Research Article A Location Selection Policy of Live …downloads.hindawi.com/journals/tswj/2013/492615.pdfResearch Article A Location Selection Policy of Live Virtual Machine Migration

Hindawi Publishing CorporationThe Scientific World JournalVolume 2013 Article ID 492615 16 pageshttpdxdoiorg1011552013492615

Research ArticleA Location Selection Policy of Live Virtual Machine Migrationfor Power Saving and Load Balancing

Jia Zhao12 Yan Ding12 Gaochao Xu12 Liang Hu12 Yushuang Dong12 and Xiaodong Fu12

1 College of Computer Science and Technology Jilin University Changchun Jilin 130000 China2 Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education Jilin University ChangchunJilin 130000 China

Correspondence should be addressed to Gaochao Xu xugcjlueducn

Received 26 July 2013 Accepted 13 September 2013

Academic Editors T Chen Q Cheng and J Yang

Copyright copy 2013 Jia Zhao et alThis is an open access article distributed under the Creative Commons Attribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Green cloud data center has become a research hotspot of virtualized cloud computing architecture And load balancing has alsobeen one of the most important goals in cloud data centers Since live virtual machine (VM) migration technology is widely usedand studied in cloud computing we have focused on location selection (migration policy) of live VM migration for power savingand load balancing We propose a novel approach MOGA-LS which is a heuristic and self-adaptive multiobjective optimizationalgorithm based on the improved genetic algorithm (GA) This paper has presented the specific design and implementation ofMOGA-LS such as the design of the genetic operators fitness values and elitismWe have introduced the Pareto dominance theoryand the simulated annealing (SA) idea into MOGA-LS and have presented the specific process to get the final solution and thusthe whole approach achieves a long-term efficient optimization for power saving and load balancing The experimental resultsdemonstrate that MOGA-LS evidently reduces the total incremental power consumption and better protects the performance ofVM migration and achieves the balancing of system load compared with the existing research It makes the result of live VMmigration more high-effective and meaningful

1 Introduction

VM technology [1 2] one of themost important technologiesin cloud computing is not only a way to implementingcloud computing such as infrastructure as a service (IaaS)[3] architecture but also the embody of the cloud computingidea whereas live VMmigration technology which is widelyused for the maintenance management in virtualized cloudcomputing data centers is the representative of the VMtechnologies When a VM needs migrating from source hostto target host for some goal or several goals generally themigration target of a VM is chosen randomly as long as thehost can accommodate it and then one can automaticallyor manually move the VM to a target host It is obviousthat the way to randomly choose a target host for a live VMmigration which some event has aroused and has more thanone available target host to meet the requirements of thatevent is not efficient in all respectsTherefore a high-efficient

location selection policy tomigrate themigrantVMs onto theright fit hosts is necessary

Nowadays power consumption of data centers has a hugeimpact on environments Researchers have been seeking tofind effective solutions to minimize power consumption ofdata centers while keeping the desired quality of service Onthe background of low-carbon world and cloud computingera researchers have already proposed the field of green cloudcomputing based on cloud computing and virtualization aswell as aiming at reducing power consumption in cloudcomputing data centers

There are a large number of VMs and tasks running onthe hosts of cloud data centers Some hosts have a heavy loadwhich has a huge impact on the service performance Andsome hosts have a relatively lighter load which results in a lowutilization of resources Therefore it is important to achieveload balancing in cloud data centers as it has covered manykey respects of cloud computing data centers

2 The Scientific World Journal

In this paper we have focused on the live VM migra-tion policy for power saving and load balancing Generallyspeaking the migrant VM has many available target hostsHowever only one target host is most suitable for the VMin order to achieve the management goals which includeminimizing the total incremental power consumption incloud data center and achieving load balancing as much aspossible in the case that the performance constraint of liveVMmigration is met

Many papers have presented some heuristic algorithmsto find optimal solutions aiming to minimize power con-sumption However most research has only taken powerconsumption into consideration but not considered loadbalancing In other word the current research has focusedon a single objective optimization for minimizing the powerconsumption In this paper we have focused on the powerconsumption and load balancing It is a multiobjectiveoptimization problem (MOP) which differs from the tradi-tional single objective optimization problem The basic ideaof the existing research is that according to the currentsituation and history of a cloud data center the controllerhas searched for a best migration policy by optimizing theobjective function of the proposed algorithm In the MOPsthe situation is different and complex It needs to optimizemultiple objectives at the same moment Generally theseobjectives influence each other and are even contradictoryTo achieve a multiobjective optimization the problem ofhow to evaluate and compare two solutions for the multiplefunctions as well as the problem of how to give each solutiona reasonable fitness value towards multiple functions must beaddressed first

It is well known that it is the Pareto dominance principlethat is used for MOPs to address the problems Accordingto the Pareto theory we know that the optimal solution ofan MOP is not a single solution but a set of Pareto optimalsolutionsThe common and classical heuristic algorithms areused to convert the MOP to a single-objective optimizationproblem by emphasizing one particular Pareto optimal solu-tion at a time When such an algorithm will be used to searchout multiple solutions it has to be executed many timeshopefully finding a different solution at each simulation runHowever the genetic algorithm (GA) as the representativeof evolution algorithms (EAs) is suitable for dealing with theMOPs since it has the ability to find multiple Pareto optimalsolutions in one single simulation run As EAs work witha population of solutions a simple EA can be extended tomaintain a diverse set of solutions With an emphasis formoving towards the true Pareto optimal region an EA can beused to find multiple Pareto optimal solutions in one singlesimulation run [4 5] In this paper we have proposed anovel heuristic GA-based location selection policy MOGA-LS of VM for power saving and load balancing whichutilizes the Pareto nondominated sorting individual densityestimation and mathematical statistics of Pareto optimalsolutions to achieve a long-term excellent power saving andload balancing optimization The final experimental resultshave demonstrated that the MOGA-LS approach not onlyreduces the total incremental power consumption but alsoachieves a better load balancing while relatively maximizing

the performance of live VMmigrationThat is it has reducedthe failure events of live VM migration and contributed toachieving better green cloud data centers with load balancing

The rest of the paper is organized as follows In Section 2we present the related work and the reasonable prerequisitesare shown clearly In Section 3 the main idea design prob-lem formulation solution representation and implementa-tion of MOGA-LS are introduced in details In Section 4 theexperimental results and analysis on CloudSim platform aregiven Finally in Section 5 we summarize the full paper andfuture work is put forward

2 Related Work

As far as we know the proposed problem which refers tofinding a fit target host for a live VM migration according tothe standard of minimizing the increment power consump-tion or load balancing has not been widely researched inthe related fields of live VM migration policy let alone bothHowever most researchers have focused on some problemswhich are similar to the proposed problem in this paperThusthe related work of the kind of problems relating to live VMmigration towards power saving and load balancing will bediscussed briefly in this section

Rusu et al in [6] have presented a cluster-wide QoS-aware technique that dynamically reconfigures the cluster toreduce energy consumption during periods of reduced loadThe proposed system consists of two important componentsnamely front end manager and a local manager While thefront end manager finds the servers which should be turnedon or off in terms of a given system load the local managerwill utilize dynamic voltage and frequency scaling (DVFS)technique to conserve energy The main shortage of theapproach is the onoff policy It relies on the table of valuesand needs offline computing However the system does notmake use of server consolidation throughVMmigration andthus its onoff policy may not be much effective

Srikantaiah et al [7] have investigated the problem ofdynamic consolidation of applications serving small statelessrequests in data centers tominimize the energy consumptionThey modeled the problem as a multidimensional bin pack-ing problem However the proposed model doesnrsquot describethe degradation of performance due to the consolidationBesides the energy drawn may rely on a particular set ofapplications combined on a computer node A heuristic forthe defined bin packing problem is proposed by the authorsThe heuristic is based upon the idea of minimizing thesum of the current allocationsrsquo Euclidean distances to theoptimal point at each server The application workload willbe allocated to a server using the proposed heuristic sincea request to execute a new application is received Withoutthe sufficient capacity of all active servers the system willswitch on a new server while reallocating all the applicationsusing the same heuristic in an arbitrary order The proposedapproach is fit for heterogeneous environments however ithas several shortcomings First the approach assumes thatall applicationsrsquo resource requirements are known in advanceand constant Second performance and energy overhead

The Scientific World Journal 3

which the authors do not take account into is causedby migration of state-full applications between nodes Thefrequent switching servers onoff also generates significantcosts which are not negligible for a real-world system

Verma et al [8] have contributed energy and migrationcost-aware application placement by exploiting the energymanagement capabilities of virtualization The authors havedesigned a new application (virtual machines) placementarchitecture called pMapper It consists of three major partsnamely a performance manager to dynamically resize theVM an energy manager for CPU throttling and a migrationmanager to identify the target host for migration using aknowledge baseThey have expounded that for energy-awarescheduling approaches estimates of energy values are notrequired and only if the scheduling algorithm has abilities infinding out which server minimizes the incremental increasein total energy owing to the new VM being placed it canplace the given VM to an appropriate host In pMappertwo algorithms are implemented One is First Fit Decreasing(FFD) by which more energy-efficient servers are utilizedfirst without balancing the load The other is incrementalFirst Fit Decreasing (iFFD) which considers the fixed targetutilization of each server and achieves server consolidationby live VM migration The proposed pMapper architectureminimizes energy and migration costs while ensuring theperformance Our approach is based on a heuristic approachwhich exploits the concept of minimizing total increase inthe incremental energy due to the new VM migrationsThe proposed architecture is simple and does not need anyknowledge base to achieve significant reduction in the energyconsumption

Li et al in [9] have proposed an approach namedEnaCloud which enables application of live placementdynamically with consideration of energy efficiency in acloud platform In EnaCloud they use a virtual machineto encapsulate the application which supports applicationsrsquoscheduling and live migration to minimize the number ofrunning machines so as to save energy In particular theapplication placement is abstracted as a bin packing problemand an energy-aware heuristic algorithm is proposed to get anappropriate solution In addition an overprovision approachis presented to deal with the varying resource demands ofapplications However the overprovision approach has riskin optimizing this problem It may cause more cost in orderto reduce cost with a certain probability

