in the cloudlet, degrading the overallperformance of the application. To avoid this performancedegradation, it is necessary for the system to adopt a VMmigration method to choose a cloudlet that is currently closerto the user to which to migrate the VM.

User mobility is not the only reason forcing a VM tomigrate. Migration can be initiated to minimize the over-provisioned resources and thus improve the overall systemobjectives. For instance, if a VM is required to be migratedfrom a cloudlet to any of the candidate cloudlets, the newcloudlet may not have the same type of VM. In that case,a VM with more resource than the current one must bechosen and provisioned in order to migrate the VM and thusminimize task-execution time. However, in this approach,the VM migration is provisioned more resources than therequired. Therefore, this over-provisioned resources greatlydecreases the system objectives, as it reduces the number ofprovisioned VMs in the cloudlets. Furthermore, the joint VMmigration approach, where a set of VMs is remapped based onthe VM task execution time and over-provisioned resources,can help to effectively increases the overall system objectives.In contrast to the joint VM migration approach, single VMmigration can only improve a particular user objectives butnot the system objectives.

In the literature, some researches have been conductedon VM migration in MCC [14], [31], [38], [42]. However,these methods are more suited for static task-execution.

Furthermore, most state-of-the-art works have only consid-ered a single VM migration approach. Though, this can helpto improve a particular user’s objectives, it can negativelyimpact not only the cloudlet system but also other users.Therefore, a joint VM migration model that effectively con-siders user mobility and the over-provisioned resources of thecloudlet has yet to be proposed in the literature.

In this work, we propose a VM migration (VMM) modelbased not only on user mobility but also on load of cloudletresources. The objective is to select the optimal cloud serverfor a mobile VM in addition to minimizing the total numberof VM migrations, reducing task-execution time. We useAnt Colony Optimization (ACO) to identify the optimal tar-get cloudlet. The main contributions of our work are statedbellow:

• We develop an Ant Colony Optimization (ACO)-basedVM migration model, in which VM are migrated tocandidate cloud servers so as tomaximize the total utilityof the MCC system.

• Mobility-aware selection of cloudlets for VM provision-ing in our proposed PRIMIO system helps significantlyto reduce service provisioning time.

• We introduce a joint VMmigration approach to optimizeboth the resource utilization and task execution time,diminishing the shortcomings of a single VM migrationapproach.

• The results of performance evaluation, depicted fromtest-bed implementation and extensive experiments,show that the proposed PRIMIO system achieves signif-icant improvements compared to state-of-the-art works.

The remaining paper is organized as follows.In Section II, we discuss on the state-of-the-art work on VMor task migration inMCC. Next, in Section III, we present thenetwork model and assumptions. After that, in Section IV, wepresent the problem formulation of our VMmigration model.In Section V-B, we present a meta-heuristic ACO-based VMmigration solution for jointly migrating a set of VMs. Sub-sequently, in Section VI, we present a numerical evaluationof our proposed model to investigate it’s effectiveness of ourproposed model compared to the state-of-the-art. Finally, wedraw conclusions andmention directions for our future worksin Section VII.

II. RELATED WORKWith the advent of the smart City, smart healthcare ser-vices are emerging to improve urban citizens’ quality oflife [18]–[23]. However, to connect and access smart health-care services, people, and sensors, seamlessly, anywhere andany time, MCC plays a vital role. In MCC services, users canoffload (-part of) a task in cloudlet for faster execution. Exist-ing literature discusses several methodologies to detect whena client should offload task in a cloudlet [4], [6], [8], [13],[30], [39], [41]. However, little works has been conductedin the field of VM migration in MCC systems. A traffic-aware, cross-site VM migration model is proposed in [31].

11888 VOLUME 5, 2017

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

In this model, when multiple VMs require migration, anarbitrary sequence of VM migration congests the bandwidthof inter-site links, thus reducing the number of successful VMmigrations. Then, the VMmigration problem is formulated asa Mixed Integer Linear Programming (MILP) problem anda heuristic algorithm is used to get an approximate optimalresult.

Besides this, a mobility induced service migration forMCC is proposed in [42]. In this work, a threshold-basedoptimal service migration policy is developed, where theVM migration problem is modeled as a Markov decisionprocess (MDP). This proposed system considered mobileusers to follow a one-dimensional asymmetric random walkmobility model. A service is migrated from one micro-cloudto another when a user is in states bounded by a set ofpredefined thresholds.

In addition, three lightweight task-migration models aredeveloped in [14]:

• Cloud-wide task migration, where the task-migrationdecision is made by a central cloud, which maximizesthe objectives of a cloud provider.

• Server-centric task migration, where all migration deci-sions are made by the server, where the task is currentlyexecuting .

• Task-based migration, where migration is initiated bythe task itself.

In this approach, the migration decision is made after eachdecision epoch, based on user’s mobility and remaining task-execution time. This proposed method considers the increas-ing data volume transfer time during task migration from onecloud to another.

Meanwhile, in our previous work, we proposed a mobility-and load-aware Genetic Algorithm-based VM migrationapproach, GAVMM [24], to minimize task-execution time.However, this approach mainly tries to minimize the pro-visioned task execution time without considering over-provisioned resources in the cloudlets. Thus, the GAVMMfails to minimize resource over-provisioning in the cloudlets.

However, all state-of-the-art works use the single VMmigration approach, where a cloudlet migrates a single VMto another cloudlet. This approach relaxes the problem for-mulation but fails to effectively optimize the whole-systemobjectives. Instead, we here use a joint VM migration, wherea set of VMs jointly migrates to a set of coudlets, allowingeffective optimization of resource usage and task-executiontime.

In summary, most VM migration methodologies do noteffectively consider user mobility alongside load condi-tion of cloudlet servers in a heterogeneous MCC sys-tem. This increases service downtime, especially for thoseapplications where the user frequently interacts with theprovisioned cloudlet. In addition, when migrating a VM,over-provisioned resources in the target cloudlet must also beconsidered; otherwise, the total number of provisioned VMsin the target cloudlet will greatly be reduced. To the best of

our knowledge, this work is the first to efficiently utilize ACOsystem to develop a VM migration approach for minimizingthe task-execution time and optimizing the cloudlets resourceusage. Moreover, we extend the VM migration model tojointly migrate a set of VMs to a set of cloudlets in orderto minimize task-execution time and to minimize resourceover-provisioning compared to single VM-based migrationapproaches.

III. MOBILE CLOUD SYSTEM ARCHITECTUREA. SYSTEM ARCHITECTUREWe assume a three-tier Mobile Cloud Computing (MCC)environment, where a set of M access points (APs) com-prise the backbone network. Tier one represents the mas-ter cloud, which consists of several public cloud providers,such as Google App Engine, and Microsoft Azure AmazonEC2. A set of high-speed interconnected cloudlets consti-tute the tier two or the backbone layer of the mobile cloudarchitecture. Smartphones, wearable devices or other mobiledevices constitute the tier three or user layer. Users accessthe nearest cloud resources using devices from tier three.Each AP is connected to any of C cloudlets, denoted asC = {C1,C2,C3.....,CL}. We present a sample networkscenario in Fig. 1. These cloudlets are connected using thebackhaul network where bandwidth between cloudlets i and jis denoted as Bi,j.

