an efficient technique for virtual machine clustering and

Research ArticleAn Efficient Technique for Virtual Machine Clustering andCommunications Using Task-Based Scheduling inCloud Computing

C Saravanakumar 1 M Geetha 2 S Manoj Kumar 3 S Manikandan 4 C Arun 5

and K Srivatsan 6

1Department of Information Technology St Josephrsquos Institute of Technology OMR Chennai Tamil Nadu 600119 India2Department of Computer Science and Engineering School of Computing College of Engineering and TechnologyFaculty of Engineering and Technology SRM Institute of Science and Technology Vadapalani Chennai Tamil Nadu India3Department of Information Technology Karpagam College of Engineering Coimbatore Tamil Nadu 641 032 India4Department of Information Technology K Ramakrishnan College of Engineering Trichy Tamilnadu 621 112 India5Department of ECE RMK College of Engineering and Technology Chennai Tamil Nadu 601 206 India6School of Electronics Engineering VIT Chennai Tamil Nadu 600127 India

Correspondence should be addressed to C Saravanakumar saravanakumarcstjosephstechnologyacin

Received 20 February 2021 Revised 6 June 2021 Accepted 1 July 2021 Published 19 July 2021

Academic Editor Pengwei Wang

Copyright copy 2021 C Saravanakumar et al )is 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

Cloud computing models use virtual machine (VM) clusters for protecting resources from failure with backup capability Clouduser tasks are scheduled by selecting suitable resources for executing the task in the VM cluster Existing VM clustering processessuffer from issues like preconfiguration downtime complex backup process and disaster management VM infrastructureprovides the high availability resources with dynamic and on-demand configuration )e proposed methodology supports VMclustering process to place and allocate VM based on the requesting task size with bandwidth level to enhance the efficiency andavailability )e proposed clustering process is classified as preclustering and postclustering based on the migration Task andbandwidth classification process classifies tasks with adequate bandwidth for execution in a VM cluster )e mapping ofbandwidth to VM is done based on the availability of the VM in the cluster )e VM clustering process uses different performanceparameters like lifetime of VM utilization of VM bucket size and task execution time )e main objective of the proposed VMclustering is that it maps the task with suitable VM with bandwidth for achieving high availability and reliability It reduces taskexecution and allocated time when compared to existing algorithms

1 Introduction

Cloud computing is a service-oriented architecture tech-nique that uses virtualized resources to perform computa-tional tasks )e cloud has a set of resources that are offeredas a means of service Cloud services are classified asSoftware as a Service (SaaS) Infrastructure as a Service(IaaS) Platform as a Service (PaaS) etc )e services aredeployed in different models to meet customer require-ments )ey are classified as private cloud public cloudcommunity cloud and hybrid cloud Private cloud resources

are shared within the organization in which the services areshared outside the organization as a public cloud Com-munity cloud is a type of cloud in which services are sharedbetween the service providers of the same category Acombination of two or more deployment models called ahybrid cloud provides the customer a service Cloud servicesare modeled by mapping the virtualization layer to theappropriate VM VMs are selected from the VM list and thenmapped to the respective request generated by the user of thecloud service)e cloud consists of a heterogeneous host in adata center that maintains a mobile resource-based access

HindawiScientific ProgrammingVolume 2021 Article ID 5586521 15 pageshttpsdoiorg10115520215586521

environment VM is a type of access that leads to a per-formance problem seen in the areas of the battery life andenergy consumption)e entire performance factor in greencomputing is used for overcoming this problem [1] Mobilecloud computing (MCC) is a mobile resource sharing servicethat allows mobile devices to have access to the appropriatecloud service It has faced challenges in terms of scalabilitysuch as computational storage services and other differentservices [2]

Mobile computing over cloud has the ability to targetparameters such as traffic quality and customer demandTraditional static cloud and dynamic cloud are compared toanalyze the workload )e static cloud allows users to accessinfrastructure services with a specific configuration whilethe dynamic cloud provides an agile response method toupdate the resource configuration Dynamic cloud has avariety of wireless nodes with device-to-device connectivityfor the achievement of a better utilization of the channel andtraffic [3] VM contention IaaS has created problems in thearea of performance )is problem is overcome through theimplementation of the data center as various ranges such assingle server with virtualization single mega data center andmultiple geo-distributed data centers [4] A researcher en-gaged in cloud suffers from issues that include energyconsumption in data centers )e data center is a key ele-ment in the cloud that handles all kinds of resources neededin the computing environment In the cloud there are twotypes of approaches that are related to hardware and soft-ware)ese approaches require reduced power consumptionin cloud resources without any situation of service un-availability [5] Cloud resources are multiplexed over VMservers designed to host cloud services in large data centers

VMmigration is the process of migrating from one locationto another leading to the performance problem arising as a resultof inference and the cost of the operation iAware imposes themultiresource supply demand model for minimizing the in-ference during the migration [6 7] Intelligent transportationsystems (ITS) are used in the vehicle cloud computing (VCC)architecture which consists of two-layered components such asthe central cloud server and the remote system control (RSC)RSC is a remote administration manager for monitoring andmanaging distributed system elements such as network com-munication lines RSC uses two local server and road side unit(RSU) components If the vehicle travels from one position toanother the VM of one RSU is moved to another RSU Itrequires continuous service response over the automated vehiclecontrol using VCC [5] )ere are two main disputes seen in theintelligent transportation systems (ITS) namely efficiency intraffic and energy quality and productivity )e data collectedfrom various sensors are overcome by using the parallelizedfusion technique )is technique follows the DempsterndashShafertheory with four components namely sensor input boot-strapping hierarchical fusion and state output [6 8] Cyber-ITSis a system that has data division scheduling and efficientsupport through the use of a genericmethodological framework)ere are two types of functions carried out by the systemnamely data-centric and operation-centric transformations)is model uses high-performance computer design with re-gion-based decoupling capability [9 10]

In this method a digital map of the global positioningsystem data is processed in a parallel manner using quad-tree-based domain decomposition technique )ese data aredivided into different subdomains with quad structure [11]Multi-CPU VM scheduling and virtual CPUs (VCPUs)scheduling have been carried out due to the availability ofvarious virtualization techniques in the cloud computingdomain [10] )e existing problems have been analyzedbased on the performance parameters which improve theefficiency in the cloud service deployment )e main ob-jective of the proposed technique is to identify the suitableVM based on the bandwidth and requesting task of allo-cating any issues in the performance of cloud to the task

)e VMs are configured in an isolated fashion whichsuffers from repetitive booting of the respective volumeswith a limited period of delivery )is problem is solved by acluster management approach based on a docker containerwith a diverse configuration [12] )e traditional placementof the docker container and the VM is carried out separatelyso that it is implemented using the container VM-PMmodel[13])e Internet of)ings plays a key role in the processingof real-time data from hardware devices that generate largequantities of data)ese data are stored in a large data centerwith big data analysis methods If the data are huge a hugenumber of servers would be used to store the received data Itfaces an expensive problem that is solved by using a cabinet-based tool called ProCon [14] Virtualization technologyoffers the benefits of the physical server operating on severalcomputers with different resources Virtual storage elimi-nates read and write delays with high IO efficiency Virtualdisks are connected to VMs for processing and storing theuser data via synchronization [15] Amazon cloud providerprovides various kinds of services to the end user in reliableand secure computing capacity AWS offers the ElasticCompute Cloud (EC2) with different versions of VM andresources )e proposed method used EC2 instances as areference for further analysis

)e rest of the paper is organized as follows Section 2depicts the architecture of the proposed system Section 3describes the phases of the proposed system Section 4presents the analysis of various metrics Section 5 describesthe proposed VM clustering process Section 6 provides theexperimental evaluation Section 7 signifies a comparison ofthe various clustering algorithms Section 8 presents theconclusions and indications for future work

2 Model of the Proposed System

Existing VM management techniques configure a VM tocloud workloads based on the energy parameters but theysuffer from a resource wastage problem in the data center)e proposed model groups the tasks based on the VM typeswith bandwidth parameters using classification methods)e VM types are customized based on the resourceavailability in the cloud)e objective of the proposed modelis to map a task to the correct VM by considering variousprocesses such as resources mapping and classification [16])e cloud requesting tasks are classified based on the metricssuch as total number of queues request count API response

