Research Article Virtual Machine Placement Algorithm for ...downloads.hindawi.com/journals/sp/2016/5612039.pdf · ciency means less energy consumption and SLA violation in data centers

Research ArticleVirtual Machine Placement Algorithm forBoth Energy-Awareness and SLA ViolationReduction in Cloud Data Centers

Zhou Zhou1 Zhigang Hu1 and Keqin Li2

1School of Software Central South University Changsha 410083 China2Department of Computer Science State University of New York New Paltz NY 12561 USA

Correspondence should be addressed to Zhou Zhou zhouzhou03201126com

Received 7 February 2016 Accepted 10 March 2016

Academic Editor Laurence T Yang

Copyright copy 2016 Zhou Zhou et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

The problem of high energy consumption is becoming more and more serious due to the construction of large-scale cloud datacenters In order to reduce the energy consumption and SLA violation a new virtual machine (VM) placement algorithm namedATEA (adaptive three-threshold energy-aware algorithm) which takes good use of the historical data from resource usage by VMsis presented In ATEA according to the load handled data center hosts are divided into four classes hosts with little load hostswith light load hosts with moderate load and hosts with heavy load ATEA migrates VMs on heavily loaded or little-loaded hoststo lightly loaded hosts while the VMs on lightly loaded and moderately loaded hosts remain unchanged Then on the basis ofATEA two kinds of adaptive three-threshold algorithm and three kinds of VMs selection policies are proposed Finally we verifythe effectiveness of the proposed algorithms by CloudSim toolkit utilizing real-world workloadThe experimental results show thatthe proposed algorithms efficiently reduce energy consumption and SLA violation

1 Introduction

Cloud computing [1 2] is derived from grid computing Atpresent cloud computing is receiving more and more atten-tion through which people can access resources in a simpleway In contrast to previous paradigms cloud computing pro-vides infrastructure as a service (IaaS) platform as a service(PaaS) and software as a service (SaaS)

On one hand the construction of a large-scale virtualizeddata centers meets the demand of computational power onthe other hand such data centers consume a great manyof electrical energy resources leading to high energy con-sumption and carbon dioxide emissions It has been reportedthat [3] in 2013 the total electricity consumption of globaldata center was more than 435 gigawatts and annual growthrate was by 15 The high energy consumption problem ofvirtualized data centers causes a series of problems includingenergy wastes low Return on the Investment (ROI) systeminstability and more carbon dioxide emissions [4]

However most hosts in data centers are in a state oflow CPU utilization Barroso and Holzle [5] took a surveyover half a year and found that most hosts in data centersoperate at lower than 50 CPU utilization Bohrer et al[6] investigated the problem of high energy consumptionand came to the same conclusion Therefore it is extremelynecessary to reduce the energy consumption of data centerswhile keeping low SLA (Service Level Agreement) violation[7]

In this paper we put forward a new VM deploymentalgorithm (ATEA) two kinds of adaptive three-thresholdalgorithm (KAM and KAI) and three kinds of VM selectionpolicies to reduce energy consumption and SLA violationWeverify the effectiveness of the proposed algorithms throughusing the CloudSim toolkit

The main contributions of the paper are summarized asfollows

(i) Proposing a novel VM deployment algorithm(ATEA) In ATEA hosts in a data center are divided

Hindawi Publishing CorporationScientific ProgrammingVolume 2016 Article ID 5612039 11 pageshttpdxdoiorg10115520165612039

2 Scientific Programming

into four classes according to their load ATEA mi-grates VMs on heavily loaded or little-loaded hosts tolightly loaded hosts while the VMs on lightly loadedand moderately loaded hosts remain unchanged

(ii) Presenting two kinds of adaptive three-thresholdalgorithm to determine the three thresholds

(iii) Putting forward three kinds of VMs selection policiesand making three paired 119905-tests

(iv) Evaluating the proposed algorithms by extensive sim-ulation using the CloudSim toolkit

The rest of this paper is organized as follows In Section 2 therelatedwork is discussed Section 3 presents the powermodelcost of VM migration SLA violation metrics and energyefficiency metrics Section 4 proposes ATEA two kinds ofadaptive three-threshold algorithm VM selection policy andVM deployment algorithm Experiments and performanceevaluation are presented in Section 5 Section 6 concludes thepaper

2 Related Work

At present there are various studies focusing on energy effi-cient resource management in cloud data centers Constraintenergy consumption algorithm [8ndash10] and energy efficiencyalgorithm [11ndash15] are twomain types of algorithms for solvingthe problem of high energy consumption in data centersThemain idea of the constraint energy consumption algorithm isto reduce the energy consumption in data centers but thistype of algorithm focuses a little on (does not consider) theSLA violation For example Lee and Zomaya [8] proposedtwo heuristic algorithms (ECS ECS + idle) to decrease theenergy consumption but the two algorithms are easy to fallinto local optimum and do not consider the SLA violationHanson et al [9] presented Dynamic Voltage and FrequencyScaling (DVFS) policy to save power in data centers Whenthe task number is large DVFS policy raises the voltage of theprocessor in order to deal with the task in time when the tasknumber is small DVFS policy decreases the voltage of pro-cessor for the purpose of saving power Kang and Ranka [10]put forward an energy-saving algorithm and they supposedthat overestimated or underestimated execution time of tasksis bad for energy-saving algorithm For overestimation theextra available time should be assigned to other tasks in orderto reduce energy consumption Similarly this energy-savingalgorithm does not consider the SLA violation Thereforethe constraint energy consumption algorithm does not meetthe requirement of users because of focusing a little on (notconsidering) the SLA violation

The goal of the energy efficiency (energy consumptionand SLA violation) algorithm is to decrease the energyconsumption and SLA violation in data centers For exampleBuyya et al [11] raised a virtual machine (VM) placementalgorithm (called Single Threshold (ST)) based on the com-bination of VM selection policies ST algorithm sets a unifiedvalue for all serversrsquo CPU utilization to make sure all serversare below this value Obviously ST algorithm can save energyconsumption and decrease the SLA violation but the SLA

violation remains at a high level Beloglazov and Buyya [12]proposed an energy efficient resource management systemwhich includes the dispatcher globalmanager localmanagerand VMM (VM Monitor) In order to improve the energyefficiency Beloglazov et al put forward a new VMmigrationalgorithm called Double Threshold (DT) [13] DT sets twothresholds to keep all hostsrsquo CPU utilization between the twothresholds However the energy consumption and SLA vio-lation for DT algorithm need to be further decreased LaterBeloglazov and Buyya [14 15] proposed an adaptive double-threshold VM placement algorithm to improve energy effi-ciency in data centers However the energy consumption indata centers remains at a high level

In our previous study [16] we proposed a three-thresholdenergy-aware algorithm named MIMT to deal with theenergy consumption and SLA violation However the threethresholds for controlling hostrsquos CPU utilization are fixedTherefore MIMT is not suitable for varying workloadTherefore it is necessary to put forward a novel VM place-ment algorithm to deal with energy consumption and SLAviolation in cloud data centers

3 The Power Model Cost ofVM Migration SLA Violation Metricsand Energy Efficiency Metrics

31 The Power Model Energy consumption by servers indata centers is connected with its CPU memory disk andbandwidth Recent studies [17 18] have illustrated that theenergy consumption by servers has a linear relationshipwith its CPU utilization even DVFS policy is appliedHowever with the decrease of hardware price multicoreCPUs and memory with large-capacity are widely equippedin servers leading to the traditional linear model not beingable to accurately describe energy consumption of serversIn order to deal with this problem we use the real dataof energy consumption offered by SPECpower benchmark(httpwwwspecorgpower ssj2008)

We have chosen two servers equipped with dual-coreCPUsThemain configuration of the two servers is as followsone is HP ProLiant G4 equipped with 186GHz (dual-core)4GB RAM the other is HP ProLiant G5 equipped with266GHz (dual-core) 4GB RAM The energy consumptionfor the two servers at different load levels is listed in Table 1[15]

32 Cost of VM Migration Proper VM migration betweenservers can reduce energy consumption and SLA violation indata centers However excessive VM migration could bringnegative impact on performance of applicationwhich runs onthe VMs Voorsluys et al [19] investigated the cost problemof VM migration and the performance degradation causedby VM can be described in

119862 = 119896 sdot int

1199050+119879119898119895

1199050

119906119895(119905) 119889119905

119879119898119895

=

119872119895

119861119895

(1)