FIGURE 1. Mobile cloud architecture.

A set of cloudlets is controlled and monitored by themaster cloud (MC). All cloudlets route their hypervisorinformation to the master clouds and they are connected tothe MC with a high-speed network connection. A cloudletserver i has fixed processing power Spk and memory Mk .Each cloudlet provisions N number of VMs, denoted asV = {V1,V2,V3.....,VN }. Users can offload their tasks toa dedicated VM. We further assume that each user is mobileand execute a task in different VMs over the task lifetime.We assume that there are no inter-dependencies among thoseVMs. In the remainder of this paper, we use task and VMinterchangeably.

VOLUME 5, 2017 11889

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

FIGURE 2. Virtual machine migration system.

B. VIRTUAL MACHINE MIGRATION SYSTEMIn Fig. (2), a computation migration system is depicted,which has two important functionality. First, the computa-tion migration between resource-constrained mobile devicesand cloudlets; Second, virtual machine migration betweentwo cloudlets to minimize the task execution time as wellas resource over-provisioning. The application offloadingmanager helps a mobile device to decide which part of themobile application should be offloaded and where [6], [8],[26], [30]. The client-request handler in a cloudlet managesincoming task execution requests from mobile devices anddispatches to the virtual machine scheduler. The VM sched-uler uses the information from hypervisor to initiate the VMscheduling. Moreover, after a specific time epoch, the VMmigration manager interacts with the VM scheduler to selecttasks and to migrate so as to reduce the task-execution time.In this work, our main concern is to develop an algorithmfor the VM migration manager for selecting a set of tasksand cloudlets, remapping tasks to minimize execution time.In the VM migration manager, there are three maincomponents-VM Selection, Joint VMMigration Engine, andVM Migration Initiator, which contribute to making optimalmigration decisions. Moreover, the user mobility managertakes local mobility management information in conjunctionwith global mobility prediction manager to decide on a setof probable cloudlets for a user. Finally, the VM migrationmanager initiates migration process for a set of tasks requir-ing faster execution.

C. ASSUMPTIONSWe assume that a cloudlet’s VM scheduler determines theamount of resources to allocate to a certain VM during itscreation and that it is allowed to update the VM size dur-ing scheduling intervals, if deemed necessary in support ofmeeting user QoS or increasing resource utilization [5], [33].We also assume that the pre-copy live VM migrationmethod [29] is used by the cloudlets, which allowsmigration of internal state, memory and application data

TABLE 1. Notations.

associated with a VM from its provisioned cloudlet to adestination one.

Moreover, in this work, we assume that users will havewalking mobility speeds. We employ a state-of-the-art usermobility prediction approach, ENDA [30], to determine thepredicted cloudlets set for each provisioned VM. However,the problem of thrashing might appear for users with highmobility speeds (e.g., vehicular mobility), which we do notconsider in this work. Notations and their descriptions arelisted in Table 1.

IV. MULTI-OBJECTIVE PROBLEM FORMULATIONMobile Cloud Computing (MCC), mobile devices offloadthe most resource intensive applications such as image pro-cessing applications, augmented reality applications etc., tothe nearest cloudlet to minimize total task-execution time.However, when a mobile user moves from one cloudlet toanother, the VM must be migrated and provisioned intoanother cloudlet to minimize the interaction time betweenthe mobile user and the provisioned VM. The resulting totalexecution time helps to meet the task deadline. When a VMis selected for migration, a cloudlet must choose that VM insuch a way to also minimize the resource over-provisioning.Therefore, a VM must be mapped to another cloudlet VM.There are several approaches to map and provisioned a VMto that of new cloudlet. Migrating a single VM based on thepreferences of a particular mobile user can help maximize theobjectives of just one user, but it fails to improve the overallsystem objective and also can have negative impact on otheruser’s objectives. Therefore, in our proposed approach, wejointly consider all VMs in a cloudlet remapping those to aset of cloudlets, if necessary.

11890 VOLUME 5, 2017

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

In the proposed VM-migration problem, both the total taskexecution time and amount of over-provisioned resourcesin the cloudlets must be minimized. Thus, this is a multi-objectives problem, where a set of VMs V are remapped toa set of cloudlets C . The proposed VM-migration problemhas two objectives: (1) to minimize total task-execution timeand (2) to minimize the resources wasted by the cloudlets.We can formulate our multi-objective VMmigration problemas follows,

Minimize : Z =∑u∈V si

∑v∈V cj,u

{α × Tu,v + (1− α)× Ro,iu,v

}× xu,v (1)

s.t.∑v∈V cj,u

xu,v ≤ 1 ∀u ∈ V si (2)

∑u∈V si

xu,v ≥ 0 ∀v ∈ V cj,u (3)

(ur ≤ vr

)∀r ∈ Rui , u ∈ V

si , v ∈ V

cj,u (4)

Tu,v × Tmaxu,v × xu,v ≤ Tdu ∀u ∈ V

si , v ∈ V

cj,u (5)

where the objectives in Eq.(1) represent the minimization ofthe total task-execution time of provisioned VMs and theminimization of resource over-provisioning in the cloudlets.xu,v is a binary variable which will be 1 when the VM uin cloudlet i is provisioned to VM v in cloudlet j; α is theweight parameter of application-types, defining how muchweight each application has in reducing task-completion timeor in optimizing system resources. Tu,v is the normalizedtotal task-execution time when VM u in cloudlet i was pro-visioned to VM v in cloudlet j, which can be defined asfollows,

Tu,v =

(T eu,v + T

tu,v)

Tmaxu,v(6)

Tmaxu,v = maxb∈V cj,u

(T eu,b + T

tu,b)

(7)

where T eu,v is the task execution time if the VM u is migratedand provisioned to VM v and T ti,j denotes the VM transfertime from cloudlet i to cloudlet j. T ti,j can be defined asfollows,

T tu,v =DuBu,v+ T q,du,v (8)

where Du is the data associated with the VM u, whichincludes both VM state and application data; Bu,v denotesthe bandwidth of the link between the cloudlets runningVMs u and v. In addition, T q,du,v denotes the VM executionqueuing time for the VM u in VM v. In the proposed VM-migration approach, the full VM image is not migrated, butrather only the data which are generated by the provisionedVM. Because, in our VM-migration approach, the applicationin the migrated VM is stopped and later resumed on themigrated cloudlet upon completion of the migration opera-tion.