2 Scientific Programming

count dispatch rate of the queue maximum size of the taskactual and scheduled execution time delay and task re-tention time )ese metrics are collected using CloudWatchmonitoring service in AWS cloud )ese metrics areexported and used for analyzing the proposed model

Figure 1 shows the architecture of the proposed system)e customer makes a request for the resource from thecloud based on their current requirement )e request isconsidered as a task executed in a VM )ere are varioustasks generated by the customer identified based on variousperspectives )e identified task is classified based on thereliability parameter in order to achieve a high performanceBandwidths are selected for the classified task so that suitableVM is allocated to the requesting task )ere are variousbandwidths available in the requested task which areidentified and classified to enable mapping the suitable VMfor service delivery )e bandwidth-VM mapping sectionselects the VM from the VM clusters which in turn areselected by the VM selector )e VMs are clusters based onthe task and bandwidth meant for providing the serviceswithout any interruption or delay )ey are clustered in therespective VMs based on the type of task requested and thehypervisor eventually provides the link between the VMsand the physical system

Figure 2 is the sequence diagram of the proposed systemIt denotes how the incidents are actually related to eachother)e activity diagram shows how the process starts andterminates the various state changes and activities that takeplace between these state changes )e first phase is the loginmodule where the user is authenticated to access the system)e username and password are provided )ese furtherprovide access to the cloud home page which consists of allthe main functionalities )is module allows only the au-thorized user to log in to the system It provides authen-tication and access only to the authorized user and allows theuser to select any of the options that include creation of avirtual machine viewing the existing machines makingtask-based separation and viewing the usage report)e firstoption in the module is helpful for the user in creation of themodule by just entering the values for the new VM Insteadof typing commands in the terminal to create a virtualmachine this module helps us create a VM by simply en-tering the values

)e second option in the module is for viewing theexisting VMs thereby enabling the user to view the type ofVM created and used Each VM has different specifica-tions )is module can view the existing VMs with theirspecification the operating system of the VM )e thirdoption in the module is task-based separation )is phasehelps separation of the task from the cloud user with thehelp of this option It separates the task in terms of CPUmemory and IO )e fourth option in the module is VMclustering Groups of VMs with similar features areclustered with the help of this option )e final option isthe report where the user can view a detailed usage of theVMs In this module the user has the ability to create newVMs for use based on requirements )e major use of thismodule is that unlike in the normal VM creation wherethe user needs to go to the terminal and type commands

the user just needs to type values for the VM creation Aname for the VM and the various parameters such as thenumber of CPU vCPU the disk name NIC and SSH areentered where all these parameters are required for thecreation of the VM template

)e template which is created is instantiated by pro-viding the disk name and the VM name thereby creating theVM )is module is used for helping the user in makingseparation on the VM based on the type of task in eachrequest )ere are three major classifications on the task thatinclude memory-based CPU-based and IO-based classifi-cation )e CPU-based classification is meant for the userwho requires high processing speed during the processing ofthe input file In this the input file is only processed but notstored )e memory-based classification is for the user tohave better memory space In this the input file is onlystored in the memory and not processed )e IO-basedclassification is used for the user to have a responsive VM Inthis option the output is only generated for the input file andnot stored in the memory

)emixed option is used for the user in the classificationof a task with more than one type Selection of this type helpsthe user in the selection of either of the two types of clas-sification method for each task )is module is used forallowing the user to view the existing VMs that are created inthe system)is view option is in a tabular view where all theexisting VMs are listed along with the specifications of eachVM)emodule is highly useful for getting knowledge of allthe VMs that have already been created in the system All theactive VMs can be viewed with the help of this module )ebasic uses of this module provide a view of the existing VMsand differentiate active VMs )e module displays the in-formation of the existing virtual machines such as their userID group and name

3 Mathematical Analysis of VM Clustering

Normally the VM is created from the physical machine(PM) through the use of a virtual machine manager (VMM))ere are various VMs maintained and controlled by theDomain_0 which acts as a supervisor performing allmapping operations with maximum availability amongother domains Five phases of the VM states exist in thecloud namely VM Creation State VM Running State VMSuspend State VM Resume State and VM Destroy State)e distribution of VM is represented in the normal dis-tribution with [minusinfin infin] )is limit is unbounded and sothis is represented with the bounded manner by changingthe limit to [minusMM] )e VM distributions are expressed in(1) VMs are considered as VM sets called VM VM1 VM2 VMn) of size N )e requesting tasks are allocated intospecific VM based on the task categories which are rep-resented as Task T1 T2 Tn )e allocation of the VMand tasks is specified as Allocation T1⟶VM1T2⟶VM2 Tn⟶VMn with the relevant size SNormal distribution is done based on the clustering with acluster size of 2 and then it increases until stopping con-ditions are reached It is represented as N1 N2 Nn withbounded distribution in equations (2)ndash(4)

Scientific Programming 3

Cloud serviceuser

Taskidentification

Taskclassification

Bandwidthselection

Bandwidthclassification

Bandwidthand virtual machine mapping

Virtual clusters

VC1 VC2 VCn

Virtual machines

Storage Memory

Virtual machine

CPU

Figure 1 Architecture of the proposed system

User System Cloud

Login ()

Create VM

VM created

View existing VMs

Requesting for VM createRequest granted

Loading VM list

Requesting separation

Request to create cluster

Separate VM based on task

Existing VM list

VM separated

VM clustered

VM clustering

Authentication success

Figure 2 Phases of the proposed VM clustering process


Distribution of VM 1

σActiveVM

21113937VM

1113969⎛⎜⎜⎜⎝ ⎞⎟⎟⎟⎠1113946α

minusαXVM minus μActiveVM( 1113857

r eminus XminusμActive VM( )

2σ2ActiveVM middot S⎛⎝ ⎞⎠ (1)

1

σActiveVM

21113937VM

1113969 1113946M


r1113946α

MXVM + μActiveVM( 1113857

s1113876 1113877 middot S (2)

1

σActiveVM

21113937VM

1113969 1113946M

minusαN1 + N21113858 1113859 middot S (3)

1

σActive VM21113937

1113968VM

1113946M

minusαN1 + N2 + middot middot middot + Nn1113858 1113859 middot S (4)

)e number of clusters is either odd or even So it ishandled based on the conditions of the VM distribution)ere are two cases to follow by performing VM clusteringprocess

Case 1 If the number of cluster selections is odd ier 2n+ 1 the VM distribution formulated is represented as

μ2n+1 2(2n+12)σ2n+1

ActiveVM1113872 11138731113937VM

1113968 1113946M

minusMN1 + N2 + middot middot middot Nn1113858 1113859 middot S

(5)

Case 2 If the number of cluster selections is even ie r 2nthe VM distribution formulated is represented as

μ2n 22n2σ2n

ActiveVM1113872 11138731113937VM

1113968 1113946M

minusMN1 + N2 + middot middot middot + Nn1113858 1113859 middot S

(6)

A VM clustering process carried out is based on thecategorization of VM as active and inactive VM (idle VM)

)is is done by considering the processor states which arerepresented in equations (7) and (8) respectively

VMActive State VMActive isin VMList state active

ϕ otherwise1113896

(7)

VMInactive State VMInactive isin VMList state idleϕ otherwise

1113896

(8)

PM and VM mapping is done by considering the uti-lization of the CPU and vCPU in both the guest and hostenvironments by finding mean value )is is shown in

PMVMmapping ratio mean (maxPMutilization

+ maxVMutilization)(9)

)e lifetime of the VM is assessed based on the timetaken from creation to process the completion stage whichis always greater than one is described in

VM life time time (VMcreation + VMexecution + VMdestroy) timegt 0

ϕ otherwise1113896 (10)

Data are stored in the cloud bucket by performing readand write operations with scalable and flexible properties byaccessing in a ubiquitous manner )e average bucket count

is calculated using the bucket count of PM as well as VM isshown in

average bucket count countPMbucket count minus VMbucket count

total bucket count1113890 11138911113896 1113897 (11)