Scientific Programming 3

Table 1 Power consumption by the two servers at different load levels in Watts

Server 0 10 20 30 40 50 60 70 80 90 100HP G4 86 894 926 96 995 102 106 108 112 114 117HP G5 937 97 101 105 110 116 121 125 129 133 135

where parameter119862 represents total performance degradationcaused by VM 119895 parameter 119896 is the coefficient of averageperformance degradation caused by VMs (in terms of theclass of web-applications the value of 119896 could be estimated asapproximately 01 (10) of the CPU utilization [19]) function119906119895(119905) corresponds to the CPU utilization of VM 119895 parameter

1199050represents start time of migration 119879

119898119895is completion time

119872119895corresponds to the total memory used by VM 119895 and 119861

119895

represents the available bandwidth

33 SLAViolationMetrics SLA violation is extremely impor-tant factor for any VMmigration algorithm At present thereare two ways to describe the SLA violations

(1) PDM [15] (Overall Performance Degradation Caused byVMMigration) It is indicated in

PDM =1

119872

119872

sum

119895=1

119862119889119895

119862119903119895

(2)

where parameter 119872 represents the number of VMs in datacenter 119862

119889119895means the estimate of the performance degrada-

tion caused by VM 119895 migration and 119862119903119895corresponds to the

total CPU capacity requested by VM 119895 during its lifetime

(2) SLATAH [15] (SLA Violation Time per Active Host) Itmeans the percentage of total SLA violation time duringwhich active hostrsquos CPU utilization has experienced 100 asindicated in

SLATAH =1

119873

119873

sum

119894=1

119879119904119894

119879119886119894

(3)

where 119873 represents the number of hosts in data center119879119904119894corresponds to the total time during which the CPU

utilization of host 119894 has experienced 100 utilization resultingin the SLA violations119879

119886119894corresponds to the total time of host

119894 being in active state The reasoning behind the SLATAH isthat the active hostrsquos CPU utilization has experienced 100utilization the VMs on the host could not be provided withthe requested CPU capacity

Both PDM and SLATAH are two effective methods toindependently evaluate the SLA violationTherefore the SLAviolation is defined as in [15]

SLA = PDM times SLATAH (4)

34 Energy Efficiency Metric Energy efficiency includes en-ergy consumption and SLA violation Improving the energyefficiency means less energy consumption and SLA violation

in data centers Therefore the metric of energy efficiency isdefined as in

119864 =1

119875 times SLA (5)

where 119864 corresponds to the energy efficiency of a data center119875 means the energy consumption of a data center and SLArepresents the SLA violation of a data center Equation (5)shows that the higher the 119864 the more the energy efficiency

4 ATEA Two Kinds of AdaptiveThree-Threshold Algorithm VM SelectionPolicy and VM Deployment Algorithm

41 ATEA VM migration is an effective method to improvethe energy efficiency in data centers However there areseveral key problems which should be dealt with (1) when ahost is supposed to be heavily loaded where some VMs fromthe host should bemigrated to another host (2)when a host isbelieved to be moderately loaded or lightly loaded resultingin a decision to keep all VMs on this host unchanged (3)when a host is believed to be little-loaded where all VMson the host must be migrated to another host (4) selectinga VM or more VMs that should be migrated from theheavily loaded (5) finding a new host to accommodate VMsmigrated from heavily loaded or little-loaded hosts

In order to solve the above problems ATEA (adaptivethree-threshold energy-aware algorithm) is proposed ATEAautomatically sets three thresholds 119879

119897 119879119898 and 119879

ℎ(0 le 119879

119897lt

119879119898

lt 119879ℎ

le 1) leading to the data center hosts to bedivided into four classes hosts with little load hosts withlight load hosts with moderate load and hosts with heavyload When CPU utilization of a host is less than or equalto 119879119897 the host is believed to be little-loaded In order to

save energy consumption all VMs on little-loaded host mustbe migrated to another host with light load and the host isswitched to sleep mode when CPU utilization of a host isbetween 119879

119897and 119879

119898 the host is considered as being lightly

loaded In order to avoid high SLA violations all VMs onlightly loaded host have to be kept unchanged the reasonis that excessive VMs migration leads to the performancedegradation and high SLA violations when CPU utilizationof a host is between 119879

119898and 119879

ℎ the host is believed to be

moderately loaded all VMs on moderately loaded host haveto be kept unchanged for the reason that excessive VMsmigration leads to the performance degradation and highSLA violations when CPU utilization of a host is greater than119879ℎ the host is considered as being heavily loaded in order

to reduce the SLA violations some VMs on heavily loadedhostmust bemigrated to another host with light load Figure 1shows the flow chart of algorithm ATEA


Begin

No Yes

No

No

Yes

Yes

Get value of hostsrsquo CPUutilization (u)

End

Obtain the value of Tl Tmand Th

u gt Tl

u gt Tm

u gt Th

All VMs on

VMs on thehost are keptunchanged

VMs on the Some VMs on thehost are kept host are migrated tounchanged

the host aremigrated toanother host

another host with

with light load

light load

Figure 1 Flow chart of ATEA

Different from out previous study [16] where the thresh-olds are fixed in this paper the three thresholds 119879

119897 119879119898 and

119879ℎin ATEA are not fixed and these values can be autoadjusted

according to workloadATEA migrates VMs that must be migrated to other

hosts while keeping some VMs unchanged In doing soATEA improves the migration efficiency of VMs ThereforeATEA is a fine-grained algorithm However two problemsshould be solved as for ATEA Firstly what are the thresholdvalues of 119879

119897 119879119898 and 119879

ℎ This problem will be discussed in

Section 42 Secondly as mentioned above some VMs onheavily loaded host must be migrated to another host withlight load Which VM should be migrated This issue will bediscussed in Section 43

TheVMplacement optimization of ATEA is illustrated inAlgorithm 1

In the first stage the algorithm inspects each host in hostlist and decides which host is heavily loaded If the host isheavily loaded (corresponding to Line 2 in Algorithm 1) thealgorithm uses the VM selection policy to choose VMswhichmust be migrated from the host (corresponding to Line 6 inAlgorithm 1) Once VMs list that should be migrated fromthe heavily loaded is created the VM deployment algorithmis invoked for the purpose of finding a new host to accommo-date the VM (corresponding to Line 7 in Algorithm 1) Func-tion ldquogetNewVmPlacement(vmsToMigrate)rdquo means to find anew host to accommodate the VM In the second stage the

algorithm inspects each host in host list and decides whichhost is little-loaded If the host is little-loaded (correspondingto Line 11 in Algorithm 1) the algorithm chooses all VMsfrom the host to migrate and finds a placement of the VMs(corresponding to Line 15-Line 16 in Algorithm 1) At last thealgorithm returns the migration map

42 Two Kinds of Adaptive Three-Threshold Algorithm Asdiscussed in Section 41 what are the threshold values of119879119897 119879119898 and 119879

ℎ To solve this problem two adaptive three-

threshold algorithms (KAM and KAI) are proposed

421 KAM (119870-Means Clustering Algorithm-Average-MedianAbsolute Deviation) For a univariate data set 119881

1 1198812 1198813

119881119899(119881119894is a hostrsquos CPU utilization at time 119894 and the size of 119899

could be determined by empirical value) the KAMalgorithmfirstly uses the 119870-means clustering algorithm to divide thedata set (119881

1 1198812 1198813 119881

119899) into 119898 groups (119866

1 1198662 119866

119898)

(the size of 119898 could be obtained by empirical value and inthis paper 119898 = 5) where 119866

119896= (119881119895119896minus1+1

119881119895119896minus1+2

119881119895119896) for

all 1 le 119896 le 5 and 0 = 1198950lt 1198951lt 1198952lt sdot sdot sdot lt 119895

5= 119899Then KAM

gets the average value of each group formalized as follows

119866119860119896

=

(119881119895119896minus1+1

+ 119881119895119896minus1+2

+ sdot sdot sdot + 119881119895119896)

(119895119896minus 119895119896minus1

) (6)