In Eq.(1), Ro,iu,v denotes the normalized over-provisionedresource, when a VM u of cloudlet i is migrated to a VM v.It can be defined as follows,

Ro,iu,v =1|Rui |

∑r∈Ru

vr − ur

ur(9)

where Rui are the resources required by the VM u in thecloudlet i. In addition ur and vr are the resourcea of type r ,provided by the VMs u and u, respectively.

Constraint (2) ensures that one VM is provisioned ormigrated to one and only one VM in a particular cloudlet.In addition, constraint (4) ensures that when the systemmigrates a VM u to a new VM v in another cloudlet, the newVM v has at least the same amount of resources in currentlyprovisioned VM in the cloudlet i. Furthermore, constraint (5)ensures that a VM can be migrated to a cloudlet where thetask execution deadline will be met.

V. META-HEURISTIC VIRTUAL MACHINE MIGRATIONNot all VMs in a cloudlet provisioned mobile-device appli-cations require migration to meet the user task-executiondeadlines or to increase the system objectives. Rather,migration of a subset of VMs could improve user expe-riences. Therefore, as a first step, the proposed PRIMIO,PRioritized meta-heurIstic virtual Machine migratIOn, sys-tem develops a candidate set of migratable VMs, V s,that can reduce applications’ overall execution time. Then,the proposed VM migration policy employs a VM map-ping procedure to select an appropriate cloudlet forprovisioning VMs.

A. SELECTING CANDIDATE SET OF VMS FOR MIGRATIONFor each provisioned VM u ∈ V , our proposed PRIMIOsystem calculates an urgency factor Uu that determines thecriticality of migrating VM. Before calculating the urgencyfactor, we have to determine the predicted candidate cloudletset,V cj,u for the VM u, from which users can get their services.There are several algorithms in the literature for predictingthe user location in the upcoming VM schedule; in thiswork, we use ENDA [30]. Recall that the key notion ofa VM migration is to minimize the service execution timeand hence to meet the service deadline of the hosted userapplication. Therefore, we calculate the urgency factor asfollows,

Uu = T p,ru × γup,r + T

p,du × γ

up,d , (10)

where the urgency factor is a linear combination of twoattributes of a task: first is the task-execution time requiredpenalty T p,ru and second is the task-execution time deadlinepenalty T p,du for a VM u. T

p,ru is the difference between

the task-execution time, which is required to execute in thecurrently provisioned VM, and the required task-executiontime. In the same way, T p,du can be defined as the differ-ence between the task-execution time and the task-executiondeadline time. These two task-execution time penalties can

VOLUME 5, 2017 11891

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

be calculated in the following way,

T p,ru = Teu − T

ru (11)

T p,du = Teu − T

du (12)

T ru =

∑i∈Aru,t=u(t)

T i,tr|Au,r |

(13)

where Au,r is the set of previous task execution time requiredby the VM u to execute task type t . T eu ,T

du and T

ru are

the task-execution time required to execute the task in theprovisioned VM, task-execution time deadline and requiredtask execution time, respectively. T du is set by user, whereT ru is defined by the system itself. In Eq.(10), γ

up,r and γ

up,d

are task-execution time penalty coefficients. These coeffi-cients defined how much weight we have to set to each taskexecution time penalties based on the system and the userpreferences. γ up,r and γ

up,d can be defined as,

γ up,r =

{1, T ru ≤ T

eu

T p,ru × βud,r , otherwise(14)

γ up,d =

{0, T eu ≤ T

du

T p,du × (1− βud,r ), otherwise(15)

where βud,r is the user application preference. With γup,r and

γ up,d , βud,r helps the system to determine when the task must

be migrate as well as gives the system a degree of flexibilityto migrate a VM from the cloudlet and provisioned in a newcloudlet. βud,r can defined as,

βud,r=

1, if d attribute is surely preferred than r(0.5, 1), if d attribute is partially preferred than r0.5, if no preference(0, 0.5), if r attribute is partially preferred than d0, if r attribute is surely preferred than d .

(16)

After that, VMs are sorted according to urgency factor indecreasing order. According to this order, candidate-VMselection algorithm chooses a set of VMs based on the avail-able resources in cloudlets accessible to each VMs. Neverthe-less, during this selection the VM migration system does notmap a VM into a cloudlet, but rather just selects a set of pro-visioned VMs from the cloudlet which need to be migrated.If the VM type or the required resources of a VM is ful-filledby any accessible VM, then it will be selected; otherwise,it must wait for the next scheduling interval. Therefore, thetotal VM resources are required by all the selected VMs mustbe less than or equal to the VM resources available in theaccessible cloulets. The prioritized VM selection method issummarized in Algorithm ( 1).

The proposed VM migration system requires each taskbe associated with two task execution deadlines: (1) task-execution deadline time, T du , which is set by the user appli-cation; and (2) the required task-execution time, T ru , which isdefined by the system itself. The task execution time, T eu , can

Algorithm 1 Candidate VM Set Selection Algorithm,at Cloudlet

INPUT: VM set V and accessible cloudlet setCu for eachVM u.OUTPUT: Candidates VM list for migration.

1: Calculate urgency priority Uu,∀u ∈ V2: Reorder the urgency priority Uu,∀u ∈ V in following

way,U1 ≥ U2 ≥ U3 ≥ ....UN

3: V s = ∅4: repeat

for each user VM u ∈ V5: for (for each VM v ∈ V cj,u) do6: if

(ur ≥ vr |r ∈ Rui

)then

7: Vvc = Vvc \ {v}8: V s = V s ∪ u9: end if10: end for11: until all VMs are provisioned or resources are utilized

be greater than the T ru ; however, it must be less than or equalto T du . These two task-execution times give the VMmigrationsystem flexibility in initiating VM migration. T ru , especially,helps the system to initiate the VM migration in order tominimize the service downtime. For instance, if T eu exceedsT ru , the system may initiate VM migration for that particularVM if enough resources are available. However, if the T eu in aVM does not exceed the T du but does exceed T

ru and if avail-

able resources are limited, then the VM migration systemgives more priority to those VMs which have higher need tobe migrated to another cloudlet. Consequently, this approachhas advantages for both the system and the user application,especially, when a small amount of time is needed to completeexecution while available resources are limited.

B. META-HEURISTIC ANT COLONY-BASEDVM MIGRATION ALGORITHMIn our proposed VMmigration approach, we jointly migrate aset of VMs to a set of cloudlets. That is, a cloudlet remaps a setof VMs to a set of cloudlets. This is a bin-packing problem,where a set of VMs is represented as bins while the cloudletsrepresent packs. These VMs must be packed with minimumnumber of cloudlets such that the total execution time of tasksis minimized; also, we have to minimize resouces wastedduring packing. Thus, it becomes an NP-hard problem [44],i.e., no algorithm can provide a guaranteed optimal solutionin polynomial time. For this reason, we have proposed ameta-heuristic Ant Colony based VMMigration algorithm to solvethe VM migration problem. We use this meta-heuristic algo-rithm since it is problem-independent and does not requirethe benefit of any specificity of the proposed problem.