Jeyarani et al [10] have proposed self-adaptive particleswarm optimization (SAPSO) for efficient virtual machineprovisioning in cloud aimed at that when mapping a set ofVM instances onto a set of servers from a dynamic resourcepool the total incremental energy drawn upon the mappingis minimal and does not compromise the performance objec-tivesThe advantage of the proposed solution is obvious It hasfocused on not only improving the performance of workloadfacilitating the cloud consumers but also developing theenergy efficient data center management to facilitate cloudproviders However the approach still may be inefficient andmay cause some additional events and costs from a long-term perspective as it does not take the future workload intoaccount Our proposed algorithm MOGA-LS is a heuristic

approach which is based on PSO one of swarm intelligencealgorithms and introduces the simulated annealing (SA) ideainto it

Jing and She [11] have proposed a novel model for thisproblem of live VM migration for load balancing based onfuzzy TOPSIS to detect the hotspots and balance load Theproposedmodel is tomigrate VMs between hosts using fuzzyTOPSIS theory to make decision over the whole of activehosts of the data center and detect the hotspotsThe proposedmodel can be a suitable tool to rank hosts However thekind of load balancing policies based on sorting hosts is notheuristic enough and has a relatively complex computationIt is not efficient for achieving load balancing

Yang et al in [12] have proposed a load balancing algo-rithm based on live VM migration The proposed algorithmincludes two major policies trigger strategy based on theload prediction and selection strategy of the destination nodebased on the multiple criteria decision Yet the kind of loadbalancing policies based on load predicting is not accurateenough and needs to maintain additional historical data asresults in the unnecessary system load

In this paperMOGA-LS has a prerequisiteWe know thatMOGA-LS is to find the target host of each VM from all 119898hosts for the 119899migrantVMsWe assume that eachVMrsquos targethost found by MOGA-LS will not be the host which the VMis moved out from The algorithm provides service to liveVM migration aiming at green cloud data center with loadbalancing Thus the fact that the VM should be moved outfor some reason is the premise of our approach Objectivelythe prerequisite is justified from a certain perspective Since aVM needs to be migrated from its source host its candidatehosts will not include its source host Otherwise it doesnot need a migration event It can be seen that for all themigrant VMs the hosts each of which is the source host ofsome migrant VM will not be the target hosts Therefore theproposed prerequisite is reasonable and does not impact theperformance and efficiency of our approach

3 Proposed MOGA-LS Approach

31 Main Idea The proposed MOGA-LS approach is aheuristic live VM migration policy which is used for thetarget location selection of live VM migration It is a mul-tiobjective optimization algorithm which is based on theimproved GA that utilizes Pareto theory to achieve theselection crossover and mutation operators of GA towardsmultiple objectives and thus to find out the set of Paretooptimal solutions of the proposed multiobjective problem bythe evolution of the population After obtaining the set ofPareto optimal solutions in order to get the final migrationpolicy we do not randomly select a solution but utilize themathematical statistics theory to obtain the final solution oflocation selection of live VMmigration by using a probabilitywheel for each VM

Specifically we have employed a novel policy which isused to generate the initial population and defined a Paretoconstraint dominance relation towards the proposed con-strained problem to compare two solutions in the improved


GA-based approach The nondominated sorting policy anddensity estimation policy used for the population has beenpresented to make each individual of the population obtainfitness values in each generation We have employed thetournament selection operator for the selection operator ofGA For the crossover and mutation operators we havedesigned the arithmetic crossover operator and the dynamicnonuniformmutation operator Besides the (120583 +120582) selectionpolicy which makes elitism that helps in achieving betterconvergence be introduced into the proposed MOGA-LSapproach is employed to generate the next population

32 Design of the MOGA-LS Approach The MOGA-LSapproach is an algorithm based onmultiobjective GA achiev-ing a live VMmigration policy for minimizing the incremen-tal power consumption caused by migrating these migrantVMs onto their target hosts and making the load of clouddata center balanced after migrating under the constraintof maximizing the performance that the number of successof live VM migration events is maximized To achieve amultiobjective GA we have introduced the concept of Paretooptimal solutions into GA and designed a constrained Paretodominance method to evaluate the individuals of populationand assign fitness values to them as well as designingGArsquos genetic operators including selection crossover andmutation operators as well as the policy of generating the newpopulation

321 Problem Formulation We now formulate the proposedproblem of migrating 119899 VMs onto 119898 hosts Its solution canbe represented by an 119899 dimension of vector each elementof which denotes the target host of the migrant VM whichits location represents We assume that there are 119898 availablehosts in the resource pool of a cloud data center and the hostsare heterogeneous and dynamic while using space sharedallocation policy The hosts change their state dynamicallyaccording to the load Similarly each VM is associated withthe required computing resource It is considered that theVMs are independent of each other and are equal in priority

The problem can be stated as follows Find a VM-hostmapping set LS of location selections such that the totalincremental power consumption caused by the migratedVMs onto hosts is minimized and after migrating theload is balanced as much as possible while maximizingthe performance by fulfilling the resource requirements ofmaximum number of VMs We define a quintuple 119878 =

PHVMPCRL LS for our problem scenario PH is aset of 119898 available physical hosts denoted by PH (119898 119905) =

PH1PH2PH3 PH119898 available at migrating start time119905 VM is a set of 119899 migrant VMs denoted by VM (119899 119905 Δ119905) =

VM1VM2VM3 VM119899 accumulated within a timewindow Δ119905 PC (119898 119905) = 1198751 1198752 119875119898 is the powerconsumption by the 119898 physical hosts in the resourcepool RL (119898 119905) = 119877CPU1 119877MEM1 119877CPU2 119877MEM2119877CPU3 119877MEM3 119877CPU119898119877MEM119898 represents the resid-ual CPU andmemory resource of each host at migrating starttime 119905 The problem is multimodal generally having morethan one location selection which meets the performance

constraint of the VM requests Our goal is to find a locationselectionwhichmeets the performance constraint whilemin-imizing the incremental power consumption andminimizingthe standard deviation of the residual load rates of all availablehosts to achieve load balancing That is the goal is to obtainall Pareto optimal solutions It can be seen as a constrainedmultiobjective Pareto optimization problem

To fulfill the performance constraint a metric denotedby 120578119903representing resource fulfillment requirement 120578

119903rep-

resents the number of success of live VM migration eventsand should be equal to 119899 as a constraint function ideally isdefined as follows

120578119903=

119899

sum

119894=1

119898

sum

119895=1

119905119895

119894= 119899 119894 isin 1 2 3 119899 119895 isin 1 2 3 119898

119903 isin 1 2 3 119904

(1)

where 120591119895119894denotes the location selection of 119894th VM on the 119895th

host and is defined as follows 119904 denotes the total number oflocation selections

120591119895

119894=

1

if VM119894has been allocated to

Host119895and rcrVM

119894le acrHost

119895

119894 isin 1 2 3 119899

0


Host119895and rcrVM

119894le acrHost

119895

119895 isin 1 2 3 119898

Invalidif VM

119894has been allocated to Host

119895

(2)

In (2) rcrVM119894denotes the minimum computing resource

requirements of 119894th VM and acrHost119895denotes the available

computing resource of 119895th hostThe first objective function is based on power consump-

tion The migration of successive VMs is represented as LS[119898 119899 119903 119905 + 119905

0(119896)] where 119903 represents any one of 119904 location

selections and 119896 is an integer increasing with successivemigrating representing a stage We can understand that if 119896is 3 the approach will migrate the third VM to the host whichis denoted at the third location of the 119903 vector The 119904 locationselections at stage 119896 are represented as LS [119898 119899 119903 119905+119905

0(119896)] 119903 isin

1 2 3 119904 and their corresponding power consumption isrepresented as 120585119903

119896 Its meaning is that according to 119903 location

vector after the system has migrated the 119896th VM to its targethost the total power consumption by the cloud data center is120585119903

119896 ideally The meaning of 120585119903

119896minus1can be imaged In this paper

these parameters of 120585119903119896can be obtained by using simulation

platform in the experiment Now the incremental powerconsumption due to migrating LS [119898 119899 119903 119905 + 119905

0(119896)] with

respect to previous migration stage LS [119898 119899 119903 119905 + 1199050(119896 minus 1)]

is defined by

Δ119875 = (120585119903

119896) minus (120585

119903

119896minus1) 119903 isin 1 2 3 119904 (3)


For better power saving the following 120575119875119903 should be mini-mized and it is denoted as follows

120575119875119903

=

119899

sum

119896=1

((120585119903

119896) minus (120585

119903

119896minus1)) 119903 isin 1 2 3 119904 (4)

Therefore the proposed MOGA-LS approach minimizes 120575119875119903for power efficiency Function (4) is the first objective forMOGA-LS

The second objective function is based on load bal-ancing In this paper we have presented the residual loadrate to measure the load situation of each host The cal-culation method of the residual load rate is describedas follows Now assume that the set of available hostsis PH (119898 119905) = PH1PH2PH3 PH119898 after a timewindow Δ119905 Within the Δ119905 the set of the accumu-lated live VM migration requests is represented as VM(119899 119905 Δ119905) = VM1VM2VM3 VM119899 After migratingall the migrant VMs based on a location selection 119903 isin

1 2 3 4 5 6 119904 the residual load 119877119894119903 of the host 119894 is

defined as follows

119877119894119903

= 120572119877119894119903

CPU + 120573119877119894119903

MEM 119894 isin 1 2 3 119898

119903 isin 1 2 3 119904

120572 + 120573 = 1

(5)

where 119877119894119903

CPU and 119877119894119903

MEM represent the residual CPU andmemory resource of host 119894 after migrating according to the

solution vector 119903 So the residual load rate 119864119894119903 of host 119894 isdefined as follows

119864119894119903

=

119877119894119903

119879119894 119894 isin 1 2 3 119898 119903 isin 1 2 3 119904 (6)

where 119879119894 can be represented as follows

119879119894

= 120572119879119894

CPU + 120573119879119894

MEM 119894 isin 1 2 3 119898 (7)

where 119879119894CPU denotes the total CPU resource of host 119894 and119879119894

MEM denotes the total memory resource of host 119894 In thispaper we think that in order to make the load of all 119898physical hosts balanced as much as possible the residualload rate 119864119894119903 of each host should be as similar as possibleafter having migrating all migrant VMs accumulated withina time window Δ119905 Therefore we have utilized the standarddeviation of all hostsrsquo residual load rates to formulate thisproblem The formula of the expectation and the standarddeviation is as follows