Require hostlist hostlist is the set of all hostsEnsure migrationMap(1) for each host in hostlist do(2) if isHostHeavilyLoaded(host) then(3) HeavilyLoadedHostsadd(host) host is heavily-loaded(4) end if(5) end for(6) vmsToMigrate = getVmsFromHeavilyLoadedHosts(HeavilyLoadedHosts)(7) migrationMapaddAll(getNewVmPlacement(vmsToMigrate))(8) HeavilyLoadedHostsclear( )(9) vmsToMigrateclear( )(10) for each host in hostlist do(11) if isHostLittleLoaded(host) then(12) LittleLoadedHostsadd(host) host is little-loaded(13) end if(14) end for(15) vmsToMigrate = getVmsToMigrateFromHosts(LittleLoadedHosts)(16) migrationMapaddAll(getNewVmPlacement(vmsToMigrate))(17) returnmigrationMap

Algorithm 1 Optimized allocation of VMs

for all 1 le 119896 le 5 Then KAM gets the Median AbsoluteDeviation (MAD) of 119866(119866

1198601 1198661198602 119866

1198605) Therefore the

MAD is defined as follows

MAD = median119860119901

(10038161003816100381610038161003816119866119860119901

minusmedian119860119902

(119866119860119902)10038161003816100381610038161003816) (7)

where 1198601 le 119860119901 le 1198605 and median119860119902(119866119860119902) is median value

of119866119860119902 Finally the three thresholds (119879

119897 119879119898 and 119879

ℎ) in ATEA

could be defined as follows

119879119897= 05 (1 minus 119903 timesMAD) (8)

119879119898= 09 (1 minus 119903 timesMAD) (9)

119879ℎ= 1 minus 119903 timesMAD (10)

where 119903 isin 119877+ represents a parameter of the algorithm that

defines how aggressively the system consolidates VMs Forexample the higher 119903 the more energy consumption butthe less SLA violations caused by VMs consolidation Thecomplexity of KAM is 119874(119898 times 119899 times 119905) where 119898 is the groupnumber 119899 denotes the data size and 119905 is the iteration number

As the value of 119881119894(119894 = 1 2 3 119899) varies from time to

time the value of 119879119897 119879119898 and 119879

ℎare also variable Therefore

KAM is an adaptive three-threshold algorithm When theworkloads are dynamic and unpredictable as compared witha fixed threshold algorithm KAM generates higher energyefficiency by setting the value of 119879

119897 119879119898 and 119879

ℎ

422 KAI (119870-Means Clustering Algorithm-Average-Inter-quartile Range) KAI is another adaptive threshold algo-rithm For a univariate data set 119881

1 1198812 1198813 119881

119899(119881119894is a

hostrsquos CPU utilization at time 119894 and the size of 119899 could bedetermined by empirical value) the KAI algorithm firstlyuses the 119870-means clustering algorithm to divide the data set(1198811 1198812 1198813 119881

119899) into119898 groups (119866

1 1198662 119866

119898) (the size of

119898 could be obtained by empirical value and in this paper119898 = 5) where119866

119896= (119881119895119896minus1+1

119881119895119896minus1+2

119881119895119896) for all 1 le 119896 le 5

and 0 = 1198950lt 1198951lt 1198952lt sdot sdot sdot lt 119895

5= 119899 Then KAI gets the

average value of each group formalized as follows

119866119860119896

=

(119881119895119896minus1+1

+ 119881119895119896minus1+2

+ sdot sdot sdot + 119881119895119896)

(119895119896minus 119895119896minus1

) (11)

for all 1 le 119896 le 5 Then KAI gets the Interquartile Range(IR) of 119866(119866

1198601 1198661198602 119866

1198605) Therefore the IR is defined as

follows

IR = 1198763minus 1198761 (12)

where 1198763is third quartiles of 119866 and 119876

1is first quartiles of 119866

Finally the three thresholds (119879119897 119879119898 and 119879

ℎ) in ATEA can be

defined as follows

119879119897= 05 (1 minus 119903 times IR) (13)

119879119898= 09 (1 minus 119903 times IR) (14)

119879ℎ= 1 minus 119903 times IR (15)


defines how aggressively the system consolidates VMs Forexample the higher 119903 the more energy consumption butthe less SLA violations caused by VMs consolidation Thecomplexity of KAI is 119874(119898 times 119899 times 119905) where 119898 is the groupnumber 119899 denotes the data size and 119905 is the iteration number


time the values of 119879119897 119879119898 and 119879


KAI is an adaptive three-threshold algorithm When theworkloads are dynamic and unpredictable as comparedwith fixed threshold algorithm KAI generates higher energyefficiency by setting the value of 119879

119897 119879119898 and 119879

ℎ


43 VM Selection Policies As described earlier in Section 41some VMs on heavily loaded host must be migrated toanother host with light loadWhich VM should be migratedIn general a hostrsquos CPU utilization and memory size affectits energy efficiency Therefore to solve this problem threekinds of VM selection policies (MMS LCU and MPCM) areproposed in this section

431 MMS (Minimum Memory Size) Policy The migrationtime of a VMwill change depending on its different memorysize A VM with less memory size means less migration timeunder the same spare network bandwidth For example aVM with 16GB memory may take 16 timesrsquo longer migrationtime than a VMwith 1GBmemory Clearly selecting the VMwith 16GB memory or the VM with 1GB memory greatlyaffects energy efficiency of data centers Therefore if a hostis being heavily loaded MMS policy selects a VM with theminimum memory size to migrate compared with the otherVMs allocated to the host MMS policy chooses a VM 119906 thatsatisfies the following condition

RAM (119906) le RAM (V) forallV isin VM119894 (16)

where VM119894means the set of VMs allocated to host 119894 and

RAM(119906) is the memory size currently utilized by the VM 119906

432 LCU (Lowest CPU Utilization) Policy As for energyefficiency in data center the CPU utilization of a host isalso another important factor Therefore if a host is beingheavily loaded LCU policy chooses a VM with the lowestCPU utilization to migrate compared with the other VMsallocated to the host LCUpolicy chooses aVM119906 that satisfiesthe following condition

Utilization (119906) le Utilization (V) forallV isin VM119894 (17)


Utilization(119906) is theCPUutilization ofVM 119906 allocated to host119894

433 MPCM (Minimum Product of Both CPUUtilization andMemory Size) Policy As hostrsquos CPU utilization and memorysize are two important factors for energy efficiency in datacenter if a host is being heavily loaded MPCM policy selectsa VM with the minimum product of both CPU utilizationand memory size to migrate compared with the other VMsallocated to the host MPCM policy chooses a VM 119906 thatsatisfies the following condition

(119906CPU times 119906memory) le (VCPU times Vmemory) forallV isin VM119894 (18)


119906CPU and 119906memory respectively represent CPU utilization andmemory size currently utilized by the VM 119906

44 VM Deployment Algorithm VM deployment can beconsidered as a bin packing problem In order to tackle theproblem a modification of the Best Fit Decreasing algorithmdenoted by Energy-Aware Best Fit Decreasing (EBFD) could

be used to deal with it As described earlier in Section 41VMs on heavily loaded host or VMs on little-loaded hostmust be migrated to another host with light load (CPUutilization of a host at 119879

119897-119879119898interval) VMs on lightly loaded

host or VMs on moderately loaded host are kept unchangedAlgorithm 2 shows the pseudocode of EBFD where 119879

119897

119879119898 and 119879

ℎare three thresholds in ATEA (the definition

of the three thresholds in Section 42) ldquovmlistrdquo is theset of all VMs ldquohostlistrdquo represents all hosts in the datacenters Line 1 (see Algorithm 2) means to sort all VMby CPU utilization in descending order Line 3 representsthat parameter ldquominimumPowerrdquo is assigned a maximumvalue Line 6 is to check whether the host is suitable toaccommodate the VM (eg hostrsquos CPU capacity memorysize and bandwidth) Function getUtilizationAfterAllocationmeans to obtain hostrsquos CPU utilization after allocating a VMLine 7 to Line 9 (see Algorithm 2) are to keep a host with lightload (CPU utilization of a host at 119879

119897-119879119898interval) Function

getPowerAfterVM is to obtain the increasing of hostrsquos energyconsumption after allocating a VM Line 11 to Line 15 areto find the host which owns the least increasing of powerconsumption caused byVMallocation Line 19 is to return thedestination hosts for accommodating VMs The complexityof the EBFD is119874(119898times119899) parameter119898 represents the numberof hosts whereas parameter 119899 corresponds to the number ofVMs which should be allocated