As swarm-optimization models are well-suited to buildingoptimal solutions in very dynamic environments [15], weemploy Ant Colony Optimization (ACO) to incorporate the

11892 VOLUME 5, 2017

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

dynamically changing cloud computing environment in theselection of optimal VM-cloudlet pair. Also, in a distributedenvironment, ACO helps to speedup computation, making itmore suitable for the VMmigration decision process. In addi-tion, each cloudlet makes VM migration decisions individu-ally by exploiting the information collected from neighboringcloudlets, mobile devices, and the master cloud. Furthermore,the ACO’s learning capability of good solution features leadsto the optimal solution and greatly decreases the selectionprobability of poor solutions [32].

The ACO is a meta-heuristic algorithm that uses the behav-ior of virtual ants to build a heuristic solution. A set ofants are created, with each ant trying to build a solution bysolving sub-problems using heuristic information. The antstry to improve their solutions by exchanging informationvia pheromones. Each ant uses this pheromone trail and thelocal heuristic information to build a local optimal solution.Finally, all ants combine their local best solutions to infer anoptimal solution.

Algorithm 2 Ant Colony Based VMMigration, at Cloudlet iINPUT: VM set V and accessible cloudlet setCu for eachVM u. System ParametersOUTPUT: VM-cloudlet pair.

1: Initialize system tunning parameter α2: Initialize system migration parameters γl , γg3: Determine Candidate VM set V s using Algorithm 14: Initialize ants set A5: Generate initial solution using FFVM Algorithm 36: Calculate initial pheromone γ07: Set maximum iteration MAX_IT8: while ( do iteration ≤MAX_IT)9: for (Ant a ∈ A) do10: k = 011: repeat12: Select a VM v in cloudlet j for VM uk ∈ V s

using Eq. 2213: k = k+114: until every VM is provisioned to a cloudlet15: for (VM k ∈ V s) do16: Update the local pheromone using Eq. 2417: end for18: end for19: Update the global pheromone using Eq. 2520: iteration = iteration+121: end while22: Return VM-cloutlet pairs

1) PHEROMONES AND INITIAL PHEROMONE CALCULATIONIn ACO algorithm, pheromones represent the desirability ofchoosing a solution. In our proposed PRIMIO algorithm, thepheromones represent the desirability of assigning a VM toa cloudlet. Each ant starts with an initial pheromone valuefor each VM to cloudlet pair. The initial solution is generated

using a greedy First Fit(FF) VM migration approach, whichis listed in Algorithm 3.

Algorithm 3 First Fit VM Migration, at Cloudlet iINPUT: VM set V and accessible cloudlet setCu for eachVM u.OUTPUT: VM-cloudlet pair in initial solution.

1: for (VM u ∈ V) do2: for (VM v in V cj,u) do3: if (Tu,v ≤ TDu And ∀r∈Rui R

D,ru ≤ R

rv ) then

4: Assign VM i to cloudlet k5: Break6: end if7: end for8: end for

The initial pheromone value is calculated by summing thetotal task-execution times and the resources over-provisionedin each cloudlet and taking the inverse of the amount. Initialpheromone value for each ant is calculated as follows,

τ0 =∑u∈V si

∑v∈V cj,u

1Tu,v +

∑r∈Rui

Ro,ru,v× yu,v (17)

where yu,v is a binary variable, which is defined as,

yu,v =

{1, if (u, v) ∈ S00, otherwise

(18)

2) CALCULATION OF HEURISTIC VALUETo build a solution each ant uses the local heuristic value toselect a cloudlet for a VM. This heuristic value defines thefavorability of choosing a cloudlet for a VM to construct thesolution. As PRIMIO tries to jointly optimize the objectivesof total task-execution time and resource over-provisioningin cloudlets. So, the local heuristic value has to be defined insuch a way that the system can optimize the task executiontime and resource over-provisioning to reflect the systemobjectives. Local heuristic is defined as,

ηu,v = α × ηEu,v + (1− α) η

Ru,v (19)

where α is the system parameter defines the relative weightbetween the objectives of total task execution time andresource over-provisioning of cloutlets. It can be tunedaccording to the system environment. ηEi,j and η

Ri,j is the

objectives of total task execution time and the resource over-provisioning respectively, if VM migrates from cloudlet i tocloudlet j. ηEi,j and η

Ri,j are defined in Eq. (20) and Eq. (21),

respectively.

ηEu,v =

u∑k=1

∑v∈V uj,u

1Tk,v

(20)

ηRu,v =∑r∈Rui

u∑k=1

∑v∈V cj,u

(1− Ro,rk,v

)(21)

VOLUME 5, 2017 11893

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

3) CLOUDLET SELECTION: PSEUDO-RANDOMPROPORTIONAL RULEWhile constructing the solution, each ant selects a cloudletfor each VM using a pseudo-random proportional rule, whichcan be defined as follows,

v =

{argmaxk∈V cj,u

([τu,k

]α×[ηu,k

]β), if q ≤ q0

s, otherwise(22)

where q is a random numbers uniformly distribute in [0,1]and q0 is the system parameter on the range [0,1]. τi,k isthe pheromone value of selecting cloudlet k for VM i. Whenq ≤ q0, (i.e., exploitation) ant selects a cloudlet j for a VM i,where the multiplication of

[τi,k]α×[ηi,k

]β gives the highestvalue among the all possible choice of cloudlets. Here, α andβ are the system parameters, denoting the relative importanceof pheromone value and the local heuristic value for choosinga cloudlet j for a VM in the cloudlet i. If the random number qis greater than q0, i.e. in case of exploration, an ant z choose acloudlet j using the probability pzi,j, which can be defined as,

pzu,v =

([τu,k]α×[ηu,k]β

)∑

k∈Vcj,u

([τu,k]α×[ηu,k]β

) , if q ≤ q00, otherwise

(23)

4) LOCAL PHEROMONE UPDATEWhen an ant chooses a VM pair to construct a solution, itimmediately updates the local pheromone value in relation tothe initial pheromone value. Local pheromones are updatedby each ant using the following relation,

τu,v(t + 1) = (1− γl)× τu,v(t)+ γlτ0 (24)

where γl is the system parameters, denoting the relativeimportance of current pheromone value at time t , τu, v(t).

5) GLOBAL PHEROMONE UPDATEGlobal pheromone values are updated for each pair of VMand cloudlet onlywhen all ants constructed their local optimalsolutions and updated the global optimal solution. The globalpheromone values are updated using the following relation,

τu,v(t + 1) = (1− γg)× τu,v(t)+ γg1τu,v (25)

where the γg is the global pheromone system parameter, thisparameters can be tuned according to the system objectives.1τu,v is the global pheromone value for the updated globalsolution, which is defined as

1τu,v =

{τu,v, if (u, v) ∈ PGS0, otherwise

(26)

where PGS is the global solution set of the selected VM u formigration and the target VM v where the migration will beoccurred.