119864 (119883) =

sum119873

119894=1119883119894

119873

120590 = radic1

119873

119873

sum

119894=1

(119883119894minus 119864 (119883))

2

(8)

By using the above two formulas the second objectivefunction can be described as follows

120590119903

=radic1

119898

119898

sum

119894=1

(

120572119877119894119903

CPU + 120573119877119894119903

MEM120572119879119894

CPU + 120573119879119894

MEMminus

sum119898

119896=1((120572119877119896119903

CPU + 120573119877119896119903

MEM) (120572119879119896

CPU + 120573119879119896

MEM))

119898

)

2

(9)

So far we have formulated the proposed problem as amultiobjective optimization problem with a constraint Thatis

Min

120575119875119903

=

119899

sum

119896=1

((120585119903

119896) minus (120585

119903

119896minus1))

120590119903

= radic1

119898

119898

sum

119894=1

(

120572119877119894119903

CPU + 120573119877119894119903

MEM120572119879119894

CPU + 120573119879119894

MEMminus

sum119898

119896=1((120572119877119896119903

CPU + 120573119877119896119903

MEM) (120572119879119896

CPU + 120573119879119896

MEM))

119898

)

2

119903 isin 1 2 3 119904

st 120578119903=

119899

sum

119894=1

119898

sum

119895=1

120601119895

119894= 119899 119894 isin 1 2 3 119899 119895 isin 1 2 3 119898 119903 isin 1 2 3 119904

(10)

322 Relevant Concepts of Pareto Optimal Solutions Ina MOP the fitness values cannot be compared betweenmultiple objectives And generally these objectives are inconflict with each other The improvement of an objectiveis often at the expense of impairing the fitness value of theother objective Thus it can be seen that the solution of a

MOP is not only one but a set consisting of many solutionswhich cannot compare with each other These solutions arecalled Pareto solutions set In all Pareto solutions sets ofa MOP the best one is Pareto optimal solutions set Toexplain this problem we have given some definitions in thefollowing


Pareto Dominance That a solution vector 119906 = (1199061 1199062 119906

119899)

dominates the other solution vector V = (V1 V2 V

119899) (119906 ≺

V) refers to that for all 119894 isin 1 2 3 119896 119891119894(119906) ⩽ 119891

119894(V) and

exist119895 isin 1 2 3 119896 st 119891119895(119906) lt 119891

119895(V) In the definition the

MOP is a minimization problem by default

Pareto Optimal Solutions That a candidate solution 119883 =

(1199091 1199092 119909

119896) isin Ω is a Pareto optimal solution refers to that

notexist1198831015840

isin Ω st1198831015840 ≺ 119883

Pareto Optimal Solutions Set 119875 = 119883 isin Ω | notexist1198831015840

isin Ω 1198831015840

≺

119883

Pareto Front PF = 119865(119883) = (1198911(119883) 119891

2(119883) 119891

119896(119883)) | 119883 isin

119875As can be seen from these definitions our goal is to find

Pareto optimal solutions setThat is theMOGA-LS approachshould return the Pareto optimal solutions set finallyWehaveutilized Pareto dominance to compare any two individualsof population As our proposed problem has a constraint wehave redefined the Pareto dominance as a constrained Paretodominance as follows A solution 119894 is said to constrained-dominate a solution 119895 if any of the following conditions istrue

(1) Solution 119894 is feasible and solution119895 is not(2) Solutions 119894 and 119895 are both infeasible but solution 119894 has

a smaller overall constraint violation

In the proposed problem this scenario refers to thatsolution 119895 results in more failure events of live VMmigrationthan solution 119894This means that if neither of the two solutionscan make all live VMmigration events accumulated within atime window Δ119905 successful the solution which results in lessfailure events of live VM migration is better since that themost migrant VMs are migrated successfully is the premiseof the proposed problem After all the MOGA-LS approachis a live VMmigration policy

(3) Solutions 119894 and 119895 are feasible and solution 119894 dominatessolution 119895

In this paper the Pareto dominance mentioned by theproposed MOGA-LS approach refers to the newly definedconstrained Pareto dominance

323 Solution Representation and GA Encoding In orderto design an efficient GA-based algorithm for finding theoptimal vector of the target hosts of all migrant VMs ina time window Δ119905 the primary problem is the solutionrepresentation as it represents a direct relationship betweenthe problem domain and the individuals in GA Duringthe applying of a GA-based algorithm we know that anindividual is denoted as a solution of the specific problemHere there are 119899 VMs to be migrated into 119898 physical hostsSo the proposed problem is an 119899 dimensional problemleading to that the individuals are represented as some 119899dimensional vectors Each dimension has a discrete set ofall possible location selections which are denoted as integersand limited to 119898 A solution vector called an individual or

a chromosome is denoted by 119909119896119894= (119909119896

1198941 119909119896

1198942 119909

119896

119894119895 119909

119896

119894119863)

where the 119909119896119894119895value represents the coded value (the no of

target host of the 119895th migrant VM) of the 119895th gene (the119895th migrant VM) of 119894th individual (the 119894th possible solutionvector) in 119896th generation

324 Design of Genetic Operators in MOGA-LS

(1) Selection Operator In the MOGA-LS approach we haveemployed the tournament selection operator Our mainidea is that the algorithm randomly chooses two groups ofindividuals from the population Each group consists of 119896individuals From the efficiency and diversity of the MOGA-LS approach to consider the tournament scale 119896 is set to 2in this paper That is the algorithm randomly chooses twogroups each of which includes two individuals from theoriginal population The two winning individuals of the twogroups are obtained by the comparison within groups In thenext step the two individuals will be used for the crossoveroperator

(2) Crossover Operator As mentioned above the GA encod-ing has employed the positive integer coding and the integeris limited to 119898 We have not utilized the widely usedSimulated Binary Crossover (SBX) crossover method that fora random given crossover point the two parent individualsexchange the sections located on both sides of the crossoverpoint However we have designed the nonuniform arithmeticcrossover operator and introduced it into our approach inorder to improve the global search ability and better keep thediversity of population Let119883119905

119894and119883119905

119895 respectively represent

the real encoding values of the crossover points of the twoparent individuals 119894 and 119895 in the 119905th generation After thecrossover the corresponding gene encoding values 119883119905+1

119894and

119883119905+1

119895of the two individuals are defined as follows

119883119905+1

119894= 120572119883119905

119894+ (1 minus 120572)119883

119905

119895

119883119905+1

119895= (1 minus 120572)119883

119905

119894+ 120572119883119905

119895

(11)

where 120572 is a parameter and is not a constant It is related tothe evolution generation number The specific definition of 120572in this paper will be described in the following

(3) Mutation Operator MOGA-LS is a heuristic approachbased onGA As a heuristic optimization algorithm it shouldhave better global search ability in the early iterations andit should have better local search and convergence ability inthe later iterations Therefore we have utilized the dynamicnonuniform mutation operator to make the scope of thegene mutation change with the increase of the generationnumber and thus to improve the search and convergenceability of MOGA-LS Now we assume that an individual (achromosome) 119883 = (119883

1 1198832 119883

119896 119883

119899) is mutated to a

new individual 1198831015840 = (1198831 1198832 119883

1015840

119896 119883

119899) If the gene

encoding value 119883119896of the mutation point 119896 ranges within


[119880119896

min 119880119896

max] the new gene1198831015840119896is determined by the following

formula

1198831015840

119896=

119883119896+ (119880119896

max minus 119883119896) (1 minus 120574(1minus(119905119879))

120576

119896)

if 119903 and (0 1) isin [0 05)

119883119896minus (119883119896minus 119880119896

min) (1 minus 120574(1minus(119905119879))

120576

119896)

if 119903 and (0 1) isin [05 1]

(12)

where 120574119896is a random number distributed uniformly between

0 and 1 119879 is the maximum number of iterations of MOGA-LS 119905 is the current number of iterations of MOGA-LS 120576 isa system parameter which is used to determine dependencydegree on the number of iterations

It is obvious that Δ119905 = (119880119896max minus 119883119896)(1 minus 120574(1minus(119905119879))

120576

119896) ranges

within [0 (119880119896max minus 119883119896)) To begin with 119905 is smaller and thusΔ119905 is larger making the gene value have an obvious mutationWith the increase of the number 119905 of iterations the value ofΔ119905 becomes gradually smaller The change of the gene valuebecomes smallerThis feature makes the operator have abilityin searching the whole space evenly in the early iterations(when 119905 is small) and precisely searching several partial areasin the later iterations To say further the mutation operatormakes the MOGA-LS approach have better global searchability in an early phase and have better local search abilityand convergence in a later phase as it is a dynamic andself-adaptive mutation operator which can be adjusted bymodifying the parameters such as 120574

119896and 120576

325 Design of Fitness Values in MOGA-LS

(1) The Fitness Value 119903119901 The design of the fitness values is the

core of GA And especially for multiobjective GA it is moreimportant since this kind of problems has multiple objectivefunctions In this paper the Pareto dominance is utilized toachieve the Pareto ranking approaches and thus to obtain thefitness values of each individual in each generation

As mentioned above the MOPs are the Pareto optimiza-tion problems which have employed the Pareto dominanceto compare and evaluate the individuals The populationis ranked according to the Pareto dominance rule andthen each solution is assigned a fitness value based on itsrank in the population not any one of its actual objectivefunction values Note that herein a lower rank correspondsto a better solution The rank of each individual refers toits nondominated rank and is called 119903

119901 Each individual

has a parameter 119899119901 whichis the number of individuals that

dominate the individual 119901 Each individual can be comparedwith every other individual in the population to find if it isdominated

The rank 119903119901of each of the individuals whose 119899

119901values

are 0 is set as 1 At this stage all individuals in the first non-dominated front are found In order to find the individualsin the next nondominated front the solutions of the firstfront are discounted temporarily and the above procedureis repeated The rank 119903

119901of each of the individuals whose 119899

119901

values are 0 is set as 2 and so forth until all the individualsof the population are ranked in this generation After thenondominated sorting the population is divided into several