5 Experiments and Performance Evaluation

51 Experiment Setup Due to the advantages of CloudSimtoolkit [20 21] such as supporting on demand dynamicresource provisioning and modeling of virtualized environ-ments and so on we choose it as the simulation toolkit forour experiments

We have simulated a data center which includes 800heterogeneous physical nodes half of which consists of HPProLiant G4 (Intel Xeon 3040 2 cores 1860MHz 4GB)and the other half are HP ProLiant G5 (Intel Xeon 3075 2cores 2660MHz 4GB) There are 1052 VMs and four kindsof VM types (High-CPU Medium Instance (2500MIPS085GB) Extra Large Instance (2000MIPS 375GB) SmallInstance (1000MIPS 17 GB) andMicro Instance (500MIPS613MB)) in the data center The characteristics of the VMscorrespond to Amazon EC2 instance types

52 Workload Data Using real workload to do experimentsis extremely significant for VM placement In this paper weutilize the workload coming from a CoMon project whichmainlymonitors infrastructure for PlanetLab [22]We obtainthe data derived from more than a thousand VMsrsquo CPUutilization and theVMs placed atmore than 500 places acrossthe world Table 2 [15] shows the characteristics of the data

53 Experimental Results and Analysis By using the work-load data (described in Section 52) two kinds of adaptivethree-threshold algorithm (KAM and KAI) and three kindsof VM selection policies (MMS LCU and MPCM) aresimulated In addition for the two adaptive three-threshold


Require 119879119897 119879119898 119879ℎ vmlist hostlist

Ensure migrationMap(1) vmListsortByCpuUtilization( )

sorted by CPU utilizatioin in descending order(2) for each vm in vmlist do(3) minimumPower = maximum

minimunPower is assigned a maximum value(4) allocatedHost = null(5) for each host in hostlist do(6) if (host is Suitable for Vm (vm)) then(7) utilization = getUtilizationAfterAllocation(host vm)(8) if ((utilization lt 119879

119897) || (utilization gt 119879

119898)) then

(9) continue(10) end if(11) EnergyConsumption = getPowerAfterVM(host vm)(12) if (EnergyConsumption ltminimumPower) then(13) minimumPower = EnergyConsumption(14) allocatedHost = host(15) end if(16) end if(17) end for(18) end for(19) return allocationHost

Algorithm 2 Energy-Aware Best Fit Decreasing (EBFD)

Table 2 Workload data characteristics (CPU utilization)

Date Number of VMs Mean St dev Quartile 1 Median Quartile 303032011 1052 1231 1709 2 6 15

algorithm we have varied the parameters (119903 (7)ndash(9)) and((11)ndash(13)) form 05 to 30 increasing by 05 Therefore thesevariations have led to 36 combinations of the algorithmsand parameters For example for KAM there are threeVM selection policies (MMS LCU and MPCM) denotedby KAM-MMS KAM-LCU and KAM-MPCM respectivelySimilarly for KAI there are also three VM selection policies(MMS LCU andMPCM) denoted by KAI-MMS KAI-LCUand KAI-MPCM respectively In the following section wewill discuss the energy consumption SLATAH PDM SLAviolations and energy efficiency for the algorithms withdifferent parameter

(1) Energy Consumption For the six algorithms (KAM-MMSKAM-LCU KAM-MPCM KAI-MMS KAI-LCU and KAI-MPCM) with different parameter (05 to 30 increasing by05) the energy consumption is shown in Figure 2 whichshows the energy consumption for the six algorithms As forKAM-MMS KAM-LCU and KAM-MPCM KAM-MPCMleads to the least energy consumption KAM-LCU the secondenergy consumption and KAM-MMS the most energy con-sumption The reason is that KAM-MPCM considers bothCPU utilization and memory size when a host is with heavyload Compared with KAM-MMS KAM-LCU leads to lessenergy consumption This can be explained by the fact thatthe processor (CPU) of a host consumes much more energythan its memory Similarly as for KAI-MMS KAI-LCU

KAM-MMSKAM-LCUKAM-MPCM

KAI-MMSKAI-LCUKAI-MPCM

10 15 20 25 3005r

0

50

100

150

200

250

Ener

gy co

nsum

ptio

n (k

Wh)

Figure 2 The energy consumption of the six algorithms

and KAI-MPCM KAI-MPCM leads to the least energyconsumption KAI-LCU the second energy consumptionand KAI-MMS the most energy consumption The reasonis that KAI-MPCM considers both CPU utilization andmemory size when a host is with heavy load Compared with




05 10 15 20 25 3000r

0

1

2

3

4

5

6

SLAT

AH

()

Figure 3 The SLATAH of the six algorithms

KAI-MMS KAI-LCU leads to less energy consumptionThiscan be also explained by the fact that the processor (CPU) ofa host consumes much more energy than its memory

(2) SLATAH The SLATAH is described in Section 33 (see(3)) For the six algorithms (KAM-MMS KAM-LCU KAM-MPCM KAI-MMS KAI-LCU and KAI-MPCM) with dif-ferent parameter (05 to 30 increasing by 05) the SLATAHis shown in Figure 3 which shows the SLATAH for the sixalgorithms In terms of KAM-MMS KAM-LCU and KAM-MPCMKAM-MMS contributes to the least SLATAHKAM-MPCM the second andKAM-LCU themostThe reason is asfollows when a host is with heavy load KAM-MMS selectsa VM with the minimum memory size to migrate leadingto less migration time Therefore KAM-MMS leads to theleast SLATAHcomparedwith KAM-LCU andKAM-MPCMCompared with KAM-MPCM KAM-LCU contributes tomuch more SLATAH This could be explained by the factthat SLATAH mainly depends on memory size but not CPUutilization Similarly as for KAI-MMS KAI-LCU and KAI-MPCM KAI-MMS contributes to the least SLATAH KAI-MPCM the second and KAI-LCU the most The reason isthat KAI-MMS causes the least migration time leading to theleast SLATAH Compared with KAI-MPCM KAI-LCU leadsto much more SLATAH This could also be explained by thefact that SLATAH mainly depends on memory size but notCPU utilization

(3) PDM The PDM is described in Section 33 (see (2))For the six algorithms (KAM-MMS KAM-LCU KAM-MPCMKAI-MMS KAI-LCU andKAI-MPCM)with differ-ent parameter (05 to 30 increasing by 05) the PDM is shownin Figure 4 which illustrates the PDM for the six algorithmsAs for KAM-MMS KAM-LCU and KAM-MPCM the PDMis the sameThis could be explained by the fact that the overallperformance degradation caused by VMs due to migration isthe same Furthermore when parameter 119903 = 05 10 and 15



05 10 15 20 25 3000r

0000

0005

0010

PDM

()

Figure 4 The PDM of the six algorithms



10 15 20 25 3005r

SLA

vio

latio

ns(10minus7)

0

10

20

30

40

50

Figure 5 The SLA violations of the six algorithms

respectively the corresponding PDM of the three algorithms(KAM-MMS KAM-LCU and KAM-MPCM) are 0 As forKAI-MMS KAI-LCU and KAI-MPCM the PDM is also thesameThis could also be explained by the fact that the overallperformance degradation caused by VMs due to migrationis the same At the same time when parameter 119903 = 05the PDM of the three algorithms (KAI-MMS KAI-LCU andKAI-MPCM) are 0

(4) SLA Violations The SLA violations are described inSection 33 (see (4)) For the six algorithms (KAM-MMSKAM-LCU KAM-MPCM KAI-MMS KAI-LCU and KAI-MPCM) with different parameter (05 to 30 increasing by05) the SLA violations are shown in Figure 5 which showsthe SLA violations for the six algorithms From (4) the SLA




0

2000

4000

6000

8000

Ener

gy effi

cien

cy

10 15 20 25 3005r

Figure 6 The energy efficiency of the six algorithms

violations depend on SLATAH (Figure 3) and PDM (Fig-ure 4) As for KAM-MMS KAM-LCU and KAM-MPCMthe PDM is the same Therefore SLA violations depend onSLATAH In terms of KAM-MMS KAM-LCU and KAM-MPCMKAM-MMS contributes to the least SLATAHKAM-MPCMthe second andKAM-LCU themost So KAM-MMScontributes to the least SLA violations KAM-MPCM the sec-ond andKAM-LCU themost Furthermore when parameter119903 = 05 10 and 15 respectively the corresponding PDMof the three algorithms (KAM-MMS KAM-LCU and KAM-MPCM) are 0 Therefore the corresponding SLA violationsof the three algorithms (KAM-MMS KAM-LCU and KAM-MPCM) are 0 By the same way as for KAI-MMS KAI-LCUand KAI-MPCM KAI-MMS contributes to the least SLAviolations KAI-MPCM the second and KAI-LCU the mostAt the same time when parameter 119903 = 05 the PDM of thethree algorithms (KAI-MMS KAI-LCU and KAI-MPCM)are 0 Therefore the SLA violations of the three algorithms(KAI-MMS KAI-LCU and KAI-MPCM) are 0