6) VM MIGRATION INITIATORAnt colony optimization optimally selects a cloudlet for aVM. After each VM is mapped to a particular cloudlet, theVMmigration initiator is invoked, which determines whetherthe PRIMIO algorithm selects the current cloudlet for the VMas a target cloudlet. If the target and current cloudlets for aVM are different then the VM migration manager initiatesa VM migration event. Otherwise, the VM is provisioned tothe current cloudlet and no VM migration event initiation isrequired.

VI. PERFORMANCE EVALUATIONIn this section, we assess the performance of our proposedPRIMIO method through test-bed experimentation. We com-pare the performance of proposed PRIMIO method with‘No Migration,’ ‘Task-centric migration’ [14], GAVMM [24]and mobility-based ‘greedy migration’ methods. In imple-menting the ‘No Migration’ method, VMs are not migratedeven though a cloudlet is overloaded or the user moves tocloudlets; the cloudlet, where the VM is provisioned, exe-cutes the task and forwards the result to the destinationcloudlet. In implementing ‘Task-centric migration,’ an indi-vidual task is migrated to another cloudlet, following thetask’s mobility pattern and cloudlet’s computational load, inorder to improve task execution time. ‘GAVMM’, employs agenetic algorithm to select a target cloudlet based on the usermobility and cloudlet sever load. On the other hand, GreedyVM Migration method migrates a VM to a cloudlet fromwhich the user is receiving the highest WiFi signal strength,without considering the load of that cloudlet initiatingVM migration.

A. EXPERIMENTAL ENVIRONMENTAs test bed to evaluate performance of PRIMIO, weused ten cloudlets, each with the following computationalresources [1.5-3.0]GHz processor, [8-16]GB memory and[250-350GB] SATA hard disk. All cloudlets were intercon-nected with [2-20]Mbps Ethernet links. Users access thecloudlets using mobile devices with varying computationcapacity. Each user device is connected with cloudlet APsthrough a WiFi IEEE 802.11g interface. In this experiment,we used two Galaxy Grand 2, two Sony Xperia Z, and sixSymphony Android location-enabled devices.

Tasks in the mobile devices are generated by followinga Poisson distribution. The tasks are text, a document, oran image which is uploaded to cloudlet for analysis. Afterthat, the result will be pushed back to the user’s mobiledevice. Here, α is a system parameter, the value of whichis determined based on the user task. In our experiment, weset the system parameter to 0.65 through rigorous simulationand also because the goal of task is to reduce the task-completion time. Based on rigorous numerical simulation andtask type, we set other experiment parameters ω0, γl and γg.The parameters used in testbed experiments are enlistedin Table 2.

11894 VOLUME 5, 2017

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

TABLE 2. Simulation parameters.

B. PERFORMANCE METRICSTo evaluate the performance of the proposed PRIMIOapproach to VM migration compared to state-of-the-art VMmigration methodologies, we consider the following metrics:• Average Task-Completion Time: calculated by averag-ing task execution time for a set of tasks executing inthe cloudlets using different VMmigration mechanisms.A VM migration method with decreasing average task-completion time increases the quality of user experience(QoE) by executing the task in the cloudlets rather thanexecuting in the mobile devices.

• Resources Over-Provisioned (%): measured by calcu-lating the amount of extra resources are provisionedcompared to the resources required. We measure thepercentage of over-provisioned resources using Eq. 9.If the percentage of over-provisioned resourcesincreases for a VM migration method, that degrades theoverall system performance.

• Successful Task Completion (%): calculated as the per-centage of tasks, which are provisioned to cloudlets andwhich successfully completed their executions withinthe predefined task execution deadline. Thismetric helpsto evaluate the probability that a VMmigration approachwill finish the execution of a task within deadline.

• VMMigration Overhead: is the ratio of the total numberof migrations are required to execute a set of tasks andthe total number of tasks finishing execution within thepredefined deadline. This metric indicates the perfor-mance of VM migration method in terms of utilizationof provisioned resources.

C. EXPERIMENTAL RESULTSWe studied the performance of our proposedPRIMIOmethodby varying user mobility, task lifetime, and data footprint.In simulation, we only considered a pedestrian walkingmodel.

1) IMPACTS OF USER MOBILITYWe gauge the impact of user mobility on average task-execution lifetime in cloudlets, over-provisioned resources,task-executions completed within deadline and the VMmigration overhead. To study the impact of user mobility, wevary user speed from [4.5 ∼ 7.0]Km/hr, which follows thepedestrian walking model.

Fig. 3 (a) shows that the average task lifetime increaseswith increased mobility speed, because it causes frequentmigration of VMs, increasing the service downtime. More-over, the communication latency between the user and thecloudlet is also increased. Hence, the graph indicates that NoMigration policy experiences higher average task-executiontime. However, our proposed VMmigration policy, PRIMIO,reduces the average task-execution lifetime as it migratesVMs based not only on the user mobility pattern but also itconsiders computational load of cloudlets. On the otherhand,other VMMigration approaches do not effectively utilize theuser mobility to initiate the VM migration decision.

Fig. 3 (b) depicts that the percentage of over-provisionedresources increases with the higher mobility speeds. As usermobility speed increases, the tasks in the cloudlets alsomigrate more frequently, thus migrating to VMs with greatercapacity than the required. The No VM migration approach,for example, does not increase the percentage of resourceover-provision, as it does not migrate any VM. Even though,all the VM migration approaches increase the resource over-provision, our proposed PRIMIO approach does not sig-nificantly increase the over-provisioned resources, becauseit considers a joint VM-migration approach. This helpsthe cloudlet system to minimize over-provisioned resourcesacross the system as a whole. By contrast, the task-based andgreedy VM migration approaches utilize a single VM migra-tion approach, making migration decisions without consider-ing the other provisioned VMs. Therefore, these VM migra-tion approaches increase the over-provisioned resource asuser mobility increases.

Fig. 3 (c) shows that the percentage of successful task exe-cution within deadline degrades with increased user mobil-ity speed. One possible explanation is that increased usermobility also increases communication delay with the VMprovisioned in a cloudlet and increase the task lifetime whichin turns miss the task execution deadline. For this reason,the No VM migration approach has a lower percentage oftasks executed within deadline. The PRIMIO VM migrationapproach migrates the VM to a cloudlet by following usermobility and cloudlet load. However, greedy and task-centricVM migration approaches can not effectively utilize usermobility. Moreover, the greedy VMmigration approach doesnot consider the load of the provisioned cloudlet.