Cuboid

f2

f1

i minus 1

Dp

i

i + 1

Figure 1 Crowding-distance calculation of the density estimation

ranks and each individual has a rank 119903119901 In other words each

of these ranks is a set consisting of several individuals

(2)The Fitness Value119863119901The individuals between these ranks

are comparable In the same rank the individuals cannot becompared with each other and are equal for being selectedFurthermore we know that maintaining a diverse populationis an important consideration in multiobjective GA to obtainsolutions uniformly distributed over the Pareto optimalsolutions set Without taking preventive measures the popu-lation tends to form relatively few clusters in multiobjectiveGA This phenomenon which will prevent the MOGA-LSapproach convergent to Pareto optimal solutions set is calledgenetic drift in genetics To further address the two problemsof the fitness values and the diversity of a population wehave designed and employed the density estimation approachaiming to obtain a uniform spread of solutions along thePareto front Its main idea is that in the same rank theindividuals in which there are fewer individuals around arebetter than others It should be selected more potentially togenerate the next generation in order to prevent genetic driftand keep the diversity of the population and thus to optimizethe MOGA-LS approach

The process of the density estimation approach is asfollows To get an estimate of the density of solutionssurrounding a particular solution in the population we needto calculate the distance of two points on either side of thispoint along each of the objectives This quantity 119863

119901serves

as an estimate of the diagonal of the cuboid formed by usingthe nearest neighbors as the vertices (call this the crowdingdistance) In this paper we have defined the crowdingdistance of the 119894th solution in its front as the diagonal lengthof the cuboid as shown in Figure 1 The computation of thecrowding distance needs to sort the population according toeach objective function value in ascending order Thereafterfor each objective function the boundary solutions (solutionswith smallest and largest function values) are assigned aninfinite distance value All other intermediate solutions areassigned a distance value equal to the corresponding diagonallength If a MOP has three objective functions the crowed


distance is the diagonal length of a cube If a MOP hastwo objective functions the crowed distance is the diagonallength of a rectangle and so forth Each objective function isnormalized before calculating the crowding distance

(3) The Fitness Value 119877119901 Now each individual in the pop-

ulation has two attributes 119903119901and 119863

119901in each generation

The Pareto rank 119903119901is the core of the MOGA-LS approach

Its convergence process is the process of converging thenondominated set of each generation to the Pareto optimalsolutions set of the final generation A rank represents a set ofindividuals where each element belongs to the nondominatedrank In other words the individuals in the same rankare equal and have the same chances to be selected forgenerating the next generation However as a matter offact the individuals in the same rank are not equal due totheir different situations of being dominated especially forthe individuals located in the ranks larger than 1 For eachindividual in the same rank their 119903

119901values are naturally

equal However the 119903119901value does not reflect the ranks of the

individuals dominating it It is obvious that if the sum of therank values of the individuals that dominate an individualis larger than that of other individuals in the same rankthe individual is not better for being used to generate thenext generation This is because its gene is not excellentenough and it is not beneficial for keeping the diversity of thepopulation Thus the information included in the 119903

119901value is

not enough We have presented a novel parameter 119877119901 It is

defined as follows

119877119901= 1 +

119905

sum

119895=1

120574119883119895

(13)

where we assume that the individual 119901 is dominated by 119905individuals 11988311198832 119883119905 whose rank values are alreadyknown as 120574

1198831 1205741198832 120574

119883119905 The dominated individuals are

penalized to reduce the population density and redundancyHere we have given an instance The multiobjective GAgenerates 11 individuals Their rank values are illustrated inFigure 2 Suppose we want to minimize two objectives119891

1and

1198912 Both 119860 and 119861 belong to rank 2 However the 119877

119901value of

individual 119860 is 5 and the 119877119901value of individual 119861 is 2 The

individual 119861 is better What is more for the individuals thathave the same ranks or the different ranks119877

119901is more suitable

as a fitness value for comparison since it has included not onlythe information of the nondominated rank of the individualbut also the information of it being dominated for penalizingit Meanwhile it can be seen that the 119877

119901value has also better

kept the diversity of the population From this perspective italso makes up for the shortcoming of the density estimationsince the density estimation approach is limited to the samerank In order to achieve the calculation of119877

119901 each individual

119894 has a parameter 119878119901 which is a set consisting of the 119903

119901values

of the individuals dominating the individual 119894Thus far in our MOGA-LS approach each individual of

the population has 3 attributes 119877119901 119903119901 and 119863

119901as its fitness

values in each generation It can be seen that our design is thatwhen comparing two individuals the algorithm compares

f2

f1

A

B

rP = 1

rP = 2

rP = 3

rP = 4

RP = 2

RP = 5

Figure 2 Individual rank values

first their 119877119901values second their 119903

119901values and last their 119863

119901

values We now define a partial order ≺119877119901120574119901119863119901

as follows

119894≺119877119901120574119901119863119901

119895 if (119894119877119901

lt 119895119877119901

)

or ((119894119877119901

= 119895119877119901

) (119894120574119901

lt 119895120574119901

))

or ((119894119877119901

= 119895119877119901

) (119894120574119901

= 119895120574119901

)

(119894119863119901

gt 119895119863119901

))

(14)

The design not only has utilized the Pareto theory toachieve multiobjective comparison but also kept the diversityof the population avoided genetic drift and made thepopulation more efficiently converged to the Pareto optimalsolutions set with no additional complexity

326 System Architecture of MOGA-LS Figure 3 depicts theproposed architecture for cloud environment It shows theposition of the controller MOGA-LS for migration locationselections and its interaction with other entities [13 14] Ina time window Δ119905 the monitor gets the requests of liveVM migration and updated with the available number ofcomputing resource such as CPUs memory and storage aswell as power consumption At the end of Δ119905 the monitortransfers the information to the controller The controllergenerates the location policy by using the proposed approachand obtained information Then it transfers the policy tothe migration controller which controls and executes livemigration of the VMs The VMs are eventually moved ontotheir target hosts

33 Implementation of MOGA-LS In this section wedescribe the specific process of MOGA-LS Details are asfollows

Initialize the MOGA-LS Algorithm The size of the populationis 119873 The number of the genes of an individual (a chromo-some) is 119899 due to the 119899 migrant VMs as mentioned above


Migration controller

Monitor

MOGA-LS

Resource information

Placement policy

Host m

Host 1

Host 2

Host clustersMigration requests

The controller for migration placement selection

Resource pool

Figure 3 The view of MOGA-LSrsquos architecture

The maximum number of iterations (evolution) is 119894max Themutation rate of the population is119872

119903 The competition scale

of the tournament selection operator is set to 2 The initialpopulation 119875

0is generated by employing the way in which

the initial archive is generated in [15]

Evolve the Population of MOGA-LS Each individual of 1198750

can be compared with every other solution to find if itis dominated and thus to obtain the 119899

119901value of each

individual by using the constrained fitness functions group(10) The algorithm sets the 120574

119901values of the individuals

whose 119899119901values are 0 as 1 At this stage all individuals in

the first nondominated front are found In order to find theindividuals in the next nondominated front the solutionsof the first front are discounted temporarily and the aboveprocedure is repeated The rank 119903

119901of each of the remaining

individuals whose 119899119901values are 0 is set as 2 and so forth

until all the individuals of the population are ranked in thisgeneration

After the nondominated sorting the population isdivided into several ranks and each individual has a rankvalue 119903

119901 Thereafter each individual of 119875

0obtains its 119877

119901value

by using (13) and its 119878119901set The algorithm first sorts the

population 1198750according to the first objective function (4)

and the second objective function (9) in ascending orderrespectivelyTheboundary solutionswith smallest and largestfunction values are assigned an infinite 119863

119901value All other

intermediate solutions are assigned a distance value 119863119901as

follows

119894119863119901

=radic(

(119894 + 1)sdot1198911minus (119894 minus 1) sdot 119891

1

119891max1

minus 119891min1

)

2

+(


2

119891max2

minus119891min2

)

2

(15)

Now each individual of the initial population 1198750has

the fitness values 119877119901 120574119901 and 119863

119901 The selection operator of

MOGA-LS begins to perform the tournament selection of 1198750

Two groups of individuals are chosen randomly Every groupconsists of two individuals In each group the individual withbetter fitness values is picked out according to the proposedpartial order 119903 ≺

119877119901120574119901119863119901

The crossover operator of MOGA-LS begins to perform the arithmetic crossover operation forthe two winning individuals of the two groups according to(11) and thus to obtain two new individuals The informationof the Pareto nondominated rank values of the individualsof the population is introduced into the arithmetic crossoveroperator used by us In the arithmetic crossover formula (11)the parameter 120572 is defined as follows

120572 =

119895120574119901

119894120574119901

+ 119895120574119901

(16)

The crossover operator coefficient 120572 is associated with thenondominated rank of each individual in the population Inthe early iterations ofMOGA-LS the 120572 value will have a greatchange However with the evolution of the population theindividuals of the population will tend to the same Paretofront and the 120572 value will tend to the constant 05

Repeat the selection and crossover operations until gen-erating a population of size 119873 The mutation operator ofMOGA-LS performs the dynamic nonuniform mutation forthe generated population according to the mutation rate and(12) to eventually obtain the offspring 119876

0of 1198750 At this time

the MOGA-LS algorithm forms a combined population 1198770=

1198750cup 1198760 The size of the population 119877

0is 2119873 Then the

population 1198770is sorted according to nondomination rank

The elitism is ensured as all previous and current populationmembers are included in 119877

0

Now the individuals belonging to the best nondominatedset 1198781are the best solutions in the combined population and


K-means clusteringNondominated sorting

Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1

(a)

Nondominated sorting Crowded distancesorting

Rejected

Crowded distancesorting

Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1120574p = 3

S3

(b)

Figure 4 The procedure of generating the new population in MOGA-LS

must be emphasized more than any other solution in thecombined population If the size of 119878

1is larger than 119873 we

utilized the 119896-means approach to cluster these individualsof 1198781for grouping the solutions into 119873 clusters The 119873

clustering centers of the 119873 clusters are chosen to form thenext population 119875

1 If the size of 119878

1is smaller than 119873

we definitely choose all members of the set for the newpopulation 119875

1 The remaining members of the population

1198751are chosen from subsequent nondominated fronts in

the order of their ranking Thus solutions from the set1198782are chosen next followed by solutions from the set 119878

3

and so forth This procedure is continued until no moresets can be accommodated Assume that the set 119878

119897is the

last nondominated set beyond which no other set can beaccommodated In general the count of solutions in all setsfrom 119878

1to 119878119897would be larger than the population size To

choose exactly119873 population members we sort the solutionsof the last front 119878