(5) Energy Efficiency For the six algorithms (KAM-MMSKAM-LCU KAM-MPCM KAI-MMS KAI-LCU and KAI-MPCM) with different parameter (05 to 30 increasingby 05) the energy efficiency (119864) is shown in Figure 6which shows the energy efficiency of the six algorithms Asdiscussed in Section 34 (5) depends on energy consumption(Figure 2) and SLA violation (Figure 5) Compared withKAM-LCUandKAM-MPCM the energy efficiency ofKAM-MMS is the most The reason is that KAM-MMS reduces themigration time of VMs In terms of energy efficiency KAM-MPCM is better than KAM-LCU This can be explained bythe fact that KAM-MPCM considers both CPU utilizationand memory size As (5) is related to energy consumption(Figure 2) and SLA violation (Figure 5) when parameter119903 = 05 10 and 15 respectively the corresponding SLAviolations of the three algorithms (KAM-MMS KAM-LCUand KAM-MPCM) are 0 Therefore the corresponding

Table 3 Comparison of the VMs select policies using paired 119905-tests

Policy 1 Policy 2 Difference 119875 valueMMS (35795) LCU (22283) 15312 119875 value lt 005MMS (35795) MPCM (31099) 4696 119875 value gt 005MPCM (31099) LCU (22283) 8816 119875 value lt 005

energy efficiency of the three algorithms (KAM-MMSKAM-LCU and KAM-MPCM) is 0 Similarly compared with KAI-LCU and KAI-MPCM the energy efficiency of KAI-MMS isthe most the reason is that KAI-MMS reduces the migrationtime of VMs In terms of energy efficiency KAI-MPCM isbetter than KAI-LCU This can be explained by the fact thatKAI-MPCM considers both CPU utilization and memorysize As (5) is related to energy consumption (Figure 2)and SLA violation (Figure 5) when parameter 119903 = 05the corresponding SLA violations of the three algorithms(KAI-MMS KAI-LCU and KAI-MPCM) are 0 Thereforethe corresponding energy efficiency of the three algorithms(KAI-MMS KAI-LCU and KAI-MPCM) is 0

Considering energy efficiency we choose the parameterof six algorithms to maximize the energy efficiency Figure 6illustrates that parameter 119903 = 20 for KAM-MMS is the best(denoted by KAM-MMS-20) parameter 119903 = 30 for KAM-LCU is the best (denoted by KAM-LCU-30) parameter 119903 =30 for KAM-MPCM is the best (denoted by KAM-MPCM-30) parameter 119903 = 10 for KAI-MMS is the best (denoted byKAI-MMS-10) parameter 119903 = 10 for KAI-LCU is the best(denoted by KAI-LCU-10) and parameter 119903 = 10 for KAI-MPCM is the best (denoted by KAI-MPCM-10)

For the two kinds of adaptive three-threshold algorithms(KAM and KAI) there are three VMs select policies whichcould be provided Does a policy that is the best comparedwith other two policies in regard to energy efficiency exist Ifit exists which VM select policy is the best In order to solvethis problem we have made three paired 119905-tests to determinewhich VM policy is the best in regard to energy efficiencyBefore using the three paired 119905-tests according to Ryan-Joinerrsquos normality test we verify the three VM select policies(MMS LCU and MPCM) with different parameter (119903) thatmaximization energy efficiency follows a normal distributionwith119875 value = 02gt 005The paired 119905-tests results are showedin Table 3 which shows that MMS leads to a statisticallysignificantly upper value of energy efficiency with 119875 valuelt 005 compared with LCU and MPCM In other words interms of energy efficiency MMS is the best VM select policycompared with LCU and MPCM Therefore KAM-MMS-20 and KAI-MMS-10 are the best combination in regardto energy efficiency In the following section we use KAM-MMS-20 and KAI-MMS-10 to make comparison with otherenergy-saving algorithms

(6) Comparison with Other Energy-Saving Algorithms In thissection the NPA (Nonpower Aware) DVFS [9] THR-MMT-10 [15] THR-MMT-08 [15] MAD-MMT-25 [15] IQR-MMT-15 [15] and MIMT [16] are chosen to make compar-ison in regard to energy efficiency The related experimentalresults are shown in Table 4


Table 4 The comparison of the energy-saving algorithms

AlgorithmEnergyefficiency(KWh)

Energyconsumption

(times10minus7)

SLAviolations

SLATAH()

PDM()

Number ofVM

migrationsNPA mdash 24108 mdash mdash mdash mdashDVFS mdash 80391 mdash mdash mdash mdashTHR-MMT-10 38 9995 2613 2697 010 19852THR-MMT-08 170 11940 492 499 010 26567MAD-MMT-25 169 11427 518 524 010 25923IQR-MMT-15 166 11708 514 508 010 26420MIMT (0ndash40ndash80) 5420 10853 17 286 001 8418MIMT (5ndash45ndash85) 4949 11226 18 322 001 8687KAM-MMS-20 9231 8333 13 173 001 6808KAI-MMS-10 6393 10428 15 203 001 7519

Table 4 illustrates the energy efficiency energy consump-tion and SLA violations for the energy-saving algorithms Asthe NPA algorithm does not take any energy-saving policyleading to 24108 KWh DVFS algorithm reduces this valueto 80391 KWh by adjusting its CPU frequency dynamicallyBecause NPA and DVFS algorithm have no VMs migrationthe symbol ldquomdashrdquo is used to represent the energy efficiency SLAviolations SLATAH PDM and number of VMsrsquo migrationIn terms of energy efficiency THR-MMT-10 and THR-MMT-08 algorithms are better thanDVFSThe reason is thatTHR-MMT-10 and THR-MMT-08 algorithms set the fixedthreshold to migrate VMs leading to upper energy efficiency

MAD-MMT-25 and IQR-MMT-15 are two kinds ofadaptive threshold algorithm which (MAD-MMT-25 andIQR-MMT-15) are better than THR-MMT-10 in regardto energy efficiency This could be described by the factthat MAD-MMT-25 and IQR-MMT-15 dynamically set twothresholds to improve the energy efficiency But comparedwith THR-MMT-08 the three algorithms (MAD-MMT-25IQR-MMT-15 and THR-MMT-08) are at the same level inregard to energy efficiencyThis could be explained by the factthat the 08 threshold value is a proper value for the THR-MMT algorithm

KAM-MMS-20 and KAI-MMS-10 algorithms are betterthan MIMT (0ndash40ndash80) and MIMT (5ndash45ndash85) Thiscould be explained by the fact that MIMT (0ndash40ndash80) andMIMT (5ndash45ndash85) set the fixed threshold to control themigration of VMs while KAM-MMS-20 and KAI-MMS-10 autoadjust threshold to control the migration of VMsaccording to the workload

Table 4 also shows that KAM-MMS-20 and KAI-MMS-10 algorithms respectively improve the energy efficiencycomparedwith IQR-MMT-15MAD-MMT-25 THR-MMT-08 and THR-MMT-10 and maintain the low SLA viola-tion of 13 times 10

minus7 In addition KAM-MMS-20 and KAI-MMS-10 algorithms respectively generate 2-3 times fewerVMsmigrations than IQR-MMT-15 MAD-MMT-25 THR-MMT-08 and THR-MMT-10 The reason is as follows Onone hand KAM-MMS-20 and KAI-MMS-10 are two kindsof adaptive three-threshold algorithm which dynamically set

three thresholds leading to the data center hosts to be dividedinto four classes (hosts with little load hosts with light loadhosts with moderate load and hosts with heavy load) andmigrate the VMs on heavily loaded and little-loaded hoststo the host with light load while the VMs on lightly loadedand moderately loaded hosts remain unchanged ThereforeKAM-MMS-20 and KAI-MMS-10 algorithm reduce themigration time and improve the migration efficiency com-pared with other energy-saving algorithms On the otherhand KAM-MMS-20 and KAI-MMS-10 respectively selectthe best VMs select policy (MMS) combination with thebest parameter (119903) compared with other energy-saving algo-rithms