VM clusters depend on various factors which includeVM states utilization of CPU in both VM and PM bucket

size lifetime of the VM and number of cores )is process iscarried out based on all clusters which are shown in


VMclusteringi ⋃n

i1Count Active VMi + CPUutilizationi + VMbucketi + VM life timei + core counti1113858 1113859 (12)

4 Analysis of Various Performance Metrics

41 VM Parameter-Level Analysis )ere are different fea-tures considered while clustering the VM for a goodmaintenance of reliability in the cloud )e capabilitiesavailable in the migration process need to keep the copy ofthe VM at the original source end Two types of copyingprocess happen in the VM namely precopying and post-copying process )e cluster process uses these techniquesfor achieving better availability )e CPU is halted duringthe migration process in the source machine as well as thedestination machine Delta-based compression of VM hasmore number of dependent VMs with respective referencesfor future VM consolidation at the target machine Data-level compression is used for getting the reduction in thecontents related to the VM at the source machine )eworkload is classified as synthesis-based and idle-basedworkload Synthesis-based workload is a preallocated loadassigned before the VM migration Idle-based workload isassigned to the task based on the demand in nature VM sizedepends upon various factors namely the number of vCPUmemory size in GB the bandwidth of the memory in GBsthe frequency of the CPU in GHz single and all core fre-quencies in GHz performance of remote memory accesstemporary storage in GB number of data disks and numberof Ethernet NICs Initially the target machine memory isconsidered as dirty pages Later it is replaced with the re-spective contents after migration Resource availability of thetarget machine is also addressed during the migration of theVM Table 1 shows the VM features in multiple perspectives

42 User Task Classification-Level Analysis Cloud user tasksare classified based on the factors that include the name ofthe user base (UB) regions and the number of requests persingle user requested task size in bytes duration in terms ofpeak hours in GMT and the average number of peak usersoffline and online Table 1 shows the classification of taskswith user parameters of different sizes )ese tasks aremapped with the VM parameters for achieving high reli-ability Cloud regions maintain the resources for users withthe corresponding user base )e requesting size and relatedtask size are analyzed in order to find average number ofusers in both peak and nonpeak hours of the particularregion )e comparison of task classification is shown inTable 2 Heterogeneous VMs are available for handling theuser-requesting task with different categories Virtual ma-chine manager schedules the task to respective VM A largenumber of requesting tasks to be handled have arrived in theclouds that are performed by a large number of VMs It isproposed that the number of virtual machines be increasedon the basis of number of tasks requested

Cloud region collects the user requests from the userbase (UB) and identifies the request size )e tasks are al-located to the respective resources in the cloud region based

on the peak and nonpeak hours through autoscaling tech-nique It handles the maximum number of VM based on thedemand Figure 3 shows the total number of tasks on theVMs over the cloud region

43 Physical Machine Analysis of Data Centers )e physicalmachine in the data centers has resources that includememory size in MB storage in MB number of processorsthe speed of the processor and VM policy such as timeshared and space shared All the resources are virtualizedand also shared by multiple VMs for task execution Table 3shows the physical resources at the data centers with variousperformance parameters)e process speed is represented asinstruction per second (IPS) )ese physical resources needVM policy for executing tasks which are either time sharedand space shared )e proposed analysis is time sharedbecause of CPU intensive operation carried out for highavailability

44 Virtual Machine Analysis of Data Centers Virtual ma-chines are created with different parameters such as thename of the data centers regions of the data center thearchitecture of the VM the platform of the VM type ofVMM the cost of the resources and physical hardwareunits Data centers perform various operations which in-clude migration process in either the same data center withmigration or different data centers with migration Linuxwith X86 architecture is used for deciding the requirednumber of physical machines at the data centers in differentregions Table 4 describes the VM attributes for task allo-cation and execution Cloud customer needs higher band-width because of the lack in the current bandwidth levelCloud providers provide sufficient bandwidth in order toretain the customers Table 5 provides the bandwidth level ofvarious domain users for effective utilization

45 Delay and Bandwidth Matrix Analysis Cloud servicesare deployed in various data centers as regions Delays seenbetween the regions are compared for efficiency )e mainobjective of this delay analysis is to identify the fast responsewith minimum response rate )e bandwidth matrix pro-vides an efficient route between the requests in the shortestpath with maximum availability Delay in the networkduring the transfer of jobs across various regions is shown inTable 6)e delay between the same region is also minimumwhereas the delay of the different regions is maximum withrespect to the distance between the regions )e same regiontransfer rate is fast when compared to different regions so itdepends on the bandwidth level and delay which is shown inTable 7 Figure 4 indicates the response rate of various cloudregions in ms (Milliseconds) Response time of various userbases is measured in three levels namely average mini-mum and maximum )e CloudAnalyst model is used for


Table 1 VM parameters

VM parameters Description

Migration algorithm (capability)

0 original precopy1 CPU-based halting

2 delta-based compression3 data-based compression

4 postcopy

Workload type 0 synthetic-based workload1 idle-based workload

VM size Size of VM at data centerPage dirty rate Total number of modified pages in the VMWorking set size )e size of dirty pagesPage transfer rate VM migration rate from one cluster to another clusterCPU utilization of VM Physical CPU utilization based on VMNetwork utilization of VM Bandwidth utilization of the VMCPU utilization on host Destination of VM clustersrsquo CPU utilizationMemory utilization on host Destination of VM clustersrsquo memory utilizationIO utilization on host Destination of VM clustersrsquo IO utilization

Table 2 User task classification and comparison

Name of the userbase Regions Requests per user

perTask size per request

(bytes)Peak hours starts

(GMT)Peak hours end

(GMT)Avg peakusers

Avg off-peakusers

UB1 0 60 100 3 9 1000 100UB2 1 60 90 4 10 1000 100UB3 2 40 75 5 8 1000 100UB4 3 55 80 7 11 1000 100UB5 4 30 85 8 9 1000 100UB6 5 45 100 6 12 1000 100

1 2 3 4 5 6Cloud regions

Number of tasksNumber of cloud regions

35000

30000

25000

20000

15000

10000

5000

0

Num

ber o

f tas

ks

Figure 3 Comparison of cloud regions and user tasks

Table 3 Data center analysis for the physical machine

ID Memory (MB) Storage (MB) Available bandwidth Number of processors Processor speed (IPS) VM policy0 204800 100000000 1000000 4 10000 Time shared1 254800 200000000 1000000 4 10000 Time shared2 304900 300000000 1000000 4 10000 Time shared3 354500 600000000 1000000 4 10000 Time shared4 224800 800000000 1000000 4 10000 Time shared


the proposed analysis with delay and bandwidth allocationof the task execution )e custom bandwidth and delaymatrix are specified on internet characteristics option withvarious regions

46Analysis ofVMClustering VMclustering is the process ofgrouping various VMs using virtual networks for achievinghigh availability of cloud resources )is clustering process isdone in a sourcemachine as well as a targetmachine)ere aretwo important concepts considered for good accuracy namelypreclustering and postclustering A preclustering process oc-curs at the source end whereas postclustering occurs at thetarget machine end Preclustering interconnects the VM alongwith the state of the processor data and VM-related pa-rameter )ese VMs then migrate to the target machinePostclustering process collects the VMs in order and finds therelationships forming a cluster by interconnecting the receivedVMs)is process resumes all the states of the VM and relatedelements into original state Table 8 shows the various VMclusters with the required parameters )ere are two types ofVMs grouped namely active and inactive Utilization of theCPU plays a vital role during the clustering process VMlifetime is considered in allocating better performance duringthe clustering process

5 Proposed VM Clustering

)e clustered virtual machines (VMs) are created from aphysical machine (PM) )e mapping of VM and PM isperformed by the hypervisor)e objective of clustering is toexecute the requesting task which is categorized with dif-ferent parameters )e bandwidth is classified and thecorresponding VMs are mapped )e VMs are clustered in away of similar categories of VM grouped together )e al-location of VMs is extremely simple and efficient for theexecution of a task )ere are large numbers of clusteredVMs that function as dynamic behavior clustering )ecompletion time of the VM is analyzed and clusters of VMsare shown in Figures 5ndash7