Fig. 3 (d) depicts that increased user mobility speed alsoincreases VM migration overhead. We do not study the NoVM migration approach here, as it migrates no VM andthus incurs no migration overhead. Although, our proposedVM migration approach PRIMIO increases VM migrationoverhead with the increased user mobility speed, it doesnot increase significantly compared to others. This is fortwo main reasons. First, PRIMIO uses user mobility infor-mation in initiating VM migration decisions, thus decreas-ing task execution lifetime and increases the percentage ofsuccessful completed of task execution within the executiondeadline. Secondly, PRIMIO considers cloudlet load in theVM migration decision process, which effectively reduces

VOLUME 5, 2017 11895

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

FIGURE 3. Impacts of mobility speed. (a) Average task completion time. (b) Resource over-provisioned. (c) Successful taskcompletion. (d) VM migration overhead.

FIGURE 4. Impacts of average task completion time. (a) Resource over-provisioned. (b) VM migration overhead.

the total number of VM migrations across the whole tasklifetime. Even though greedy and task-centric VM migrationapproaches consider the user mobility and thus reduce thetotal number of VM migration, neither could significantlyincrease the percentage of successful task execution withindeadline, therefore failing to reduce VMmigration overhead.

2) IMPACTS OF TASK LIFETIMEWe measure the impacts of task lifetime on the percentageof over-provisioned resources and VM migration overhead

by varying the task lifetime on the cloudlets from [6 ∼ 14]minutes. During this simulation, we set user mobility speedto approximately 5.5 Km/h.

Fig. 4 (a) indicates that the percentage of over-provisionedresources increases with average task lifetime. Neverthe-less, the over-provisioned resources do not increase signif-icantly in PRIMIO, because, it employs a joint VM migra-tion approach that effectively utilizes the system resources.On the other hand, the Greedy VM migration approachincreases over-provisioned resources more rapidly as it only

11896 VOLUME 5, 2017

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

FIGURE 5. Impacts of data footprints. (a) Average task completion time. (b) Successful task completion.

considers the user mobility without looking at cloudletresources.Meanwhile, the task-centric VMmigrationmethodtakes a single VM migration approach, which fails toeffectively utilize cloudlet resources. Finally, the No VMmigration approach, eventhough it does not increase theover-provisioned resources, but it negatively impacts on otheraspects.

Fig. 4 (b) depicts that increased task execution lifetime inthe cloudlets also increases VM migration overhead. Obvi-ously, a task that executes for a longer period of time wouldincur more VM migration overhead. However, if migrationincreases rapidly with regard to task lifetime, it degrades bothsystem performance and user application’s quality of service.This graph indicates that, compared to our proposed PRIMIOVMmigration approach, other methodsmore greatly increaseVM migration overhead. Moreover, VM migration overheadin thePRIMIO approach reaches a steady state when task exe-cution lifetime is close to 10 minutes. Because, PRIMIO uti-lizes both the joint-VMmigration approach and user mobilityto initiate VM migration decisions.

3) IMPACTS OF DATA FOOTPRINTSWe studied impacts of data footprint, which is the data asso-ciated with the VM, on average task lifetime, successful taskexecution within deadline and VM migration overhead ofdifferent VMmigration approaches. During this study, we setthe usermobility speed to approximately 5.5Km/h.We variedthe data footprint size from [0.5 ∼ 10]× 106.Fig. 5 (a) shows that average task execution lifetime

increase as data footprint increases. Nonetheless, comparedto the other VM migration approaches, PRIMIO does notsignificantly increase the average task lifetime, because itconsiders a data-transfer penalty in the VM migration deci-sion process. Moreover, in the graph we can see that theaverage task lifetime reaches a steady state when the datafootprint size is greater than approximately 4 × 106. Largerthan this footprint, the PRIMIO method refrains from trans-ferring a VM closer to the user in order to reduce the totaltask execution time. On the other hand, other VM migration

approaches only consider user mobility when imitating VMmigration and, thus they increase the average task lifetime.

Fig. 5 (b) depicts that increased data footprint invreselyaffects the percentage of task completion within the deadline.However, the task completion percentage does not decreasein our proposed PRIMIO compared to the other approaches.As previously stated PRIMIO considers the data transfertime before initiating any VM migration decision. On theother hand, the percentage of task completion in the No VMmigration approach decreases significantly, as users movefrom one access point to another, the communication delayincreases, and thus it increases transfer time of final taskresult. Therefore, No VM migration approach significantlyreduces the percentage of tasks completed within deadline.

FIGURE 6. Impacts of data footprints: VM migration overhead.

Fig. 6 shows that increased data footprints also increaseVM migration overhead. The greedy and task-centric VMmigration policies that do not effectively consider migra-tion time, increase not only total number of VM migra-tions but also decrease number of tasks completed withindeadline. Nonetheless, our proposed method PRIMIO doesnot increase the VM migration overhead compared to otherpolices, mainly because PRIMIO decreases the total number

VOLUME 5, 2017 11897

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

of VM migrations as the data footprints size increases andalso increases the total number of tasks completed withindeadline. Thus, the proposed algorithm demonstrated itssuperiority over existing algorithms.

VII. CONCLUSION AND FUTURE WORKIn this paper, we have proposed a mobility- and resource-aware joint virtual-machine migration model for hetero-geneous mobile cloud computing systems to improve theperformance of mobile Smart health care applications in aSmart City environment. Here, we address research chal-lenges to reduce task-completion times as well as to reduceresource over-provisioning in mobile cloud computing thatexecutes both computationally and BigData-intensive health-care tasks. The proposed PRIMIOmodel initiates VMmigra-tion by considering user mobility and computational load of acloudlet. As PRIMIO exploits users’ mobility in achieving anoptimal solution, it effectively brings cloud resources closerto the user. At the same time, PRIMIO considers the loadof the cloudlet to which the system wants to migrate theVM, thereby reducing total number of migrations acrossthe entire task-execution time. Furthermore, we consideredrates of resource over-provisioning during VM migration,allowing the overall system to utilize computing resourcesoptimally. Left for our future work is how to further optimizethe task-computation time and data-access latency throughconsidering the presence of fog clouds and crowd sourcing.In this aspect, we will explore the emerging edge computingtechnology to further optimize the task execution time whileconsidering mobility and context-awareness.

REFERENCES[1] J. H. Abawajy and M. M. Hassan, ‘‘Federated Internet of Things and cloud

computing pervasive patient health monitoring system,’’ IEEE Commun.Mag., vol. 55, no. 1, pp. 48–53, Jan. 2017.

[2] E. Ahmed et al., ‘‘Enabling mobile and wireless technologies for smartcities,’’ IEEE Commun. Mag., vol. 55, no. 1, pp. 74–75, May 2017.

[3] M. Basiri, A. Z. Azim, and M. Farrokhi, ‘‘Smart city solution for sustain-able urban development,’’Eur. J. Sustain. Develop., vol. 6, no. 1, pp. 71–84,2017.

[4] N. Bobroff, A. Kochut, and K. Beaty, ‘‘Dynamic placement of virtualmachines for managing SLA violations,’’ in Proc. 10th IFIP/IEEE Int.Symp. Integr. Netw. Manage. (IM), May 2007, pp. 119–128.