6 Conclusions

This paper proposes a novel VMs deployment algorithmbased on the combination of adaptive three-threshold algo-rithm and VMs selection policies This paper shows thatdynamic thresholds are more energy efficient than fixedthreshold The proposed algorithms are expected to beapplied in real-world cloud platforms aiming at reducing theenergy costs for cloud data centers

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

This work was supported by the National Natural Sci-ence Foundation (nos 61572525 and 61272148) the PhDPrograms Foundation of Ministry of Education of China(no 20120162110061) the Fundamental Research Funds forthe Central Universities of Central South University (no2014zzts044) the Hunan Provincial Innovation Foundationfor Postgraduate (no CX2014B066) and China ScholarshipCouncil


References

[1] H Khazaei J Misic V B Misic and S Rashwand ldquoAnalysis ofa pool management scheme for cloud computing centersrdquo IEEETransactions on Parallel and Distributed Systems vol 24 no 5pp 849ndash861 2013

[2] J Carretero and J G Blas ldquoIntroduction to cloud computingplatforms and solutionsrdquo Cluster Computing vol 17 no 4 pp1225ndash1229 2014

[3] L Wang and S U Khan ldquoReview of performance metrics forgreen data centers a taxonomy studyrdquo Journal of Supercomput-ing vol 63 no 3 pp 639ndash656 2013

[4] D Rincon A Agustı-Torra J F Botero et al ldquoA novel collabo-ration paradigm for reducing energy consumption and carbondioxide emissions in data centresrdquo The Computer Journal vol56 no 12 pp 1518ndash1536 2013

[5] L A Barroso and U Holzle ldquoThe case for energy-proportionalcomputingrdquo Computer vol 40 no 12 pp 33ndash37 2007

[6] P Bohrer E N Elnozahy T Keller et al ldquoThe case for powermanagement in web serversrdquo in Power Aware ComputingComputer Science pp 261ndash289 Springer Berlin Germany2002

[7] S K Garg S Versteeg and R Buyya ldquoA framework for rankingof cloud computing servicesrdquo Future Generation ComputerSystems vol 29 no 4 pp 1012ndash1023 2013

[8] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011

[9] H Hanson S W Keckler S Ghiasi K Rajamani F Rawsonand J Rubio ldquoThermal response to DVFS analysis with an IntelPentium Mrdquo in Proceedings of the International Symposium onLow Power Electronics and Design (ISLPED rsquo07) pp 219ndash224August 2007

[10] J Kang and S Ranka ldquoDynamic slack allocation algorithms forenergy minimization on parallel machinesrdquo Journal of Paralleland Distributed Computing vol 70 no 5 pp 417ndash430 2010

[11] R Buyya R Ranjan and R N Calheiros ldquoModeling andsimulation of scalable cloud computing environments and theCloudSim toolkit challenges and opportunitiesrdquo in Proceedingsof the International Conference on High Performance Computingand Simulation (HPCS rsquo09) pp 1ndash11 Leipzig Germany June2009

[12] A Beloglazov and R Buyya ldquoEnergy efficient resource man-agement in virtualized cloud data centersrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 826ndash831 IEEE MelbourneAustralia May 2010

[13] A Beloglazov J Abawajy and R Buyya ldquoEnergy-awareresource allocation heuristics for efficient management of datacenters for Cloud computingrdquo Future Generation ComputerSystems vol 28 no 5 pp 755ndash768 2012

[14] A Beloglazov and R Buyya ldquoAdaptive threshold-basedapproach for energy-efficient consolidation of virtual machinesin cloud data centersrdquo in Proceedings of the ACM 8thInternational Workshop on Middleware for Grids Cloudsand e-Science (MGC rsquo10) pp 1ndash6 Bangalore India December2010

[15] A Beloglazov and R Buyya ldquoOptimal online deterministicalgorithms and adaptive heuristics for energy and performanceefficient dynamic consolidation of virtual machines in cloud

data centersrdquo Concurrency Computation Practice and Experi-ence vol 24 no 13 pp 1397ndash1420 2012

[16] Z Zhou Z-G Hu T Song and J-Y Yu ldquoA novel virtualmachine deployment algorithm with energy efficiency in cloudcomputingrdquo Journal of Central South University vol 22 no 3pp 974ndash983 2015

[17] X FanW-DWeber and L A Barroso ldquoPower provisioning fora warehouse-sized computerrdquo in Proceedings of the 34th AnnualInternational Symposium on Computer Architecture (ISCA rsquo07)pp 13ndash23 ACM June 2007

[18] D Kusic J O Kephart J E Hanson N Kandasamy andG Jiang ldquoPower and performance management of virtualizedcomputing environments via lookahead controlrdquo Cluster Com-puting vol 12 no 1 pp 1ndash15 2009

[19] W Voorsluys J Broberg S Venugopal and R Buyya ldquoCostof virtual machine live migration in clouds a performanceevaluationrdquo inCloud Computing First International ConferenceCloudCom 2009 Beijing China December 1ndash4 2009 Proceed-ings vol 5931 of Lecture Notes in Computer Science pp 254ndash265Springer Berlin Germany 2009

[20] R N Calheiros R Ranjan A Beloglazov C A F De Rose andR Buyya ldquoCloudSim a toolkit for modeling and simulationof cloud computing environments and evaluation of resourceprovisioning algorithmsrdquo Software Practice and Experience vol41 no 1 pp 23ndash50 2011

[21] BWickremasinghe R N Calheiros and R Buyya ldquoCloudAna-lyst a cloudsim-based visual modeller for analysing cloud com-puting environments and applicationsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 446ndash452 IEEEPerth Australia April 2010

[22] K S Park and V S Pai ldquoCoMon a mostly-scalable monitoringsystem for PlanetLabrdquoACM SIGOPS Operating Systems Reviewvol 40 no 1 pp 65ndash74 2006

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Distributed Sensor Networks


Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014


ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014


Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia


Biomedical Imaging


ArtificialNeural Systems

Advances in


RoboticsJournal of



Computational Intelligence and Neuroscience

Industrial EngineeringJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in



into four classes according to their load ATEA mi-grates VMs on heavily loaded or little-loaded hosts tolightly loaded hosts while the VMs on lightly loadedand moderately loaded hosts remain unchanged

(ii) Presenting two kinds of adaptive three-thresholdalgorithm to determine the three thresholds

(iii) Putting forward three kinds of VMs selection policiesand making three paired 119905-tests

(iv) Evaluating the proposed algorithms by extensive sim-ulation using the CloudSim toolkit

The rest of this paper is organized as follows In Section 2 therelatedwork is discussed Section 3 presents the powermodelcost of VM migration SLA violation metrics and energyefficiency metrics Section 4 proposes ATEA two kinds ofadaptive three-threshold algorithm VM selection policy andVM deployment algorithm Experiments and performanceevaluation are presented in Section 5 Section 6 concludes thepaper

2 Related Work

At present there are various studies focusing on energy effi-cient resource management in cloud data centers Constraintenergy consumption algorithm [8ndash10] and energy efficiencyalgorithm [11ndash15] are twomain types of algorithms for solvingthe problem of high energy consumption in data centersThemain idea of the constraint energy consumption algorithm isto reduce the energy consumption in data centers but thistype of algorithm focuses a little on (does not consider) theSLA violation For example Lee and Zomaya [8] proposedtwo heuristic algorithms (ECS ECS + idle) to decrease theenergy consumption but the two algorithms are easy to fallinto local optimum and do not consider the SLA violationHanson et al [9] presented Dynamic Voltage and FrequencyScaling (DVFS) policy to save power in data centers Whenthe task number is large DVFS policy raises the voltage of theprocessor in order to deal with the task in time when the tasknumber is small DVFS policy decreases the voltage of pro-cessor for the purpose of saving power Kang and Ranka [10]put forward an energy-saving algorithm and they supposedthat overestimated or underestimated execution time of tasksis bad for energy-saving algorithm For overestimation theextra available time should be assigned to other tasks in orderto reduce energy consumption Similarly this energy-savingalgorithm does not consider the SLA violation Thereforethe constraint energy consumption algorithm does not meetthe requirement of users because of focusing a little on (notconsidering) the SLA violation