Table 4 Data center analysis for virtual machines

Name Region Arch OS VMM Cost per VM $hr Memory cost $s Storage cost $s Data transfer cost $GB Physical HW unitsDC1 0 x86 Linux Xen 01 005 01 01 5DC2 1 x86 Linux Xen 02 005 03 025 1

Table 5 Comparison of bandwidth utilization

Domain of service providers Current bandwidth utilization () Expected bandwidth level ()Banking and insurance 61 19Telecommunications 73 10IT services 39 16Cloud services 62 4Education services 36 4Government services 41 10

Table 6 Comparison of the delay matrix

Regionsregion R0 R1 R2 R3 R4 R5R0 250 1000 1500 2500 2500 1000R1 1000 250 2500 5000 3500 2000R2 1500 2500 250 1500 1500 2000R3 2500 5000 1500 250 5000 5000R4 2500 3500 1500 5000 250 5000R5 1000 2000 2000 5000 5000 250

Table 7 Comparison of the bandwidth matrix


700

600

500

400

300

200

100

0

Resp

onse

tim

e in

ms

UB1 UB2 UB3 UB4 UB5 UB6User base

Avg (ms)Min (ms)Max (ms)

Figure 4 Comparison of response time of the cloud region


6 Experimental Evaluation

Experimental setup is done based on the CloudSim andCloudAnalyst model with a cloudlet for task specifica-tion VM management service allocation of resources(CPU storage bandwidth and memory) and VMprovisioning service )e data center (DC) host and VMare created with suitable cloud brokers in CloudSim and

analysis of parameters in CloudAnalyst DC is configuredwith x86 architecture Linux OS Xen Hypervisorhardware unit specific region data center name and costof data transfer and resources )e parameter values ofthe VM configuration are time-shared VM policy 25 GBof memory storage size of 1 TB 1Mbps of bandwidthnumber of tasks of 100 VM size of 10000 and speed of28 GHz

Number of virtual machines

200

180

160

140

120

100

80

60

40

20

0

Tim

e (se

c)

Resource IDVMIDTime

Start timeFinish time

Figure 5 Time comparison of VM cluster 2

Table 8 Parameters of VM clustering process

Clusters ActiveVMs

InactiveVMs

Maximum CPUutilization

Average CPUutilization

VM core countbucket

VM memorybucket

VM lifetime

0 558300 1673700 9177688 072887 8 32 30983331 424500 4254000 3787926 332535 4 32 0252 113310 1133700 0304368 022055 4 32 0166663 0 2591400 9857342 334005 2 4 7198334 228300 2298000 8258144 187629 2 4 041666

700

600

500

400

300

200

100

0

Tim

e (se

c)


Resource IDVMIDTime




1400

1200

1000

800

600

400

200

ndash200

0

Tim

e (se

c)


Resource IDVMIDTime



120

100

80

60

40

20

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(a)

20

15

10

5

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(b)

Figure 8 Continued


250

200

150

100

50

0

Reso

urce

s

10 20 30 40 50 60Time (minutes)

(c)

Figure 8 Analysis of VM clusters with task execution (a) VM cluster 1 (b) VM cluster 2 (c) VM cluster 3

30252015100500

Virt

ual m

achi

nes

Pow

er (k

W)

120

100

80

60

40

20

00 10 20 30 40 50

Time (minutes)Customer 1

Customer 2

Customer 3

(a)

200

150

100

50

0

Clou

dlet

s

Pow

er (k

W)

3035

25201510

50

0 10 20 30 40 50Time (minutes)Customer 1

Customer 2

Customer 3

(b)

100

75

50

25

0

ndash25

Cos

ts (m

u)

Pow

er (k

W) 15

20

10

5

00 10 20 30 40 50


Customer 2

Customer 3

(c)

Figure 9 Comparison of VM allocation execution and cost with power consumption (a) VM allocation vs power consumption (b) Taskexecution vs power consumption (c) Data center cost vs power consumption


Cloud report is a model that is used for the simulation ofthe cloud services with various supporting packages to enablethe analysis of the cloud-computing environment [17] )ismode has meta-data for allocating resources to the data centerssuch as host policy VM with configuration and cloud brokerpolices Analysis of the resource allocation and related con-sumption of energy is made with different parameters at thedata center )e VMs are allocated to the data center based onthe task category and analyzed with various configurations[18 19] )e user requires knowledge of how the resources areutilized with their task at the data center)e execution time ofthe task from the requesting time to the completion time is alsoanalyzed to provide satisfaction to the customer )ree VMclusters are considered for the proposed evaluation of thealgorithm )ree clusters are modeled in the proposed systemfor executing tasks at the resources of the cloud Resourceutilization is analyzed based on the CPU RAM and bandwidthin the VM clusters as shown in Figure 8

)e energy consumption of VM before and after allo-cation of tasks over the data center is represented in Figure 9)e power consumption is high in cluster 1 and cluster 3 asthe number of task assignments to the VM is large whencompared to cluster 2

A comparison of task allocation to the VM is shown inFigure 10 )e VMs are created and scheduled by the VMmonitor which is responsible for the task assignment)e taskswhich are considered as cloudlets are configured with thecustomer requirement and achieved efficiency of a high order

7 Results and Discussion

)e ATOM-based VM clustering method helps in theevaluation of the performance measurement with parame-ters like accuracy and reliability with precision recall andF-measure as 9608 9510 and 9559 respectively [20])e Piccolo system is used for minimizing the traffic that

VM0VM1

VM2

VM3VM4

VM5

60

50

40

30

20

10

0

Clou

dlet

s

(a)

60

50

40

30

20

10

0

Clou

dlet

s

VM0VM1

VM2

VM3VM4

VM5

(b)

Figure 10 Comparison of VM vs task allocation

120

100

80

60

40

20

0

Task

allo

tmen

t tim

e (se

c)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

Cluster-aware algorithmProposed algorithm

Figure 11 Comparison of task allocation time


occurs during the rollback of the entire VM clusters )ishelps obtain less time and network-related overheadHowever it suffers an overload problem due to the handlingof entire VM clusters [21 22] Cloud radio access network(C-RAN) allocates the VM with low cost and maximumperformance but needs examination with constraints likethose relating to data and capacity [23] Cluster-aware VMclustering with two-tier model clusters does the most fre-quent communication among VMs in the cloud It containstwo stages of the clustering process namely host andpartitioned oriented )is method is restricted to a homo-geneous cloud and does not have application to heteroge-neous cloud [24 25] ACTor algorithm matches with thehistorical pattern to the VM clustering process It uses thepassive way of clustering VM with dynamic task and re-sources and collocated VM [26 27])e proposed algorithm

compared with the cluster-aware algorithm based on thefactors such as allocation time execution time and taskcompletion time of various clusters is shown inFigures 11ndash13 respectively )e proposed algorithm takesless time for task execution compared to cluster-awarealgorithm

8 Conclusion and Future Work

)e focus of the proposed system is on the clustering of VMbased on various performance parameters like the type of therequesting task and bandwidth )e requesting tasks areclassified into CPU-based storage-based and IO-basedmixed types To start with the requesting task is validatedand then the tasks are allocated to the task classificationprocess )e task classification process categorizes the task

8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs


Figure 12 Comparison of task execution time

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)




into different types depending upon the properties whichexist in the task )ese tasks are clustered using the clus-tering algorithm in which the categories are grouped to-gether )ere are two types of clustering process beingcarried out namely preclustering and postclustering )ebandwidth is also clustered based on the task in the taskcluster )ere are two types of clusters maintained in theproposed technique namely the task cluster and thebandwidth cluster with the same features )e VM is clas-sified and mapped to the suitable requesting task for exe-cution )e main objective of the proposed technique is tomap the requesting task to a suitable VM in which thelatency in the service handling is minimized with higherefficiency

Data Availability

)e proposed model uses the CloudSim package which isavailable at httpsgithubcomCloudslabcloudsim )edata are selected for implementation of this model

Conflicts of Interest

All authors declare that there are no conflicts of interest

Authorsrsquo Contributions

All authors have participated in conception and designanalysis and interpretation of the data drafting of the articleand revising it critically for important intellectual contentand approval of the final version