119897according to the crowded distance 119863

119901of

the density estimation in descending order and choose thebest solutions needed to make all population seats occupied

The procedure of the policy which is used for forming thenext population and ensures elitism of population evolutionin MOGA-LS is shown in Figure 4 The new population1198751of size 119873 is now used for selection crossover and

mutation to create a new population 1198761of size 119873 Repeat

the above process until meeting the maximum number 119894maxof iterations That is the maximum generation number 119894maxof population evolution is reached At that point the finalpopulation obtained by evolving should be 119875

119894maxminus1

Get the Final Solution Vector After the last round of iterationsin MOGA-LS the final population 119875

119894maxminus1is formed The

population 119875119894maxminus1

is sorted according to nondominationrank Its first nondominated set (120574

119901= 1) is the final Pareto

optimal solutions set Although the set of solutions vectorsis the final solutions set of the multiobjective optimizationproblem proposed by us on the location selection policy oflive VM migration the migrant VMs cannot be migratedaccording to multiple solutions

Generally the way to address this problem is to choosea solution randomly or according to the context from thePareto optimal set However as for the proposed Pareto mul-tiobjective optimization problem of live VMmigration policyin cloud environment MOGA-LS is a heuristic approachwhich does not have and does not need any context infor-mation and it should be a self-adaptive algorithm Besidesrandomly choosing a solution from the Pareto optimal set isnonefficient and one-sided

In this paper we have presented a novel approach toaddress the problem of selecting a solution from the obtainedset of Pareto optimal solutions Now assume that there are119871 solutions vectors in the Pareto optimal set The 119871 solutionsvectors can constitute a matrix Θ of 119871 lowast 119899 where each rowvector is a solution vector belonging to the Pareto optimalset Each of the 119899 column vectors of the matrix Θ represents119871 solutions of the location selection of the correspondingVM For each of the 119899 migrant VMs it has a solution spaceconsisting of 119871 solutions The 119871 solutions should consist ofa few kinds of solutions The probability of each solution is


Host1110

Host 9

70

Host1920

(a) The probability wheel of VM 1

Host 710

Host1030

Host2560

(b) The probability wheel of VM 2

Figure 5 Examples of the probability wheel

obtained by calculating their frequency of occurrence in the119871 solutions according to probability theory andmathematicalstatistics

At this point each migrant VM has a probability wheelas shown in Figure 5 A pointer randomly rotates on theprobability wheel The host in which the area where thepointer finally stops represents is the final target host of livemigration In the specific implementation a random numberlimited between 0 and 1 can be used to find the final locationof live VMmigration by finding the range where the randomnumber is By utilizing this approach each migrant VM getsthe final solution of location selection of live VM migrationAn 119899 dimension of solution vector is returned to migrationcontroller as an optimal migration policy of the 119899 migrantVMs

Let us consider a problem in the proposed approachWe know that the solution vector of the proposed problemought to be a sequence of integer numbers which denotethe hosts in the resource pool However each individualof the initial population is randomly initialized What ismore is that the crossover and mutation functions have somecoefficients which are limited between 0 and 1Thus althoughthe genes of each individual are limited to an integer typein the implementation the problem that the solution is notan integer still persists To address this problem we do notlimit the elements to an integer type but employ the smallestposition value (SPV) rules presented by in [16 17] In shortit is a rule which converts each individual given by the GAto a valid solution vector fit for the proposed problem Theprocess of applying the SPV rules into the proposed approachcan be understood as follows First the hosts in the resourcepool should be numbered from 0 to 119898 minus 1 Second after theinitial population is initialized all the genes of each individualare sorted in ascending order and then are numbered from 0to 119899 minus 1 and have a modulo operation of 119898 For instance ina time window Δ119905 there are four VMs to be migrated in aresource pool which has three hosts If an individual is (minus121329 minus012 126) it will be converted to (0 3 1 2) firstly andthen to (0 0 1 2) At this point the solution vector is usefuland meaningful It represents that the target host of VM 0 ishost 0The target host of VM 1 is host 0The target host of VM2 is host 1The target host of VM3 is host 2 In theMOGA-LS

all the solutions vectors (individuals) refer to the vectorswhich the original vectors have been converted to accordingto the SPV rules

Starting from this intuition the MOGA-LS algorithmwhich aims at live VM migration policy gives considerationto both the power saving optimization and the load balancingoptimization to achieve a more efficient cloud data centerby designing and utilizing the GA During the course ofthe study we have found that the two goals compete witheach other to further improve itself Thus the GA cannot beutilized directly into our proposed problem At this pointbased on some inherent contradictory relationship betweenthe two goals we have thought of game theory which isaiming to deal with and research this kind of problems In theprocess that we further study game theory andmultiobjectiveoptimization we have found the Pareto dominance theoryBased on the research on the Pareto dominance theory wehave proposed a novel multiobjective GA fit for being usedto address the proposed problem and designed its geneticoperators and fitness values as well as the policy of generatingthe next populationThe specific process ofMOGA-LS is alsopresented In the MOGA-LS algorithm each subalgorithmand subprocess are specially designed to better achieve ourproposed two optimization goals Therefore MOGA is anobviously efficient heuristic algorithm for location selectionpolicy of live VM migration from the experimental anddesignrsquos points of view

In the proposed MOGA-LS approach we have intro-duced the simulated annealing (SA) idea [18] into the processof selecting the final solution from the obtained Paretooptimal set Since the MOGA-LS algorithm is based on thePareto dominance theory the returned solution of MOGA-LS should be a set of Pareto optimal solutions (Pareto optimalfront) In our problem scenario it is very natural that only onefinal solution should be obtained as each VM needs to finallyidentify which host it will be migrated onto clearly In the setof Pareto optimal solutions it is obvious that all the solutionsare equal and not comparable Thus the idea of comparingthese solutions through other respects is meaningless andputting the cart before the horse If so we may as welldirectly add a new goal in the proposed MOP Finally theproblem still persists Thus we think the way that a final


solution is selected by comparing all the solutions of thePareto optimal front does notworkThewidely employedwayis to randomly selecting a solution from the Pareto optimalfront Its rationality lies in that all the Pareto optimal solutionsare optimal equal and not comparable as mentioned aboveAlthough themethod is simple and addresses this problem ina way randomly choosing is only an expedient and obviouslynot efficient Furthermore our problem scenario is also notrelated to any context information to help MOGA-LS selectthe final solution and the MOGA-LS approach is a self-adaptive heuristic algorithm which needs to achieve theselection of the final solution autonomously To addressthis problem in MOGA-LS we have utilized the SA ideato convert the seeming disadvantage that multiple optimalsolutions (Pareto optimal solutions set) are returned at last tothe advantage that each VM finally obtains a sample space ofoptimal solutions and can bemigrated to some optimal targethost with a certain probability and thus to achieve a long-term optimization by implementing the SA ideaThe specificprocess is shown above

There are many parameters in the proposed MOGA-LSalgorithmMost of them have great influences onMOGA-LSThe maximum number 119894max of iterations of the populationevolution and the population size 119873 of MOGA-LS are twoimportant parameters as well as their values are relatedto the efficiency and accuracy of MOGA-LS For instanceif the population size 119873 increases the computation time(convergence time) of MOGA-LS will increase and thepossibility that the algorithm is converged to the Paretooptimal front will also increase That is the global searchability will strengthen Moreover the required generationnumber 119894max of population evolution will decrease with thepopulation size 119873 increase The maximum number 119894max ofiterations of the population evolution should be neither toosmall nor too large Evidently the size of the 119894max value shouldmake the first Pareto nondominated solutions set of thepopulation of the MOGA-LS algorithm exactly converged tothe Pareto optimal front as much as possible Therefore boththe 119894max value and the 119873 value are the experience problemsand large numbers of experiments need to be performedto obtain fit values of them In the proposed MOGA-LSapproach 119873 is set to 20 and 119894max is set to 100 based on alarge number of experiments Also the mutation rate 119872

119903

is another important parameter in MOGA-LS since it has adirect impact on the diversity and convergence of populationinMOGA-LS It is difficult for GA to assign themutation rate119872119903a fit value as it is related to a number of factors Thus it

is an open problem which needs be studied further In thispaper the119872

119903value is set to 04That is there are 8 individuals

to be mutated in each round of iterations in our proposedMOGA-LS approach

4 Evaluation

In this section we have experimentally verified the proposedMOGA-LS approach The experiments have included evalu-ating the effect of power saving and load balancing ofMOGA-LS and verifying the performance of MOGA-LS migration

policy on the failure rate of migration events and SLA viola-tion Besides an auxiliary experiment has been conducted toget a best power management policy in the cloud data centerimplementing the proposed MOGA-LS policy In order tosimulate a dynamic cloud data center we have utilized anevent driven simulator named CloudSim toolkit [19] TheCloudSim framework enables the accounting of the totalpower consumed by the system during the simulation periodIn one word the calculating of power consumption has beenachieved by the CloudSim platform which has provided aclass including the methods getPower() CloudSim allowssuch simulation scenarios by supporting dynamic creation ofdifferent kinds of entities and can add and remove data centerentities at run time This functionality has achieved simulat-ing dynamic cloud environment where system componentscan join fail or leave the system randomly [19] OnCloudSimplatform we compare the proposed MOGA-LS approachwith random migration policy and optimal migration policyby power consumption the degree of load balancing andthe number of invalid VM migration We have preparedfour different kinds of experiments to evaluate and test theproposed approach The experimental results demonstratethat the proposed approach not only has a better powersaving and load balancing but also has a better migrationperformance Especially for a large number of migrationevents the MOGA-LS approach shows the stability and thebetter performance

41 Experimental Scenarios On CloudSim platform aresource pool consisting of 100 hosts is created These hostshave varying computing resource Twenty-four batches ofvirtual machine migration requests containing 13 requestsrandomly belonging to different hosts and with differentresource requirements are created The proposed MOGA-LSmodule is invoked and fetches the resource informationand state of the cloud resource pool periodically Δ119905 is setas 600 seconds and the migration events of each batch areuniformly distributed within an hour which includes 6Δ119905 sBesides the hostrsquos resource change rate is set as 1 time perhalf an hour