The goal of the energy efficiency (energy consumptionand SLA violation) algorithm is to decrease the energyconsumption and SLA violation in data centers For exampleBuyya et al [11] raised a virtual machine (VM) placementalgorithm (called Single Threshold (ST)) based on the com-bination of VM selection policies ST algorithm sets a unifiedvalue for all serversrsquo CPU utilization to make sure all serversare below this value Obviously ST algorithm can save energyconsumption and decrease the SLA violation but the SLA

violation remains at a high level Beloglazov and Buyya [12]proposed an energy efficient resource management systemwhich includes the dispatcher globalmanager localmanagerand VMM (VM Monitor) In order to improve the energyefficiency Beloglazov et al put forward a new VMmigrationalgorithm called Double Threshold (DT) [13] DT sets twothresholds to keep all hostsrsquo CPU utilization between the twothresholds However the energy consumption and SLA vio-lation for DT algorithm need to be further decreased LaterBeloglazov and Buyya [14 15] proposed an adaptive double-threshold VM placement algorithm to improve energy effi-ciency in data centers However the energy consumption indata centers remains at a high level

In our previous study [16] we proposed a three-thresholdenergy-aware algorithm named MIMT to deal with theenergy consumption and SLA violation However the threethresholds for controlling hostrsquos CPU utilization are fixedTherefore MIMT is not suitable for varying workloadTherefore it is necessary to put forward a novel VM place-ment algorithm to deal with energy consumption and SLAviolation in cloud data centers

3 The Power Model Cost ofVM Migration SLA Violation Metricsand Energy Efficiency Metrics

31 The Power Model Energy consumption by servers indata centers is connected with its CPU memory disk andbandwidth Recent studies [17 18] have illustrated that theenergy consumption by servers has a linear relationshipwith its CPU utilization even DVFS policy is appliedHowever with the decrease of hardware price multicoreCPUs and memory with large-capacity are widely equippedin servers leading to the traditional linear model not beingable to accurately describe energy consumption of serversIn order to deal with this problem we use the real dataof energy consumption offered by SPECpower benchmark(httpwwwspecorgpower ssj2008)

We have chosen two servers equipped with dual-coreCPUsThemain configuration of the two servers is as followsone is HP ProLiant G4 equipped with 186GHz (dual-core)4GB RAM the other is HP ProLiant G5 equipped with266GHz (dual-core) 4GB RAM The energy consumptionfor the two servers at different load levels is listed in Table 1[15]

32 Cost of VM Migration Proper VM migration betweenservers can reduce energy consumption and SLA violation indata centers However excessive VM migration could bringnegative impact on performance of applicationwhich runs onthe VMs Voorsluys et al [19] investigated the cost problemof VM migration and the performance degradation causedby VM can be described in

119862 = 119896 sdot int

1199050+119879119898119895

1199050

119906119895(119905) 119889119905

119879119898119895

=

119872119895

119861119895

(1)



Server 0 10 20 30 40 50 60 70 80 90 100HP G4 86 894 926 96 995 102 106 108 112 114 117HP G5 937 97 101 105 110 116 121 125 129 133 135





119895




PDM =1

119872

119872

sum

119895=1

119862119889119895

119862119903119895

(2)






SLATAH =1

119873

119873

sum

119894=1

119879119904119894

119879119886119894

(3)









119864 =1






119897 119879119898 and 119879

ℎ(0 le 119879

119897lt

119879119898

lt 119879ℎ



119897and 119879



119898and 119879





Begin

No Yes

No

No

Yes

Yes


End


u gt Tl

u gt Tm

u gt Th

All VMs on




another host with

with light load

light load



119897 119879119898 and




119897 119879119898 and 119879










1 1198812 1198813



1 1198812 1198813 119881

119899) into 119898 groups (119866

1 1198662 119866

119898)


119896= (119881119895119896minus1+1

119881119895119896minus1+2

119881119895119896) for


5= 119899Then KAM


119866119860119896

=

(119881119895119896minus1+1

+ 119881119895119896minus1+2

+ sdot sdot sdot + 119881119895119896)

(119895119896minus 119895119896minus1

) (6)





1198601 1198661198602 119866



MAD = median119860119901

(10038161003816100381610038161003816119866119860119901


(119866119860119902)10038161003816100381610038161003816) (7)



119897 119879119898 and 119879

ℎ) in ATEA


119879119897= 05 (1 minus 119903 timesMAD) (8)

119879119898= 09 (1 minus 119903 timesMAD) (9)

119879ℎ= 1 minus 119903 timesMAD (10)







119897 119879119898 and 119879

ℎ


1 1198812 1198813 119881

119899(119881119894is a


119899) into119898 groups (119866

1 1198662 119866



119896= (119881119895119896minus1+1

119881119895119896minus1+2

119881119895119896) for all 1 le 119896 le 5




119866119860119896

=

(119881119895119896minus1+1

+ 119881119895119896minus1+2

+ sdot sdot sdot + 119881119895119896)

(119895119896minus 119895119896minus1

) (11)


1198601 1198661198602 119866


follows

IR = 1198763minus 1198761 (12)




ℎ) in ATEA can be

defined as follows

119879119897= 05 (1 minus 119903 times IR) (13)

119879119898= 09 (1 minus 119903 times IR) (14)

119879ℎ= 1 minus 119903 times IR (15)







119897 119879119898 and 119879

ℎ



















119897

119879119898 and 119879















119897) || (utilization gt 119879

119898)) then









10 15 20 25 3005r

0

50

100

150

200

250

Ener

gy co

nsum

ptio

n (k

Wh)






05 10 15 20 25 3000r

0

1

2

3

4

5

6

SLAT

AH

()







05 10 15 20 25 3000r

0000

0005

0010

PDM

()




10 15 20 25 3005r

SLA

vio

latio

ns(10minus7)

0

10

20

30

40

50







0

2000

4000

6000

8000

Ener

gy effi

cien

cy

10 15 20 25 3005r













Energyconsumption

(times10minus7)

SLAviolations

SLATAH()

PDM()

Number ofVM








6 Conclusions


Competing Interests


Acknowledgments



References































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





Server 0 10 20 30 40 50 60 70 80 90 100HP G4 86 894 926 96 995 102 106 108 112 114 117HP G5 937 97 101 105 110 116 121 125 129 133 135





119895




PDM =1

119872

119872

sum

119895=1

119862119889119895

119862119903119895

(2)






SLATAH =1

119873

119873

sum

119894=1

119879119904119894

119879119886119894

(3)









119864 =1






119897 119879119898 and 119879

ℎ(0 le 119879

119897lt

119879119898

lt 119879ℎ



119897and 119879



119898and 119879





Begin

No Yes

No

No

Yes

Yes


End


u gt Tl

u gt Tm

u gt Th

All VMs on




another host with

with light load

light load



119897 119879119898 and




119897 119879119898 and 119879










1 1198812 1198813



1 1198812 1198813 119881

119899) into 119898 groups (119866

1 1198662 119866

119898)


119896= (119881119895119896minus1+1

119881119895119896minus1+2

119881119895119896) for


5= 119899Then KAM


119866119860119896

=

(119881119895119896minus1+1

+ 119881119895119896minus1+2

+ sdot sdot sdot + 119881119895119896)

(119895119896minus 119895119896minus1

) (6)





1198601 1198661198602 119866



MAD = median119860119901

(10038161003816100381610038161003816119866119860119901


(119866119860119902)10038161003816100381610038161003816) (7)



119897 119879119898 and 119879

ℎ) in ATEA


119879119897= 05 (1 minus 119903 timesMAD) (8)

119879119898= 09 (1 minus 119903 timesMAD) (9)

119879ℎ= 1 minus 119903 timesMAD (10)







119897 119879119898 and 119879

ℎ


1 1198812 1198813 119881

119899(119881119894is a


119899) into119898 groups (119866

1 1198662 119866



119896= (119881119895119896minus1+1

119881119895119896minus1+2

119881119895119896) for all 1 le 119896 le 5




119866119860119896

=

(119881119895119896minus1+1

+ 119881119895119896minus1+2

+ sdot sdot sdot + 119881119895119896)

(119895119896minus 119895119896minus1

) (11)


1198601 1198661198602 119866


follows

IR = 1198763minus 1198761 (12)