References

[1] Y Tang Y Hu and L Zhang ldquoA classification-based virtualmachine placement algorithm in mobile cloud computingrdquoKSII Transactions on Internet and Information Systemsvol 10 no 5 pp 1998ndash2014 2016

[2] M R Rahimi J Ren C H Liu A V Vasilakos andN Venkatasubramanian ldquoMobile cloud computing a surveystate of art and future directionsrdquo Mobile Networks andApplications vol 19 no 2 pp 133ndash143 2014

[3] M Jo T Maksymyuk B Strykhalyuk and C-H CholdquoDevice-to-device-based heterogeneous radio access networkarchitecture for mobile cloud computingrdquo IEEE WirelessCommunications vol 22 pp 50ndash58 2015

[4] F Xu F Liu H Jin and V V Athanasios ldquoManagingperformance overhead of virtual machines in cloud com-puting a survey state of the art and future directionsrdquoProceedings of the IEEE vol 102 pp 11ndash31 2013

[5] H Yao C Bai D Zeng Q Liang and Y Fan ldquoMigrate or notExploring virtual machine migration in roadside cloudlet-based vehicular cloudrdquo Concurrency and ComputationPractice and Experience vol 27 no 18 pp 5780ndash5792 2015

[6] F Xu F Liu L Liu H Jin and B Li ldquoiAware making livemigration of virtual machines interference-aware in thecloudrdquo IEEE Transactions on Computers vol 63 no 12pp 3012ndash3025 2014

[7] H Ghiasi and M Ghobaei-Arani ldquoSmart virtual machineplacement using learning automata to reduce power con-sumption in cloud data centersrdquo Smart Computing Reviewvol 5 no 6 pp 532ndash562 2015

[8] Y Xia X Li and Z Shan ldquoParallelized fusion on m trans-portation data a case study in CyberITSrdquo InternationalJournal of Intelligent Systems vol 28 no 6 pp 540ndash564 2013

[9] Y Xia T Zhang and S Wang ldquoA generic methodologicalframework for cyber-ITS using cyber-infrastructure in ITSdata analysis casesrdquo Fundamenta Informaticae vol 133 no 1pp 35ndash53 2014

[10] K T Raghavendra ldquoVirtual CPU scheduling techniques forkernel based virtual machine (KVM)rdquo in Proc of IEEE In-ternational Conference on Cloud Computing in EmergingMarkets pp 1ndash6 IEEE Bangalore India October 2013

[11] Y Xia Y Liu Z Ye W Wu and M Zhu ldquoQuadtree-baseddomain decomposition for parallel map-matching on GPSdatardquo in Proceedings of 15th IEEE Conference on IntelligentTransportation Systems Conference pp 808ndash813 IEEE An-chorage AK USA September 2012

[12] Y Mao J Oak A Pompili D Beer T Han and P HuldquoDraps dynamic and resource-aware placement scheme fordocker containers in a heterogeneous clusterrdquo in Proceedingsof the 2017 IEEE 36th International Performance Computingand Communications Conference (IPCCC) December 2017

[13] R Zhang A Zhong B Dong and F Tian ldquoContainer-VM-PMarchitecture a novel architecture for docker container placementrdquoin Proceedings of the International Conference on Cloud Com-puting pp 128ndash140 Springer Saettle WA USA June 2018

[14] Y Fu S Zhang J Terrero Y Mao G Liu S Li et alldquoProgress-based container scheduling for short-lived appli-cations in a kubernetes clusterrdquo in Proceedings of the 2019IEEE International Conference on Big Data (Big Data)pp 278ndash287 IEEE Los Angeles CA USA December 2019

[15] M Lee and Y I Eom ldquoTransaction-aware data cluster allo-cation scheme for qcow2-based virtual disksrdquo in Proceedingsof the 2020 IEEE International Conference on Big Data andSmart Computing (BigComp) pp 1ndash6 IEEE Busan KoreaFebruary 2020

[16] S F Piraghaj R N Calheiros J Chan A V Dastjerdi andR Buyya ldquoVirtual machine customization and task mappingarchitecture for efficient allocation of cloud data center re-sourcesrdquoAeComputer Journal vol 59 no 2 pp 208ndash224 2016

[17] Q Liu C Weng M Li and Y Luo ldquoAn in-VM measuringframework for increasing virtual machine security in cloudsrdquoIEEE Security amp Privacy vol 8 no 6 pp 56ndash62 2010

[18] L Yu and C Weng M Li and Y Luo SNPdisk An efficientpara-virtualization snapshot mechanism for virtual disks inprivate cloudsrdquordquo IEEE Network vol 25 no 4 pp 20ndash26 2011

[19] P Graubner M Schmidt and B Freisleben ldquoEnergy-efficientvirtual machine consolidationrdquo IT Professional vol 15 no 2pp 28ndash34 2013

[20] M Du and F Li ldquoATOM efficient tracking monitoring andorchestration of cloud resourcesrdquo IEEE Transactions onParallel and Distributed Systems vol 28 no 8 2017

[21] L Cui Z Hao and X Yun ldquoPiccolo a fast and efficient rollbacksystem for virtual machine clustersrdquo IEEE Transactions on Paralleland Distributed Systems vol 28 no 8 2017

[22] X Bu J Rao and C-Z Xu ldquoCoordinated self-configurationof virtual machines and appliances using a model-freelearning approachrdquo IEEE Transactions on Parallel and Dis-tributed Systems vol 24 no 4 pp 681ndash690 2013

[23] J Tang PT Wee Q S Q Tony and L Ben ldquosystem costminimization in cloud RAN with limited fronthaul capacityrdquoIEEE Transactions onWireless Communications vol 16 no 52017

[24] W Zhang Y Chen X Gao Q Z ZhichaoMo and Z LuldquoCluster-Aware virtual machine collaborative migration in


media cloudrdquo IEEE Transactions on Parallel and DistributedSystems vol 28 no 10 2017

[25] C Saravanakumar and C Arun ldquoEfficient idle virtual ma-chine management for heterogeneous cloud using commondeployment modelrdquo KSII Transactions on Internet and In-formation Systems vol 10 no 4 pp 1501ndash1518 2016

[26] J Vallone R Birke and L Y Chen ldquoMaking neighbors quietan approach to detect virtual resource contentionrdquo IEEETransactions on Services Computing vol 13 no 5 pp 843ndash856 2020

[27] S T Teixeira R Calheiros and D Gomes ldquoCloud reports anextensible simulation tool for energy-aware cloud computingenvironmentsrdquo in Cloud Computing Computer Communi-cations and Networks Z Mahmood Ed Springer New YorkNY USA 2014


environment VM is a type of access that leads to a per-formance problem seen in the areas of the battery life andenergy consumption)e entire performance factor in greencomputing is used for overcoming this problem [1] Mobilecloud computing (MCC) is a mobile resource sharing servicethat allows mobile devices to have access to the appropriatecloud service It has faced challenges in terms of scalabilitysuch as computational storage services and other differentservices [2]

Mobile computing over cloud has the ability to targetparameters such as traffic quality and customer demandTraditional static cloud and dynamic cloud are compared toanalyze the workload )e static cloud allows users to accessinfrastructure services with a specific configuration whilethe dynamic cloud provides an agile response method toupdate the resource configuration Dynamic cloud has avariety of wireless nodes with device-to-device connectivityfor the achievement of a better utilization of the channel andtraffic [3] VM contention IaaS has created problems in thearea of performance )is problem is overcome through theimplementation of the data center as various ranges such assingle server with virtualization single mega data center andmultiple geo-distributed data centers [4] A researcher en-gaged in cloud suffers from issues that include energyconsumption in data centers )e data center is a key ele-ment in the cloud that handles all kinds of resources neededin the computing environment In the cloud there are twotypes of approaches that are related to hardware and soft-ware)ese approaches require reduced power consumptionin cloud resources without any situation of service un-availability [5] Cloud resources are multiplexed over VMservers designed to host cloud services in large data centers