42 Comparison in Power Saving Theexperiment is designedfor verifying the efficiency and availability of MOGA-LSin power saving due to the location selection of live VMmigration during the long-term operation of a cloud datacenter In this scenario we compare MOGA-LS with randommigration policy dynamic load balancing (DLB) policyimplementing the idea of load balancing in [20] and theStdPSO migration policy which is an optimal migration pol-icy based on standard particle swarm optimization for onlypower saving by power consumption during the long-termoperation of the simulated cloud data center As illustrated inFigure 6 the cloud data center implementing DLBmigrationpolicy and the cloud data center implementing randommigration policy consume more power while the clouddata center implementing MOGA-LS and the cloud datacenter implementing the StdPSOmigration policies consumerelatively less power This is because both the latter two


0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO

Figure 6 Comparison of random-migration DLB StdPSO andMOGA-LS in power consumption

010

015

020

025

030

035

040

045

050

1 2 4 53Time (weeks)

Random-migration DLBMOGA-LSStdPSO

Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing

Figure 7 Comparison of StdPSO random-migration DLB andMOGA-LS in load balancing

are the heuristic approaches for power saving Further thecloud data center implementing DLBmigration policy causesmore power consumption than that of random migrationpolicy The reason for this is that the DLB migration policyhas only taken load balancing into account Achieving thebalancing of load is to improve the service performanceof cloud data center What it has been considering is theuser side The migration policy for only the above goalnaturally causes relatively more power consumption than therandom migration policy since the two objectives powerconsumption and load balancing are competing with eachother and the random migration policy does not sway thebalance on either side

Besides as can be seen in Figure 6 the cloud data centerimplementing the StdPSO migration policy has a less totalincremental power consumption than the cloud data centerimplementing the MOGA-LS migration policy The MOGA-LS approach has considered not only power saving but alsoload balancing Just from the point of view on power-savingeffect the MOGA-LS policy is not better than the StdPSOpolicy as mentioned above However the power-saving effectof the MOGA-LS policy is very close to that of the StdPSOpolicy in Figure 6 The MOGA-LS algorithm has achievedmany proposed optimization policy and thus had a betterconvergence to the Pareto optimal front and has utilized theSA idea to achieve a long-term optimization effect effectivelyavoiding the local optimization Therefore although it hasalso considered load balancing it obviously still has a betterpower-saving effect than the random migration policy andthe DLB policy Moreover its power saving effect approx-imates that of the StdPSO algorithm for optimally savingpower To sum up the proposed MOGA-LS approach is anefficient migration policy of live VM migration for powersaving

43 Comparison in Load Balancing In the experiment sce-nario it is designed for verifying the efficiency and availabilityof MOGA-LS in load balancing due to the location selectionof live VM migration during the long-term operation of acloud data center We have compared the degrees of loadbalancing in the cloud data centers implementing the StdPSOpolicy random migration policy MOGA-LS migration pol-icy and DLBmigration policy respectively at the end of eachof five weeks Here we have employed the standard deviationvalue indicated above and used to measure the degree of loadbalancing to conduct the experiment Obviously a smallerstandard deviation value represents that the cloud data centerhas the better balancing of load We know that in a clouddata center there is the dedicated load balancing algorithmrunning regularly in order to actively achieve the balancingof system load Thus the degree of load balancing of a clouddata center will be getting better and better to some extentwith the time passing As shown in Figure 7 all the fourpoliciesrsquo standard deviation values measuring the balancingof load have been tending to become smaller to some extentHowever at the end of each week the obtained four policiesrsquostandard deviation values have shown that there are thecertain gaps between the degrees of load balancing of clouddata centers In Figure 7 the cloud data center implementingthe StdPSO policy has the worst degree of load balancingThe reason is similar to the previous experiment The clouddata center implementing the random migration policy hasthe second worst effect of load balancing relatively since itdoes not have any information The DLB migration policyleads to the best effect of load balancing relativelyMeanwhilethe MOGA-LS approach is close to it and has the secondbest effect of load balancing As can be seen in Figure 7 thefour migration policies are divided into two levels of loadbalancing In other words the MOGA-LS migration policyand the DLBmigration policy are the same level although theMOGA-LS algorithm is not as excellent as theDLB algorithm


10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45

The number of VM migration requests

The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts

Random-migrationDLB MOGA-LS

StdPSO

Figure 8 Comparison of the number of failures in VM migrationevents

It is obvious that the MOGA-LS approach has already beenquite efficient for load balancing To sum up the proposedMOGA-LS approach is an efficient and feasible migrationpolicy of live VMmigration for load balancing

44 Comparison of the Number of Failures in VM MigrationEvents In the experiment scenario the dynamic host failureis simulated by CloudSim by scheduling some host failureevents and host shut down events to occur during the locationselection interval These events have caused some failuresin VM migration due to the nonavailability of the selectedhosts for some VM migration requests As illustrated inFigure 8 we have compared the MOGA-LS policy with therandom migration policy the DLB migration policy and thementioned StdPSO migration policy In the simulated clouddata center with the increase of the number of VMmigrationrequests the cloud data centers respectively implementingrandom migration DLB and StdPSO result in more numberof invalid VM migrations whereas the MOGA-LS performsbetter in finding the fit hosts in the dynamic resource poolThis is because the memory data is outdated in the StdPSODLB and random migration policy They cannot have anadjustment with the environment changed Conversely sincethe SA idea has been introduced into it the MOGA-LSpolicy based on the GA is quite efficient in detecting the hostfailures during the interval as well as has a fit adjustmentin a better manner by searching out the new availablehosts that can meet the resource requirements of the VMmigration requestsTheMOGA-LS approach can better meetthe performance requirement of live VM migration andreduces the failure numbers of live VM migration since theMOGA-LS algorithm has seen the migration performanceas the constraint objective The design of the MOGA-LSapproach is rational and efficient After all it is the migrationpolicy of VMs that we are discussing And it should firstmakelive migration of VMs succeed as much as possible

0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)

Workload in the hosts ()

StdPSOMOGA-LS

Figure 9 Comparison of the incremental cost due to SLA violationin VM migration with varying percentage of load in the cloud datacenter

45 Comparison of the Incremental Cost due to SLA ViolationIn this experimental scenario we compare the incrementalcost of StdPSO and MOGA-LS due to service level agree-ments (SLA) violation with varying percentage of load inthe cloud data center The load mentioned refers to theload of the whole cloud data center and is not the loadof some physical host As illustrated in Figure 9 with theincrease of the percentage of workload in the cloud datacenter the incremental cost of MOGA-LS is less than that ofStdPSO due to SLA violation We know if the workload inthe hosts is heavier achieving the balancing of load is moreimportant for better service performance and to thus meetthe SLA The above experiment has shown that the StdPSOpolicy and the MOGA-LS policy belong to the same level forpower savingMOGA-LS is just slightly inferior than StdPSOHowever since the proposed MOGA-LS approach has alsoachieved the balancing of load it has ability in making thecloud data centers increase resource utilization and providebetter service performance to users Thus its cost due toSLA violation is less So to say the MOGA-LS approach hasnot only made a quite excellent power-saving effect cometrue to contribute to the green cloud data centers but also isconsidered the most important user experience and achievedload balancing and thus to enhance the resource utilizationand service capability of green cloud data centers To sumup the MOGA-LS algorithm is an efficient location selectionpolicy of live VMmigration

46 Tradeoff between Power Saving and Performance Ful-fillment The experimental scenario is conducted to findthe optimal power management policy which balances thebenefits due to power saving and performance fulfillment


Since live VMmigration events are time critical VM requestsand the cloud service provider should meet strict SLA com-pliance any violations in SLA in terms of performance lossof VMs will result in penalty cost on the provider of a clouddata center The performance loss mentioned may be causedby power saving liveVMmigration network bandwidth andthroughput and so forthHowever this paper has not focusedon these problems This experiment is designed to only findout a better power management policy suitable for the cloudenvironment implementing the proposed MOGA-LS to havea trade-off between the benefit of power saving and penaltycost due to SLA violation In other words the experimentaims to find a power management policy which makes theproposed MOGA-LS approach more efficient in the clouddata centers and makes the cloud data centers implementingthe MOGA-LS migration policy have a better power-savingeffect and satisfy the users There are four different policiesto be formulated The first policy is onoff policy whereinall idle hosts are switched off It can be seen that the policygives the best power saving but it causes the high penaltycost obviously The single-DSS policy is the second policywherein all idle hosts are switched to deep sleep state Itresults in an increase in the power consumption cost butthe penalty cost is reduced significantly The third policy issingle-SSS wherein all idle hosts are switched to shallow sleepstate There is no penalty cost as SLA violation is absent inthis policy However the enormous increase in the powerconsumption cost is caused The multiple-SS is the fourthpolicy wherein some of the idle hosts are kept in deep sleepand others are kept in shadow sleep state according to a short-term prediction technology From its meaning and as shownin Figure 10 it can be known that themultiple-SS policy givesthe optimal cost trade-off relatively

5 Conclusion and Future Work

In this paper a novel location selection policy MOGA-LSof live VM migration is proposed and we give its mainidea design implementation and evaluation It employs theimproved GA-based approach and the Pareto dominanceidea In the improved GA-based approach we have designedthe selection crossover the crossover operator and the muta-tion operator of MOGA-LS Moreover we have designed thefitness values of MOGA-LS and given how to obtain themAlso in order to make the MOGA-LS approach have theelitism we have designed and employed a novel (120583 + 120582)

selection policy based on Pareto nondominated sorting togenerate the next population and thus to further optimizationthe MOGA-LS algorithm

It is noteworthy that in the proposedMOGA-LS approachwe do not randomly get the initial population but employa novel optimization method utilizing the hill-climbingtechnique and clustering technique In this paper we do notgive its details Besides in the process of generating the nextpopulation we have introduced 119870-means clustering into theprocess to address this problem that the number of the indi-viduals of the first nondominated set may be larger than 119873In the proposedMOGA-LS we have also utilized the SA ideato obtain the final solution vector and achieve a long-term

05

101520253035

Onoff Single-DSS Multiple-SS Single-SSSThe power policies

In penalty cost due to SLA violationIn power consumption cost

Incr

ease

in co

st (

)

Figure 10 The trade-off between power and performance withdifferent power management policies

optimization Also aiming to connect the improved GA-based algorithm and the SA process we take use of the prob-ability theory and mathematical statistics as well as the char-acteristics (the returned solution is the set of Pareto optimalsolutions) of the algorithm itself to obtain and process data