[5] R. Buyya, C. S. Yeo, and S. Venugopal, ‘‘Market-oriented cloud com-puting: Vision, hype, and reality for delivering it services as comput-ing utilities,’’ in Proc. 10th IEEE Int. Conf. High Perform. Comput.Commun. (HPCC), Sep. 2008, pp. 5–13.

[6] B.-G. Chun, S. Ihm, P. Maniatis, M. Naik, and A. Patti, ‘‘Clonecloud:Elastic execution between mobile device and cloud,’’ in Proc. 6th Conf.Comput. Syst., 2011, pp. 301–314.

[7] R. Cimler, J. Matyska, and V. Sobeslav, ‘‘Cloud based solution for mobilehealthcare application,’’ in Proc. 18th Int. Database Eng. Appl. Symp.,2014, pp. 298–301.

[8] E. Cuervo et al., ‘‘Maui: Making smartphones last longer with codeoffload,’’ in Proc. 8th Int. Conf. Mobile Syst., Appl., Services, 2010,pp. 49–62.

[9] A. V. Dastjerdi and R. Buyya, ‘‘Fog computing: Helping the Internetof Things realize its potential,’’ Computer, vol. 49, no. 8, pp. 112–116,Aug. 2016.

[10] H. T. Dinh, C. Lee, D. Niyato, and P. Wang, ‘‘A survey of mobilecloud computing: Architecture, applications, and approaches,’’ Wire-less Commun. Mobile Comput., vol. 13, no. 18, pp. 1587–1611,Dec. 2013.

[11] C. Doukas, T. Pliakas, and I. Maglogiannis, ‘‘Mobile healthcare informa-tion management utilizing cloud computing and Android OS,’’ in Proc.Annu. Int. Conf. IEEE Eng. Med. Biol., Aug. 2010, pp. 1037–1040.

[12] E.-M. Fong and W.-Y. Chung, ‘‘Mobile cloud-computing-based health-care service by noncontact ecg monitoring,’’ Sensors, vol. 13, no. 12,pp. 16451–16473, 2013.

[13] L. Gkatzikis and I. Koutsopoulos, ‘‘Migrate or not? Exploiting dynamictask migration in mobile cloud computing systems,’’ IEEE WirelessCommun., vol. 20, no. 3, pp. 24–32, Jun. 2013.

[14] L. Gkatzikis and I. Koutsopoulos, ‘‘Mobiles on cloud nine: Efficient taskmigration policies for cloud computing systems,’’ in Proc. IEEE 3rd Int.Conf. Cloud Netw. (CloudNet), Oct. 2014, pp. 204–210.

[15] M. Guntsch, M. Middendorf, and H. Schmeck, ‘‘An ant colony optimiza-tion approach to dynamic TSP,’’ in Proc. 3rd Annu. Conf. Genetic Evol.Comput., 2001, pp. 860–867.

[16] M. M. Hassan, ‘‘Cost-effective resource provisioning for multimediacloud-based e-health systems,’’ Multimedia Tools Appl., vol. 74, no. 14,pp. 5225–5241, 2015.

[17] M. M. Hassan, H. S. Albakr, and H. Al-Dossari, ‘‘A cloud-assisted Internetof Things framework for pervasive healthcare in smart city environment,’’in Proc. 1st Int. Workshop Emerg. Multimedia Appl. Services Smart Cities,2014, pp. 9–13.

[18] M. M. Hassan, K. Lin, X. Yue, and J. Wan, ‘‘A multimedia healthcaredata sharing approach through cloud-based body area network,’’ FutureGenerat. Comput. Syst., vol. 66, pp. 48–58, Jan. 2017.

[19] M. S. Hossain, M. Moniruzzaman, G. Muhammad, A. Ghoneim, andA. Alamri, ‘‘Big data-driven service composition using parallel clusteredparticle swarm optimization in mobile environment,’’ IEEE Trans. ServiceComput., vol. 9, no. 5, pp. 806–817, May 2016.

[20] M. S. Hossain, S. A. Hossain, A. Alamri, and M. A. Hossain, ‘‘Ant-basedservice selection framework for a smart home monitoring environment,’’Multimedia Tools Appl., vol. 67, no. 2, pp. 433–453, 2013.

[21] M. S. Hossain, ‘‘Patient state recognition system for healthcare usingspeech and facial expressions,’’ J. Med. Syst., vol. 40, no. 12, p. 272, 2016.

[22] M. S. Hossain and G. Muhammad, ‘‘Cloud-assisted industrial Internetof Things (IIoT)—Enabled framework for health monitoring,’’ Comput.Netw., vol. 101, pp. 192–202, Jun. 2016.

[23] M. S. Hossain and G. Muhammad, ‘‘Healthcare big data voice pathologyassessment framework,’’ IEEE Access, vol. 4, pp. 7806–7815, 2016.

[24] M. Islam, A. Razzaque, and J. Islam, ‘‘A genetic algorithm for virtualmachine migration in heterogeneous mobile cloud computing,’’ in Proc.Int. Conf. Netw. Syst. Security (NSysS), Jan. 2016, pp. 1–6.

[25] A. R. Khan, M. Othman, S. A. Madani, and S. U. Khan, ‘‘A survey ofmobile cloud computing application models,’’ IEEE Commun. SurveysTuts., vol. 16, no. 1, pp. 393–413, Feb. 2014.

[26] S. Kosta, A. Aucinas, P. Hui, R. Mortier, and X. Zhang, ‘‘ThinkAir:Dynamic resource allocation and parallel execution in the cloud for mobilecode offloading,’’ in Proc. IEEE INFOCOM, Mar. 2012, pp. 945–953.

[27] P. Kulkarni and T. Farnham, ‘‘Smart city wireless connectivity consider-ations and cost analysis: Lessons learnt from smart water case studies,’’IEEE Access, vol. 4, pp. 660–672, 2016.

[28] R. Kumari et al., ‘‘Application offloading using data aggregation in mobilecloud computing environment,’’ in Leadership, Innovation Entrepreneur-ship as Driving Forces Global Economy. Switzerland: Springer, 2017,pp. 17–29.

[29] P. G. J. Leelipushpam and J. Sharmila, ‘‘Live VM migration techniquesin cloud environment : A survey,’’ in Proc. IEEE Conf. Inf. Commun.Technol. (ICT), Apr. 2013, pp. 408–413.

[30] J. Li, K. Bu, X. Liu, and B. Xiao, ‘‘Enda: Embracing network inconsistencyfor dynamic application offloading in mobile cloud computing,’’ in Proc.2nd ACM SIGCOMMWorkshop Mobile Cloud Comput., 2013, pp. 39–44.

[31] J. Liu, Y. Li, D. Jin, L. Su, and L. Zeng, ‘‘Traffic aware cross-site virtualmachinemigration in futuremobile cloud computing,’’Mobile Netw. Appl.,vol. 20, no. 1, pp. 62–71, Feb. 2015.