VMmigration is the process of migrating from one locationto another leading to the performance problem arising as a resultof inference and the cost of the operation iAware imposes themultiresource supply demand model for minimizing the in-ference during the migration [6 7] Intelligent transportationsystems (ITS) are used in the vehicle cloud computing (VCC)architecture which consists of two-layered components such asthe central cloud server and the remote system control (RSC)RSC is a remote administration manager for monitoring andmanaging distributed system elements such as network com-munication lines RSC uses two local server and road side unit(RSU) components If the vehicle travels from one position toanother the VM of one RSU is moved to another RSU Itrequires continuous service response over the automated vehiclecontrol using VCC [5] )ere are two main disputes seen in theintelligent transportation systems (ITS) namely efficiency intraffic and energy quality and productivity )e data collectedfrom various sensors are overcome by using the parallelizedfusion technique )is technique follows the DempsterndashShafertheory with four components namely sensor input boot-strapping hierarchical fusion and state output [6 8] Cyber-ITSis a system that has data division scheduling and efficientsupport through the use of a genericmethodological framework)ere are two types of functions carried out by the systemnamely data-centric and operation-centric transformations)is model uses high-performance computer design with re-gion-based decoupling capability [9 10]

In this method a digital map of the global positioningsystem data is processed in a parallel manner using quad-tree-based domain decomposition technique )ese data aredivided into different subdomains with quad structure [11]Multi-CPU VM scheduling and virtual CPUs (VCPUs)scheduling have been carried out due to the availability ofvarious virtualization techniques in the cloud computingdomain [10] )e existing problems have been analyzedbased on the performance parameters which improve theefficiency in the cloud service deployment )e main ob-jective of the proposed technique is to identify the suitableVM based on the bandwidth and requesting task of allo-cating any issues in the performance of cloud to the task

)e VMs are configured in an isolated fashion whichsuffers from repetitive booting of the respective volumeswith a limited period of delivery )is problem is solved by acluster management approach based on a docker containerwith a diverse configuration [12] )e traditional placementof the docker container and the VM is carried out separatelyso that it is implemented using the container VM-PMmodel[13])e Internet of)ings plays a key role in the processingof real-time data from hardware devices that generate largequantities of data)ese data are stored in a large data centerwith big data analysis methods If the data are huge a hugenumber of servers would be used to store the received data Itfaces an expensive problem that is solved by using a cabinet-based tool called ProCon [14] Virtualization technologyoffers the benefits of the physical server operating on severalcomputers with different resources Virtual storage elimi-nates read and write delays with high IO efficiency Virtualdisks are connected to VMs for processing and storing theuser data via synchronization [15] Amazon cloud providerprovides various kinds of services to the end user in reliableand secure computing capacity AWS offers the ElasticCompute Cloud (EC2) with different versions of VM andresources )e proposed method used EC2 instances as areference for further analysis

)e rest of the paper is organized as follows Section 2depicts the architecture of the proposed system Section 3describes the phases of the proposed system Section 4presents the analysis of various metrics Section 5 describesthe proposed VM clustering process Section 6 provides theexperimental evaluation Section 7 signifies a comparison ofthe various clustering algorithms Section 8 presents theconclusions and indications for future work

2 Model of the Proposed System

Existing VM management techniques configure a VM tocloud workloads based on the energy parameters but theysuffer from a resource wastage problem in the data center)e proposed model groups the tasks based on the VM typeswith bandwidth parameters using classification methods)e VM types are customized based on the resourceavailability in the cloud)e objective of the proposed modelis to map a task to the correct VM by considering variousprocesses such as resources mapping and classification [16])e cloud requesting tasks are classified based on the metricssuch as total number of queues request count API response












Cloud serviceuser

Taskidentification

Taskclassification

Bandwidthselection



Virtual clusters

VC1 VC2 VCn

Virtual machines

Storage Memory

Virtual machine

CPU


User System Cloud

Login ()

Create VM

VM created

View existing VMs


Loading VM list




Existing VM list

VM separated

VM clustered

VM clustering





σActiveVM

21113937VM

1113969⎛⎜⎜⎜⎝ ⎞⎟⎟⎟⎠1113946α




1

σActiveVM

21113937VM

1113969 1113946M


r1113946α


s1113876 1113877 middot S (2)

1

σActiveVM

21113937VM

1113969 1113946M


1

σActive VM21113937

1113968VM

1113946M




μ2n+1 2(2n+12)σ2n+1

ActiveVM1113872 11138731113937VM

1113968 1113946M


(5)


μ2n 22n2σ2n

ActiveVM1113872 11138731113937VM

1113968 1113946M


(6)




ϕ otherwise1113896

(7)


1113896

(8)














VMclusteringi ⋃n
















4 postcopy







(GMT)Peak hours end

(GMT)Avg peakusers

Avg off-peakusers

UB1 0 60 100 3 9 1000 100UB2 1 60 90 4 10 1000 100UB3 2 40 75 5 8 1000 100UB4 3 55 80 7 11 1000 100UB5 4 30 85 8 9 1000 100UB6 5 45 100 6 12 1000 100



35000

30000

25000

20000

15000

10000

5000

0

Num

ber o

f tas

ks

















700

600

500

400

300

200

100

0

Resp

onse

tim

e in

ms









200

180

160

140

120

100

80

60

40

20

0

Tim

e (se

c)

Resource IDVMIDTime




Clusters ActiveVMs

InactiveVMs



VM core countbucket

VM memorybucket

VM lifetime

0 558300 1673700 9177688 072887 8 32 30983331 424500 4254000 3787926 332535 4 32 0252 113310 1133700 0304368 022055 4 32 0166663 0 2591400 9857342 334005 2 4 7198334 228300 2298000 8258144 187629 2 4 041666

700

600

500

400

300

200

100

0

Tim

e (se

c)


Resource IDVMIDTime




1400

1200

1000

800

600

400

200

ndash200

0

Tim

e (se

c)


Resource IDVMIDTime



120

100

80

60

40

20

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(a)

20

15

10

5

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(b)

Figure 8 Continued


250

200

150

100

50

0

Reso

urce

s

10 20 30 40 50 60Time (minutes)

(c)


30252015100500

Virt

ual m

achi

nes

Pow

er (k

W)

120

100

80

60

40

20

00 10 20 30 40 50


Customer 2

Customer 3

(a)

200

150

100

50

0

Clou

dlet

s

Pow

er (k

W)

3035

25201510

50


Customer 2

Customer 3

(b)

100

75

50

25

0

ndash25

Cos

ts (m

u)

Pow

er (k

W) 15

20

10

5

00 10 20 30 40 50


Customer 2

Customer 3

(c)








VM0VM1

VM2

VM3VM4

VM5

60

50

40

30

20

10

0

Clou

dlet

s

(a)

60

50

40

30

20

10

0

Clou

dlet

s

VM0VM1

VM2

VM3VM4

VM5

(b)


120

100

80

60

40

20

0

Task

allo

tmen

t tim

e (se

c)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs








8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs



1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)





Data Availability






References









































Cloud serviceuser

Taskidentification

Taskclassification

Bandwidthselection



Virtual clusters

VC1 VC2 VCn

Virtual machines

Storage Memory

Virtual machine

CPU


User System Cloud

Login ()

Create VM

VM created

View existing VMs


Loading VM list




Existing VM list

VM separated

VM clustered

VM clustering





σActiveVM

21113937VM

1113969⎛⎜⎜⎜⎝ ⎞⎟⎟⎟⎠1113946α




1

σActiveVM

21113937VM

1113969 1113946M


r1113946α


s1113876 1113877 middot S (2)

1

σActiveVM

21113937VM

1113969 1113946M


1

σActive VM21113937

1113968VM

1113946M




μ2n+1 2(2n+12)σ2n+1

ActiveVM1113872 11138731113937VM

1113968 1113946M


(5)


μ2n 22n2σ2n

ActiveVM1113872 11138731113937VM

1113968 1113946M


(6)




ϕ otherwise1113896

(7)


1113896

(8)














VMclusteringi ⋃n
















4 postcopy







(GMT)Peak hours end

(GMT)Avg peakusers

Avg off-peakusers

UB1 0 60 100 3 9 1000 100UB2 1 60 90 4 10 1000 100UB3 2 40 75 5 8 1000 100UB4 3 55 80 7 11 1000 100UB5 4 30 85 8 9 1000 100UB6 5 45 100 6 12 1000 100