MOGA-LS achieves the high-efficiency of power con-sumption and the stability of requirement performance Itnot only minimizes the incremental power consumptionof a cloud data center but also achieves the balancing ofsystem load while minimizing the number of failure inVM migration events relatively In the proposed MOGA-LS approach there are some open problems which need bestudied further and experience problems which need manyexperiments to gradually get a better solutionThemaximumnumber 119894max of iterations of the population evolution and thepopulation size119873 of MOGA-LS are the experience problemsand need to perform several experiments to obtain the fitvalues to thus make the MOGA-LS approach efficient andfeasible The mutation rate 119872

119903of our approach is an open

problem which needs be further widely researched In thispaper all the parameters are set to the fit values respectivelyTo evaluate the MOGA-LS approach we have conductedseveral experiments on the CloudSim platform The finalexperimental results show that MOGA-LS has an excellentpower-saving effect and achieves the balancing of systemload while having a high success rate of live VM migrationevents Therefore MOGA-LS is an effective heuristic andself-adaptive location selection policy of live VM migrationfor power conservation and load balancing

Aiming to further improve the performance of MOGA-LS we plan to study the robustness of MOGA-LS in the nextstep work MOGA-LS should have abilities in dealing withsome sudden matters and be combined with the mechanismof live VM migration for achieving a more efficient hybridmanagement In the next work and experiments the experi-ence problems and the open problems presented in this paperwill also be researched further

Acknowledgments

The authors would like to thank the editors and anonymousreviewers for their valuable comments


References

[1] P Barham B Dragovic K Fraser et al ldquoXen and the art ofvirtualizationrdquo in Proceedings of the 19th ACM Symposium onOperating Systems Principles (SOSP rsquo03) pp 164ndash177 BoltonLanding NY USA October 2003

[2] Y Li W Li and C Jiang ldquoA survey of virtual machine systemcurrent technology and future trendsrdquo in Proceedings of the 3rdInternational Symposium on Electronic Commerce and Security(ISECS rsquo10) pp 332ndash336 Guangzhou China July 2010

[3] M Armbrust A Fox R Griffith et al ldquothe clouds a view ofcloud computingrdquo Tech Rep EECS-2009-28 2009

[4] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002

[5] A Konak D W Coit and A E Smith ldquoMulti-objectiveoptimization using genetic algorithms a tutorialrdquo ReliabilityEngineering and System Safety vol 91 no 9 pp 992ndash1007 2006

[6] C Rusu A Ferreira C Scordino AWatson RMelhem andDMosse ldquoEnergy-efficient real-time heterogeneous server clus-tersrdquo in Proceedings of the 12th IEEE Real-Time and EmbeddedTechnology and Applications Symposium pp 418ndash428 San JoseCalif USA April 2006

[7] S Srikantaiah A Kansal and F Zhao ldquoEnergy aware consolida-tion for cloud computingrdquo Cluster Computing vol 12 pp 1ndash152009

[8] A Verma P Ahuja and A Neogi ldquoPMapper power andmigra-tion cost aware application placement in virtualized systems(PMapper rsquo08)rdquo in Proceedings of the 9th ACMIFIPUSENIXInternational Conference onMiddleware pp 243ndash264 SpringerNew York NY USA

[9] B Li J Li J Huai T Wo Q Li and L Zhong ldquoEnaCloudan energy-saving application live placement approach for cloudcomputing environmentsrdquo in IEEE International Conference onCloud Computing (CLOUD rsquo09) pp 17ndash24 Bangalore IndiaSeptember 2009

[10] R Jeyarani N Nagaveni and R V Ram ldquoSelf adaptive particleswarm optimization for efficient virtual machine provisioningin cloudrdquo International Journal of Intelligent Information Tech-nologies vol 7 no 2 pp 25ndash44 2011

[11] S Jing and K She ldquoA novel model for load balancing in clouddata centerrdquo Journal of Convergence InformationTechnology vol6 no 4 pp 171ndash179 2011

[12] Z Yang G Dong and C Li ldquoAlgorithm study of load balancebased on live migration of VM in cloud computingrdquo Interna-tional Journal of Advancements in Computing Technology vol 5no 3 pp 481ndash489 2013

[13] R Jeyarani R V Ram and N Nagaveni ldquoImplementation ofefficient light weight internal scheduler for high throughputgrid environmentrdquo in Proceedings of the National Conferenceon Advanced Computing in Computer Applications (NCACCArsquo09) E Jeyakumar and R Rangarajan Eds pp 283ndash289Coimbatore India 2009

[14] S Sadhasivam R Jayarani NNagaveni andRV Ram ldquoDesignand implementation of an efficient two-level scheduler for cloudcomputing environmentrdquo in Proceedings of the InternationalConference on Advances in Recent Technologies in Communica-tion and Computing (ARTCOM rsquo09) pp 884ndash886 KottayamIndia October 2009

[15] S Bandyopadhyay S Saha U Maulik and K Deb ldquoA simu-lated annealing-based multiobjective optimization algorithm

AMOSArdquo IEEE Transactions on Evolutionary Computation vol12 no 3 pp 269ndash283 2008

[16] M F Tasgetiren M Sevkli Y C Liang and G GencyilmazldquoParticle swarm optimization algorithm for permutation flow-shop sequencing problemrdquo in Ant Colony Optimization andSwarm vol 3172 of Lecture Notes in Computer Science pp 382ndash389 Springer Berlin Germany 2004

[17] M F Tasgetiren Y C Liang M Sevkli and G GencyilmazldquoParticle swarmoptimization algorithm for singlemachine totalweighted tardiness problemrdquo in Proceedings of the Congress onEvolutionary Computation (CEC rsquo04) pp 1412ndash1419 PortlandOre USA June 2004

[18] S Kirkpatrick C D Gelatt Jr and M P Vecchi ldquoOptimizationby simulated annealingrdquo Science vol 220 no 4598 pp 671ndash6801983

[19] 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 vol 41 no 1 pp 23ndash50 2011

[20] M H Willebeek-LeMair and A P Reeves ldquoStrategies fordynamic load balancing on highly parallel computersrdquo IEEETransactions on Parallel and Distributed Systems vol 4 no 9pp 979ndash993 1993

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



In this paper we have focused on the live VM migra-tion policy for power saving and load balancing Generallyspeaking the migrant VM has many available target hostsHowever only one target host is most suitable for the VMin order to achieve the management goals which includeminimizing the total incremental power consumption incloud data center and achieving load balancing as much aspossible in the case that the performance constraint of liveVMmigration is met

Many papers have presented some heuristic algorithmsto find optimal solutions aiming to minimize power con-sumption However most research has only taken powerconsumption into consideration but not considered loadbalancing In other word the current research has focusedon a single objective optimization for minimizing the powerconsumption In this paper we have focused on the powerconsumption and load balancing It is a multiobjectiveoptimization problem (MOP) which differs from the tradi-tional single objective optimization problem The basic ideaof the existing research is that according to the currentsituation and history of a cloud data center the controllerhas searched for a best migration policy by optimizing theobjective function of the proposed algorithm In the MOPsthe situation is different and complex It needs to optimizemultiple objectives at the same moment Generally theseobjectives influence each other and are even contradictoryTo achieve a multiobjective optimization the problem ofhow to evaluate and compare two solutions for the multiplefunctions as well as the problem of how to give each solutiona reasonable fitness value towards multiple functions must beaddressed first

It is well known that it is the Pareto dominance principlethat is used for MOPs to address the problems Accordingto the Pareto theory we know that the optimal solution ofan MOP is not a single solution but a set of Pareto optimalsolutionsThe common and classical heuristic algorithms areused to convert the MOP to a single-objective optimizationproblem by emphasizing one particular Pareto optimal solu-tion at a time When such an algorithm will be used to searchout multiple solutions it has to be executed many timeshopefully finding a different solution at each simulation runHowever the genetic algorithm (GA) as the representativeof evolution algorithms (EAs) is suitable for dealing with theMOPs since it has the ability to find multiple Pareto optimalsolutions in one single simulation run As EAs work witha population of solutions a simple EA can be extended tomaintain a diverse set of solutions With an emphasis formoving towards the true Pareto optimal region an EA can beused to find multiple Pareto optimal solutions in one singlesimulation run [4 5] In this paper we have proposed anovel heuristic GA-based location selection policy MOGA-LS of VM for power saving and load balancing whichutilizes the Pareto nondominated sorting individual densityestimation and mathematical statistics of Pareto optimalsolutions to achieve a long-term excellent power saving andload balancing optimization The final experimental resultshave demonstrated that the MOGA-LS approach not onlyreduces the total incremental power consumption but alsoachieves a better load balancing while relatively maximizing

the performance of live VMmigrationThat is it has reducedthe failure events of live VM migration and contributed toachieving better green cloud data centers with load balancing

The rest of the paper is organized as follows In Section 2we present the related work and the reasonable prerequisitesare shown clearly In Section 3 the main idea design prob-lem formulation solution representation and implementa-tion of MOGA-LS are introduced in details In Section 4 theexperimental results and analysis on CloudSim platform aregiven Finally in Section 5 we summarize the full paper andfuture work is put forward

2 Related Work

As far as we know the proposed problem which refers tofinding a fit target host for a live VM migration according tothe standard of minimizing the increment power consump-tion or load balancing has not been widely researched inthe related fields of live VM migration policy let alone bothHowever most researchers have focused on some problemswhich are similar to the proposed problem in this paperThusthe related work of the kind of problems relating to live VMmigration towards power saving and load balancing will bediscussed briefly in this section

Rusu et al in [6] have presented a cluster-wide QoS-aware technique that dynamically reconfigures the cluster toreduce energy consumption during periods of reduced loadThe proposed system consists of two important componentsnamely front end manager and a local manager While thefront end manager finds the servers which should be turnedon or off in terms of a given system load the local managerwill utilize dynamic voltage and frequency scaling (DVFS)technique to conserve energy The main shortage of theapproach is the onoff policy It relies on the table of valuesand needs offline computing However the system does notmake use of server consolidation throughVMmigration andthus its onoff policy may not be much effective