[32] J. Montgomery, M. Randall, and T. Hendtlass, ‘‘Structural advantages forant colony optimisation inherent in permutation scheduling problems,’’ inProc. 18th Int. Conf. Innov. Appl. Artif. Intell., 2005, pp. 218–228.

[33] Z. L. Phyo and T. Thein, ‘‘Correlation based vms placement resourceprovision,’’ Int. J. Comput. Sci. Inf. Technol., vol. 5, no. 1, p. 95, 2013.

[34] M. Rahimi, J. Ren, C. Liu, A. Vasilakos, and N. Venkatasubramanian,‘‘Mobile cloud computing: A survey, state of art and future directions,’’Mobile Netw. Appl., vol. 19, no. 2, pp. 133–143, 2014.

11898 VOLUME 5, 2017

Md. M. Islam et al.: Mobile Cloud-Based Big Healthcare Data Processing in Smart Cities

[35] C. C. Sasan Adibi and N.Wickramasinghe, ‘‘CCmH: The cloud computingparadigm for mobile health (mHealth),’’ Int. J. Soft Comput. Softw. Eng.,vol. 3, no. 3, pp. 403–410, 2013.

[36] M. Satyanarayanan, P. Bahl, R. Caceres, and N. Davies, ‘‘The case for VM-based cloudlets in mobile computing,’’ IEEEPervas. Comput., vol. 8, no. 4,pp. 14–23, Oct. 2009.

[37] M. Sneps-Sneppe and D. Namiot. (2016). ‘‘On mobile cloud for smart cityapplications.’’ [Online]. Available: https://arxiv.org/abs/1605.02886

[38] T. Taleb and A. Ksentini, ‘‘An analytical model for follow me cloud,’’in Proc. IEEE Global Commun. Conf. (GLOBECOM), Dec. 2013,pp. 1291–1296.

[39] H. N. Van, F. Tran, and J.-M. Menaud, ‘‘Sla-aware virtual resource man-agement for cloud infrastructures,’’ in Proc. 9th IEEE Int. Conf. Comput.Inf. Technol. (CIT), vol. 1, Oct. 2009, pp. 357–362.

[40] U. Varshney, Pervasive Computing and Healthcare. Boston, MA, USA:Springer, 2009, pp. 39–62.

[41] L. Wang, F. Zhang, A. V. Vasilakos, C. Hou, and Z. Liu, ‘‘Joint virtualmachine assignment and traffic engineering for green data center net-works,’’ SIGMETRICS Perform. Eval. Rev., vol. 41, no. 3, pp. 107–112,Jan. 2014.

[42] S. Wang, R. Urgaonkar, T. He, M. Zafer, K. Chan, and K. Leung,‘‘Mobility-induced service migration in mobile micro-clouds,’’ in Proc.IEEE Military Commun. Conf. (MILCOM), Oct. 2014, pp. 835–840.

[43] I. Yaqoob, I. A. T. Hashem, Y. Mehmood, A. Gani, S. Mokhtar, andS. Guizani, ‘‘Enabling communication technologies for smart cities,’’ IEEECommun. Mag., vol. 55, no. 1, pp. 112–120, Jan. 2017.

[44] Q. Zhang, L. Cheng, and R. Boutaba, ‘‘Cloud computing: State-of-the-artand research challenges,’’ J. Internet Services Appl., vol. 1, no. 1, pp. 7–18,2010.

MD. MOFIJUL ISLAM received the B.S. andM.S. degrees in computer science and engineeringfrom the Department of Computer Science andEngineering, University of Dhaka, Banglabesh. Hewas a Software Engineer with Tiger It Ltd. He iscurrently a Lecturer (part-time) with the Depart-ment of Computer Science and Engineering, Uni-versity of Dhaka. He is involved in programmingtraining, mobile apps, and different software con-test team-building activities. His research inter-

ests include cloud computing and mobile cloud computing, human com-puter interaction, human-centered computing, data science and informationretrieval, big data and visualization, and user experience design.

MD. ABDUR RAZZAQUE (SM’12) receivedthe B.S. and M.S. degrees from the Universityof Dhaka, Bangladesh, and the Ph.D. degree inwireless networking from Kyung Hee University,SouthKorea, in 2009. Hewas a Research Professorwith the College of Electronics and Information,Kyung Hee University, from 2010 to 2011. He isa Professor with the Department of Computer Sci-ence and Engineering, University of Dhaka, wherehe is also the Group Leader of the Green Network-

ing Research Group. He has published a number of research papers in theIEEE/ACM/Springer conferences, journals, and books. His research interestsare in the areas of modeling, analysis, and optimization of wireless net-working protocols and architectures, wireless sensor networks, body sensornetworks, cooperative communications, sensor data clouds, and cognitiveradio networks. He is a TPC Member of the IEEE HPCC, ICOIN, ADM,ICUFN, and NSyS. He is a senior member of the IEEE Communica-tions Society, the IEEE Computer Society, the Internet Society, the PacificTelecommunications Council, and KIPS.

MOHAMMAD MEHEDI HASSAN (M’12)received the Ph.D. degree in computer engineeringfromKyung Hee University, South Korea, in 2011.He is an Associate Professor with the InformationSystems Department, College of Computer andInformation Sciences (CCIS), King Saud Univer-sity (KSU), Riyadh, Saudi Arabia. He has authoredmore than 100 research papers in internationaljournals and conferences. His research areas ofinterest are cloud federation, multimedia cloud,

sensor-cloud, Internet of Things, big data, mobile cloud, cloud security,IPTV, sensor network, 5G network, social network, publish/subscribe sys-tem, and recommender system. He received the Best Paper Award from theCloudComp Conference in China in 2014. He also received the Excellencein Research Award from CCIS, KSU, in 2015 and 2016, respectively.He has served as the Chair and a Technical Program Committee Member innumerous international conferences/workshops, such as the IEEEHPCC, theACMBodyNets, the IEEE ICME, the IEEE ScalCom, the ACMMultimedia,ICA3PP, the IEEE ICC, TPMC, and IDCS. He was the Guest Editor ofseveral international ISI-indexed journals.

WALAA NAGY ISMAIL is currently pursuing the Ph.D. degree with theInformation Systems Department, College of Computer and InformationSciences, King SaudUniversity, Riyadh, Saudi Arabia. Her research interestsinclude pervasive health care, anomaly detection, cloud computing, mobilehealth care, and data mining.

BIAO SONG (M’12) received the Ph.D. degreein computer engineering from Kyung Hee Uni-versity, South Korea, in 2012. He is currently anAssistant Professor with the College of Computerand Information Science, King Saud University,Saudi Arabia. His current research interests arecloud computing, remote display technologies,and dynamic VM resource allocation.

VOLUME 5, 2017 11899