35000

30000

25000

20000

15000

10000

5000

0

Num

ber o

f tas

ks

















700

600

500

400

300

200

100

0

Resp

onse

tim

e in

ms









200

180

160

140

120

100

80

60

40

20

0

Tim

e (se

c)

Resource IDVMIDTime




Clusters ActiveVMs

InactiveVMs



VM core countbucket

VM memorybucket

VM lifetime

0 558300 1673700 9177688 072887 8 32 30983331 424500 4254000 3787926 332535 4 32 0252 113310 1133700 0304368 022055 4 32 0166663 0 2591400 9857342 334005 2 4 7198334 228300 2298000 8258144 187629 2 4 041666

700

600

500

400

300

200

100

0

Tim

e (se

c)


Resource IDVMIDTime




1400

1200

1000

800

600

400

200

ndash200

0

Tim

e (se

c)


Resource IDVMIDTime



120

100

80

60

40

20

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(a)

20

15

10

5

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(b)

Figure 8 Continued


250

200

150

100

50

0

Reso

urce

s

10 20 30 40 50 60Time (minutes)

(c)


30252015100500

Virt

ual m

achi

nes

Pow

er (k

W)

120

100

80

60

40

20

00 10 20 30 40 50


Customer 2

Customer 3

(a)

200

150

100

50

0

Clou

dlet

s

Pow

er (k

W)

3035

25201510

50


Customer 2

Customer 3

(b)

100

75

50

25

0

ndash25

Cos

ts (m

u)

Pow

er (k

W) 15

20

10

5

00 10 20 30 40 50


Customer 2

Customer 3

(c)








VM0VM1

VM2

VM3VM4

VM5

60

50

40

30

20

10

0

Clou

dlet

s

(a)

60

50

40

30

20

10

0

Clou

dlet

s

VM0VM1

VM2

VM3VM4

VM5

(b)


120

100

80

60

40

20

0

Task

allo

tmen

t tim

e (se

c)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs








8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs



1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)





Data Availability






References
































σActiveVM

21113937VM

1113969⎛⎜⎜⎜⎝ ⎞⎟⎟⎟⎠1113946α




1

σActiveVM

21113937VM

1113969 1113946M


r1113946α


s1113876 1113877 middot S (2)

1

σActiveVM

21113937VM

1113969 1113946M


1

σActive VM21113937

1113968VM

1113946M




μ2n+1 2(2n+12)σ2n+1

ActiveVM1113872 11138731113937VM

1113968 1113946M


(5)


μ2n 22n2σ2n

ActiveVM1113872 11138731113937VM

1113968 1113946M


(6)




ϕ otherwise1113896

(7)


1113896

(8)














VMclusteringi ⋃n
















4 postcopy







(GMT)Peak hours end

(GMT)Avg peakusers

Avg off-peakusers

UB1 0 60 100 3 9 1000 100UB2 1 60 90 4 10 1000 100UB3 2 40 75 5 8 1000 100UB4 3 55 80 7 11 1000 100UB5 4 30 85 8 9 1000 100UB6 5 45 100 6 12 1000 100



35000

30000

25000

20000

15000

10000

5000

0

Num

ber o

f tas

ks

















700

600

500

400

300

200

100

0

Resp

onse

tim

e in

ms









200

180

160

140

120

100

80

60

40

20

0

Tim

e (se

c)

Resource IDVMIDTime




Clusters ActiveVMs

InactiveVMs



VM core countbucket

VM memorybucket

VM lifetime

0 558300 1673700 9177688 072887 8 32 30983331 424500 4254000 3787926 332535 4 32 0252 113310 1133700 0304368 022055 4 32 0166663 0 2591400 9857342 334005 2 4 7198334 228300 2298000 8258144 187629 2 4 041666

700

600

500

400

300

200

100

0

Tim

e (se

c)


Resource IDVMIDTime




1400

1200

1000

800

600

400

200

ndash200

0

Tim

e (se

c)


Resource IDVMIDTime



120

100

80

60

40

20

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(a)

20

15

10

5

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(b)

Figure 8 Continued


250

200

150

100

50

0

Reso

urce

s

10 20 30 40 50 60Time (minutes)

(c)


30252015100500

Virt

ual m

achi

nes

Pow

er (k

W)

120

100

80

60

40

20

00 10 20 30 40 50


Customer 2

Customer 3

(a)

200

150

100

50

0

Clou

dlet

s

Pow

er (k

W)

3035

25201510

50


Customer 2

Customer 3

(b)

100

75

50

25

0

ndash25

Cos

ts (m

u)

Pow

er (k

W) 15

20

10

5

00 10 20 30 40 50


Customer 2

Customer 3

(c)








VM0VM1

VM2

VM3VM4

VM5

60

50

40

30

20

10

0

Clou

dlet

s

(a)

60

50

40

30

20

10

0

Clou

dlet

s

VM0VM1

VM2

VM3VM4

VM5

(b)


120

100

80

60

40

20

0

Task

allo

tmen

t tim

e (se

c)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs








8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs



1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)





Data Availability






References































VMclusteringi ⋃n
















4 postcopy







(GMT)Peak hours end

(GMT)Avg peakusers

Avg off-peakusers

UB1 0 60 100 3 9 1000 100UB2 1 60 90 4 10 1000 100UB3 2 40 75 5 8 1000 100UB4 3 55 80 7 11 1000 100UB5 4 30 85 8 9 1000 100UB6 5 45 100 6 12 1000 100



35000

30000

25000

20000

15000

10000

5000

0

Num

ber o

f tas

ks

















700

600

500

400

300

200

100

0

Resp

onse

tim

e in

ms









200

180

160

140

120

100

80

60

40

20

0

Tim

e (se

c)

Resource IDVMIDTime




Clusters ActiveVMs

InactiveVMs



VM core countbucket

VM memorybucket

VM lifetime

0 558300 1673700 9177688 072887 8 32 30983331 424500 4254000 3787926 332535 4 32 0252 113310 1133700 0304368 022055 4 32 0166663 0 2591400 9857342 334005 2 4 7198334 228300 2298000 8258144 187629 2 4 041666

700

600

500

400

300

200

100

0

Tim

e (se

c)


Resource IDVMIDTime




1400

1200

1000

800

600

400

200

ndash200

0

Tim

e (se

c)


Resource IDVMIDTime



120

100

80

60

40

20

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(a)

20

15

10

5

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(b)

Figure 8 Continued


250

200

150

100

50

0

Reso

urce

s

10 20 30 40 50 60Time (minutes)

(c)


30252015100500

Virt

ual m

achi

nes

Pow

er (k

W)

120

100

80

60

40

20

00 10 20 30 40 50


Customer 2

Customer 3

(a)

200

150

100

50

0

Clou

dlet

s

Pow

er (k

W)

3035

25201510

50


Customer 2

Customer 3

(b)

100

75

50

25

0

ndash25

Cos

ts (m

u)

Pow

er (k

W) 15

20

10

5

00 10 20 30 40 50


Customer 2

Customer 3

(c)








VM0VM1

VM2

VM3VM4

VM5

60

50

40

30

20

10

0

Clou

dlet

s

(a)

60

50

40

30

20

10

0

Clou

dlet

s

VM0VM1

VM2

VM3VM4

VM5

(b)


120

100

80

60

40

20

0

Task

allo

tmen

t tim

e (se

c)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs








8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs



1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)





Data Availability






References




































4 postcopy







(GMT)Peak hours end

(GMT)Avg peakusers

Avg off-peakusers

UB1 0 60 100 3 9 1000 100UB2 1 60 90 4 10 1000 100UB3 2 40 75 5 8 1000 100UB4 3 55 80 7 11 1000 100UB5 4 30 85 8 9 1000 100UB6 5 45 100 6 12 1000 100



35000

30000

25000

20000

15000

10000

5000

0

Num

ber o

f tas

ks

















700

600

500

400

300

200

100

0

Resp

onse

tim

e in

ms









200

180

160

140

120

100

80

60

40

20

0

Tim

e (se

c)

Resource IDVMIDTime




Clusters ActiveVMs

InactiveVMs



VM core countbucket

VM memorybucket

VM lifetime

0 558300 1673700 9177688 072887 8 32 30983331 424500 4254000 3787926 332535 4 32 0252 113310 1133700 0304368 022055 4 32 0166663 0 2591400 9857342 334005 2 4 7198334 228300 2298000 8258144 187629 2 4 041666