ℎ) in ATEA can be

defined as follows

119879119897= 05 (1 minus 119903 times IR) (13)

119879119898= 09 (1 minus 119903 times IR) (14)

119879ℎ= 1 minus 119903 times IR (15)







119897 119879119898 and 119879

ℎ



















119897

119879119898 and 119879















119897) || (utilization gt 119879

119898)) then









10 15 20 25 3005r

0

50

100

150

200

250

Ener

gy co

nsum

ptio

n (k

Wh)






05 10 15 20 25 3000r

0

1

2

3

4

5

6

SLAT

AH

()







05 10 15 20 25 3000r

0000

0005

0010

PDM

()




10 15 20 25 3005r

SLA

vio

latio

ns(10minus7)

0

10

20

30

40

50







0

2000

4000

6000

8000

Ener

gy effi

cien

cy

10 15 20 25 3005r













Energyconsumption

(times10minus7)

SLAviolations

SLATAH()

PDM()

Number ofVM








6 Conclusions


Competing Interests


Acknowledgments



References































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




Begin

No Yes

No

No

Yes

Yes


End


u gt Tl

u gt Tm

u gt Th

All VMs on




another host with

with light load

light load



119897 119879119898 and




119897 119879119898 and 119879










1 1198812 1198813



1 1198812 1198813 119881

119899) into 119898 groups (119866

1 1198662 119866

119898)


119896= (119881119895119896minus1+1

119881119895119896minus1+2

119881119895119896) for


5= 119899Then KAM


119866119860119896

=

(119881119895119896minus1+1

+ 119881119895119896minus1+2

+ sdot sdot sdot + 119881119895119896)

(119895119896minus 119895119896minus1

) (6)





1198601 1198661198602 119866



MAD = median119860119901

(10038161003816100381610038161003816119866119860119901


(119866119860119902)10038161003816100381610038161003816) (7)



119897 119879119898 and 119879

ℎ) in ATEA


119879119897= 05 (1 minus 119903 timesMAD) (8)

119879119898= 09 (1 minus 119903 timesMAD) (9)

119879ℎ= 1 minus 119903 timesMAD (10)







119897 119879119898 and 119879

ℎ


1 1198812 1198813 119881

119899(119881119894is a


119899) into119898 groups (119866

1 1198662 119866



119896= (119881119895119896minus1+1

119881119895119896minus1+2

119881119895119896) for all 1 le 119896 le 5




119866119860119896

=

(119881119895119896minus1+1

+ 119881119895119896minus1+2

+ sdot sdot sdot + 119881119895119896)

(119895119896minus 119895119896minus1

) (11)


1198601 1198661198602 119866


follows

IR = 1198763minus 1198761 (12)




ℎ) in ATEA can be

defined as follows

119879119897= 05 (1 minus 119903 times IR) (13)

119879119898= 09 (1 minus 119903 times IR) (14)

119879ℎ= 1 minus 119903 times IR (15)







119897 119879119898 and 119879

ℎ



















119897

119879119898 and 119879















119897) || (utilization gt 119879

119898)) then









10 15 20 25 3005r

0

50

100

150

200

250

Ener

gy co

nsum

ptio

n (k

Wh)






05 10 15 20 25 3000r

0

1

2

3

4

5

6

SLAT

AH

()







05 10 15 20 25 3000r

0000

0005

0010

PDM

()




10 15 20 25 3005r

SLA

vio

latio

ns(10minus7)

0

10

20

30

40

50







0

2000

4000

6000

8000

Ener

gy effi

cien

cy

10 15 20 25 3005r













Energyconsumption

(times10minus7)

SLAviolations

SLATAH()

PDM()

Number ofVM








6 Conclusions


Competing Interests


Acknowledgments



References































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in







1198601 1198661198602 119866



MAD = median119860119901

(10038161003816100381610038161003816119866119860119901


(119866119860119902)10038161003816100381610038161003816) (7)



119897 119879119898 and 119879

ℎ) in ATEA


119879119897= 05 (1 minus 119903 timesMAD) (8)

119879119898= 09 (1 minus 119903 timesMAD) (9)

119879ℎ= 1 minus 119903 timesMAD (10)







119897 119879119898 and 119879

ℎ


1 1198812 1198813 119881

119899(119881119894is a


119899) into119898 groups (119866

1 1198662 119866



119896= (119881119895119896minus1+1

119881119895119896minus1+2

119881119895119896) for all 1 le 119896 le 5




119866119860119896

=

(119881119895119896minus1+1

+ 119881119895119896minus1+2

+ sdot sdot sdot + 119881119895119896)

(119895119896minus 119895119896minus1

) (11)


1198601 1198661198602 119866


follows

IR = 1198763minus 1198761 (12)




ℎ) in ATEA can be

defined as follows

119879119897= 05 (1 minus 119903 times IR) (13)

119879119898= 09 (1 minus 119903 times IR) (14)

119879ℎ= 1 minus 119903 times IR (15)







119897 119879119898 and 119879

ℎ



















119897

119879119898 and 119879















119897) || (utilization gt 119879

119898)) then









10 15 20 25 3005r

0

50

100

150

200

250

Ener

gy co

nsum

ptio

n (k

Wh)






05 10 15 20 25 3000r

0

1

2

3

4

5

6

SLAT

AH

()







05 10 15 20 25 3000r

0000

0005

0010

PDM

()




10 15 20 25 3005r

SLA

vio

latio

ns(10minus7)

0

10

20

30

40

50







0

2000

4000

6000

8000

Ener

gy effi

cien

cy

10 15 20 25 3005r













Energyconsumption

(times10minus7)

SLAviolations

SLATAH()

PDM()

Number ofVM








6 Conclusions


Competing Interests


Acknowledgments



References































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





















119897

119879119898 and 119879















119897) || (utilization gt 119879

119898)) then









10 15 20 25 3005r

0

50

100

150

200

250

Ener

gy co

nsum

ptio

n (k

Wh)






05 10 15 20 25 3000r

0

1

2

3

4

5

6

SLAT

AH

()







05 10 15 20 25 3000r

0000

0005

0010

PDM

()




10 15 20 25 3005r

SLA

vio

latio

ns(10minus7)

0

10

20

30

40

50







0

2000

4000

6000

8000

Ener

gy effi

cien

cy

10 15 20 25 3005r













Energyconsumption

(times10minus7)

SLAviolations

SLATAH()

PDM()

Number ofVM








6 Conclusions


Competing Interests


Acknowledgments



References































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in








119897) || (utilization gt 119879

119898)) then









10 15 20 25 3005r

0

50

100

150

200

250

Ener

gy co

nsum

ptio

n (k

Wh)






05 10 15 20 25 3000r

0

1

2

3

4

5

6

SLAT

AH

()







05 10 15 20 25 3000r

0000

0005

0010

PDM

()




10 15 20 25 3005r

SLA

vio

latio

ns(10minus7)

0

10

20

30

40

50







0

2000

4000

6000

8000

Ener

gy effi

cien

cy

10 15 20 25 3005r













Energyconsumption

(times10minus7)

SLAviolations

SLATAH()

PDM()

Number ofVM








6 Conclusions


Competing Interests


Acknowledgments



References































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






05 10 15 20 25 3000r

0

1

2

3

4

5

6

SLAT

AH

()







05 10 15 20 25 3000r

0000

0005

0010

PDM

()




10 15 20 25 3005r

SLA

vio

latio

ns(10minus7)

0

10

20

30

40

50







0

2000

4000

6000

8000

Ener

gy effi

cien

cy

10 15 20 25 3005r













Energyconsumption

(times10minus7)

SLAviolations

SLATAH()

PDM()

Number ofVM








6 Conclusions


Competing Interests


Acknowledgments



References































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






0

2000

4000

6000

8000

Ener

gy effi

cien

cy

10 15 20 25 3005r













Energyconsumption

(times10minus7)

SLAviolations

SLATAH()

PDM()

Number ofVM








6 Conclusions


Competing Interests


Acknowledgments



References































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






Energyconsumption

(times10minus7)

SLAviolations

SLATAH()

PDM()

Number ofVM








6 Conclusions


Competing Interests


Acknowledgments



References































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




References































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in










Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



Documents

Research Article Virtual Machine Placement Algorithm for ...downloads.hindawi.com/journals/sp/2016/5612039.pdf · ciency means less energy consumption and SLA violation in data centers