700

600

500

400

300

200

100

0

Tim

e (se

c)


Resource IDVMIDTime




1400

1200

1000

800

600

400

200

ndash200

0

Tim

e (se

c)


Resource IDVMIDTime



120

100

80

60

40

20

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(a)

20

15

10

5

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(b)

Figure 8 Continued


250

200

150

100

50

0

Reso

urce

s

10 20 30 40 50 60Time (minutes)

(c)


30252015100500

Virt

ual m

achi

nes

Pow

er (k

W)

120

100

80

60

40

20

00 10 20 30 40 50


Customer 2

Customer 3

(a)

200

150

100

50

0

Clou

dlet

s

Pow

er (k

W)

3035

25201510

50


Customer 2

Customer 3

(b)

100

75

50

25

0

ndash25

Cos

ts (m

u)

Pow

er (k

W) 15

20

10

5

00 10 20 30 40 50


Customer 2

Customer 3

(c)








VM0VM1

VM2

VM3VM4

VM5

60

50

40

30

20

10

0

Clou

dlet

s

(a)

60

50

40

30

20

10

0

Clou

dlet

s

VM0VM1

VM2

VM3VM4

VM5

(b)


120

100

80

60

40

20

0

Task

allo

tmen

t tim

e (se

c)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs








8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs



1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)





Data Availability






References











































700

600

500

400

300

200

100

0

Resp

onse

tim

e in

ms









200

180

160

140

120

100

80

60

40

20

0

Tim

e (se

c)

Resource IDVMIDTime




Clusters ActiveVMs

InactiveVMs



VM core countbucket

VM memorybucket

VM lifetime

0 558300 1673700 9177688 072887 8 32 30983331 424500 4254000 3787926 332535 4 32 0252 113310 1133700 0304368 022055 4 32 0166663 0 2591400 9857342 334005 2 4 7198334 228300 2298000 8258144 187629 2 4 041666

700

600

500

400

300

200

100

0

Tim

e (se

c)


Resource IDVMIDTime




1400

1200

1000

800

600

400

200

ndash200

0

Tim

e (se

c)


Resource IDVMIDTime



120

100

80

60

40

20

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(a)

20

15

10

5

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(b)

Figure 8 Continued


250

200

150

100

50

0

Reso

urce

s

10 20 30 40 50 60Time (minutes)

(c)


30252015100500

Virt

ual m

achi

nes

Pow

er (k

W)

120

100

80

60

40

20

00 10 20 30 40 50


Customer 2

Customer 3

(a)

200

150

100

50

0

Clou

dlet

s

Pow

er (k

W)

3035

25201510

50


Customer 2

Customer 3

(b)

100

75

50

25

0

ndash25

Cos

ts (m

u)

Pow

er (k

W) 15

20

10

5

00 10 20 30 40 50


Customer 2

Customer 3

(c)








VM0VM1

VM2

VM3VM4

VM5

60

50

40

30

20

10

0

Clou

dlet

s

(a)

60

50

40

30

20

10

0

Clou

dlet

s

VM0VM1

VM2

VM3VM4

VM5

(b)


120

100

80

60

40

20

0

Task

allo

tmen

t tim

e (se

c)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs








8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs



1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)





Data Availability






References



































200

180

160

140

120

100

80

60

40

20

0

Tim

e (se

c)

Resource IDVMIDTime




Clusters ActiveVMs

InactiveVMs



VM core countbucket

VM memorybucket

VM lifetime

0 558300 1673700 9177688 072887 8 32 30983331 424500 4254000 3787926 332535 4 32 0252 113310 1133700 0304368 022055 4 32 0166663 0 2591400 9857342 334005 2 4 7198334 228300 2298000 8258144 187629 2 4 041666

700

600

500

400

300

200

100

0

Tim

e (se

c)


Resource IDVMIDTime




1400

1200

1000

800

600

400

200

ndash200

0

Tim

e (se

c)


Resource IDVMIDTime



120

100

80

60

40

20

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(a)

20

15

10

5

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(b)

Figure 8 Continued


250

200

150

100

50

0

Reso

urce

s

10 20 30 40 50 60Time (minutes)

(c)


30252015100500

Virt

ual m

achi

nes

Pow

er (k

W)

120

100

80

60

40

20

00 10 20 30 40 50


Customer 2

Customer 3

(a)

200

150

100

50

0

Clou

dlet

s

Pow

er (k

W)

3035

25201510

50


Customer 2

Customer 3

(b)

100

75

50

25

0

ndash25

Cos

ts (m

u)

Pow

er (k

W) 15

20

10

5

00 10 20 30 40 50


Customer 2

Customer 3

(c)








VM0VM1

VM2

VM3VM4

VM5

60

50

40

30

20

10

0

Clou

dlet

s

(a)

60

50

40

30

20

10

0

Clou

dlet

s

VM0VM1

VM2

VM3VM4

VM5

(b)


120

100

80

60

40

20

0

Task

allo

tmen

t tim

e (se

c)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs








8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs



1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)





Data Availability






References































1400

1200

1000

800

600

400

200

ndash200

0

Tim

e (se

c)


Resource IDVMIDTime



120

100

80

60

40

20

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(a)

20

15

10

5

0

Reso

urce

s

0 10 20 30 40 50Time (minutes)

(b)

Figure 8 Continued


250

200

150

100

50

0

Reso

urce

s

10 20 30 40 50 60Time (minutes)

(c)


30252015100500

Virt

ual m

achi

nes

Pow

er (k

W)

120

100

80

60

40

20

00 10 20 30 40 50


Customer 2

Customer 3

(a)

200

150

100

50

0

Clou

dlet

s

Pow

er (k

W)

3035

25201510

50


Customer 2

Customer 3

(b)

100

75

50

25

0

ndash25

Cos

ts (m

u)

Pow

er (k

W) 15

20

10

5

00 10 20 30 40 50


Customer 2

Customer 3

(c)








VM0VM1

VM2

VM3VM4

VM5

60

50

40

30

20

10

0

Clou

dlet

s

(a)

60

50

40

30

20

10

0

Clou

dlet

s

VM0VM1

VM2

VM3VM4

VM5

(b)


120

100

80

60

40

20

0

Task

allo

tmen

t tim

e (se

c)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs








8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs



1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)





Data Availability






References































250

200

150

100

50

0

Reso

urce

s

10 20 30 40 50 60Time (minutes)

(c)


30252015100500

Virt

ual m

achi

nes

Pow

er (k

W)

120

100

80

60

40

20

00 10 20 30 40 50


Customer 2

Customer 3

(a)

200

150

100

50

0

Clou

dlet

s

Pow

er (k

W)

3035

25201510

50


Customer 2

Customer 3

(b)

100

75

50

25

0

ndash25

Cos

ts (m

u)

Pow

er (k

W) 15

20

10

5

00 10 20 30 40 50


Customer 2

Customer 3

(c)








VM0VM1

VM2

VM3VM4

VM5

60

50

40

30

20

10

0

Clou

dlet

s

(a)

60

50

40

30

20

10

0

Clou

dlet

s

VM0VM1

VM2

VM3VM4

VM5

(b)


120

100

80

60

40

20

0

Task

allo

tmen

t tim

e (se

c)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs








8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs



1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)





Data Availability






References




































VM0VM1

VM2

VM3VM4

VM5

60

50

40

30

20

10

0

Clou

dlet

s

(a)

60

50

40

30

20

10

0

Clou

dlet

s

VM0VM1

VM2

VM3VM4

VM5

(b)


120

100

80

60

40

20

0

Task

allo

tmen

t tim

e (se

c)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs








8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs



1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)





Data Availability






References



































8000

7000

6000

5000

4000

3000

2000

1000

0

Task

exec

utio

n tim

e (μs

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53Number of VMs



1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51Number of VMs

4000

30003500

2500200015001000

5000C

ost o

f tas

k ex

ecut

ion

(sec

)





Data Availability






References
































Data Availability






References































an efficient technique for virtual machine clustering and

Documents