Srikantaiah et al [7] have investigated the problem ofdynamic consolidation of applications serving small statelessrequests in data centers tominimize the energy consumptionThey modeled the problem as a multidimensional bin pack-ing problem However the proposed model doesnrsquot describethe degradation of performance due to the consolidationBesides the energy drawn may rely on a particular set ofapplications combined on a computer node A heuristic forthe defined bin packing problem is proposed by the authorsThe heuristic is based upon the idea of minimizing thesum of the current allocationsrsquo Euclidean distances to theoptimal point at each server The application workload willbe allocated to a server using the proposed heuristic sincea request to execute a new application is received Withoutthe sufficient capacity of all active servers the system willswitch on a new server while reallocating all the applicationsusing the same heuristic in an arbitrary order The proposedapproach is fit for heterogeneous environments however ithas several shortcomings First the approach assumes thatall applicationsrsquo resource requirements are known in advanceand constant Second performance and energy overhead























119903rep-


120578119903=

119899

sum

119894=1

119898

sum

119895=1

119905119895

119894= 119899 119894 isin 1 2 3 119899 119895 isin 1 2 3 119898

119903 isin 1 2 3 119904

(1)



120591119895

119894=

1


Host119895and rcrVM

119894le acrHost

119895

119894 isin 1 2 3 119899

0


Host119895and rcrVM

119894le acrHost

119895

119895 isin 1 2 3 119898

Invalidif VM


119895

(2)







0(119896)] 119903 isin








0(119896)] with


is defined by

Δ119875 = (120585119903

119896) minus (120585

119903

119896minus1) 119903 isin 1 2 3 119904 (3)



120575119875119903

=

119899

sum

119896=1

((120585119903

119896) minus (120585

119903

119896minus1)) 119903 isin 1 2 3 119904 (4)




defined as follows

119877119894119903

= 120572119877119894119903

CPU + 120573119877119894119903

MEM 119894 isin 1 2 3 119898

119903 isin 1 2 3 119904

120572 + 120573 = 1

(5)

where 119877119894119903

CPU and 119877119894119903



119864119894119903

=

119877119894119903

119879119894 119894 isin 1 2 3 119898 119903 isin 1 2 3 119904 (6)


119879119894

= 120572119879119894

CPU + 120573119879119894

MEM 119894 isin 1 2 3 119898 (7)



119864 (119883) =

sum119873

119894=1119883119894

119873

120590 = radic1

119873

119873

sum

119894=1

(119883119894minus 119864 (119883))

2

(8)


120590119903

=radic1

119898

119898

sum

119894=1

(

120572119877119894119903

CPU + 120573119877119894119903

MEM120572119879119894

CPU + 120573119879119894

MEMminus

sum119898

119896=1((120572119877119896119903

CPU + 120573119877119896119903

MEM) (120572119879119896

CPU + 120573119879119896

MEM))

119898

)

2

(9)


Min

120575119875119903

=

119899

sum

119896=1

((120585119903

119896) minus (120585

119903

119896minus1))

120590119903

= radic1

119898

119898

sum

119894=1

(

120572119877119894119903

CPU + 120573119877119894119903

MEM120572119879119894

CPU + 120573119879119894

MEMminus

sum119898

119896=1((120572119877119896119903

CPU + 120573119877119896119903

MEM) (120572119879119896

CPU + 120573119879119896

MEM))

119898

)

2

119903 isin 1 2 3 119904

st 120578119903=

119899

sum

119894=1

119898

sum

119895=1

120601119895

119894= 119899 119894 isin 1 2 3 119899 119895 isin 1 2 3 119898 119903 isin 1 2 3 119904

(10)





119899)


119899) (119906 ≺


119894(V) and

exist119895 isin 1 2 3 119896 st 119891119895(119906) lt 119891




(1199091 1199092 119909


notexist1198831015840

isin Ω st1198831015840 ≺ 119883


isin Ω 1198831015840

≺

119883

Pareto Front PF = 119865(119883) = (1198911(119883) 119891

2(119883) 119891

119896(119883)) | 119883 isin










1198941 119909119896

1198942 119909

119896

119894119895 119909

119896

119894119863)






119894and119883119905



119894and

119883119905+1


119883119905+1

119894= 120572119883119905

119894+ (1 minus 120572)119883

119905

119895

119883119905+1

119895= (1 minus 120572)119883

119905

119894+ 120572119883119905

119895

(11)



1 1198832 119883

119896 119883


new individual 1198831015840 = (1198831 1198832 119883

1015840

119896 119883

119899) If the gene



[119880119896

min 119880119896


formula

1198831015840

119896=

119883119896+ (119880119896


120576

119896)

if 119903 and (0 1) isin [0 05)

119883119896minus (119883119896minus 119880119896

min) (1 minus 120574(1minus(119905119879))

120576

119896)

if 119903 and (0 1) isin [05 1]

(12)




120576

119896) ranges


119896and 120576









119901values



119901


Cuboid

f2

f1

i minus 1

Dp

i

i + 1







119901serves












119901value is


defined as follows

119877119901= 1 +

119905

sum

119895=1

120574119883119895

(13)


1198831 1205741198832 120574



1and


119901value of









119901values





f2

f1

A

B

rP = 1

rP = 2

rP = 3

rP = 4

RP = 2

RP = 5




119901


as follows

119894≺119877119901120574119901119863119901

119895 if (119894119877119901

lt 119895119877119901

)

or ((119894119877119901

= 119895119877119901

) (119894120574119901

lt 119895120574119901

))

or ((119894119877119901

= 119895119877119901

) (119894120574119901

= 119895120574119901

)

(119894119863119901

gt 119895119863119901

))

(14)







Monitor

MOGA-LS


Placement policy

Host m

Host 1

Host 2



Resource pool









119901value of each










0obtains its 119877

119901value






follows

119894119863119901

=radic(


1

119891max1

minus 119891min1

)

2

+(


2

119891max2

minus119891min2

)

2

(15)






119877119901120574119901119863119901


120572 =

119895120574119901

119894120574119901

+ 119895120574119901

(16)









0




Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1

(a)


Rejected


Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1120574p = 3

S3

(b)












3


119897is the





119901of





119894maxminus1










Host1110

Host 9

70

Host1920


Host 710

Host1030

Host2560












119903





4 Evaluation






0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in

























119903rep-


120578119903=

119899

sum

119894=1

119898

sum

119895=1

119905119895

119894= 119899 119894 isin 1 2 3 119899 119895 isin 1 2 3 119898

119903 isin 1 2 3 119904

(1)



120591119895

119894=

1


Host119895and rcrVM

119894le acrHost

119895

119894 isin 1 2 3 119899

0


Host119895and rcrVM

119894le acrHost

119895

119895 isin 1 2 3 119898

Invalidif VM


119895

(2)







0(119896)] 119903 isin








0(119896)] with


is defined by

Δ119875 = (120585119903

119896) minus (120585

119903

119896minus1) 119903 isin 1 2 3 119904 (3)



120575119875119903

=

119899

sum

119896=1

((120585119903

119896) minus (120585

119903

119896minus1)) 119903 isin 1 2 3 119904 (4)




defined as follows

119877119894119903

= 120572119877119894119903

CPU + 120573119877119894119903

MEM 119894 isin 1 2 3 119898

119903 isin 1 2 3 119904

120572 + 120573 = 1

(5)

where 119877119894119903

CPU and 119877119894119903



119864119894119903

=

119877119894119903

119879119894 119894 isin 1 2 3 119898 119903 isin 1 2 3 119904 (6)


119879119894

= 120572119879119894

CPU + 120573119879119894

MEM 119894 isin 1 2 3 119898 (7)



119864 (119883) =

sum119873

119894=1119883119894

119873

120590 = radic1

119873

119873

sum

119894=1

(119883119894minus 119864 (119883))

2

(8)


120590119903

=radic1

119898

119898

sum

119894=1

(

120572119877119894119903

CPU + 120573119877119894119903

MEM120572119879119894

CPU + 120573119879119894

MEMminus

sum119898

119896=1((120572119877119896119903

CPU + 120573119877119896119903

MEM) (120572119879119896

CPU + 120573119879119896

MEM))

119898

)

2

(9)


Min

120575119875119903

=

119899

sum

119896=1

((120585119903

119896) minus (120585

119903

119896minus1))

120590119903

= radic1

119898

119898

sum

119894=1

(

120572119877119894119903

CPU + 120573119877119894119903

MEM120572119879119894

CPU + 120573119879119894

MEMminus

sum119898

119896=1((120572119877119896119903

CPU + 120573119877119896119903

MEM) (120572119879119896

CPU + 120573119879119896

MEM))

119898

)

2

119903 isin 1 2 3 119904

st 120578119903=

119899

sum

119894=1

119898

sum

119895=1

120601119895

119894= 119899 119894 isin 1 2 3 119899 119895 isin 1 2 3 119898 119903 isin 1 2 3 119904

(10)





119899)


119899) (119906 ≺


119894(V) and

exist119895 isin 1 2 3 119896 st 119891119895(119906) lt 119891




(1199091 1199092 119909


notexist1198831015840

isin Ω st1198831015840 ≺ 119883


isin Ω 1198831015840

≺

119883

Pareto Front PF = 119865(119883) = (1198911(119883) 119891

2(119883) 119891

119896(119883)) | 119883 isin










1198941 119909119896

1198942 119909

119896

119894119895 119909

119896

119894119863)






119894and119883119905



119894and

119883119905+1


119883119905+1

119894= 120572119883119905

119894+ (1 minus 120572)119883

119905

119895

119883119905+1

119895= (1 minus 120572)119883

119905

119894+ 120572119883119905

119895

(11)



1 1198832 119883

119896 119883


new individual 1198831015840 = (1198831 1198832 119883

1015840

119896 119883

119899) If the gene



[119880119896

min 119880119896


formula

1198831015840

119896=

119883119896+ (119880119896


120576

119896)

if 119903 and (0 1) isin [0 05)

119883119896minus (119883119896minus 119880119896

min) (1 minus 120574(1minus(119905119879))

120576

119896)

if 119903 and (0 1) isin [05 1]

(12)




120576

119896) ranges


119896and 120576









119901values



119901


Cuboid

f2

f1

i minus 1

Dp

i

i + 1







119901serves












119901value is


defined as follows

119877119901= 1 +

119905

sum

119895=1

120574119883119895

(13)


1198831 1205741198832 120574



1and


119901value of









119901values





f2

f1

A

B

rP = 1

rP = 2

rP = 3

rP = 4

RP = 2

RP = 5




119901


as follows

119894≺119877119901120574119901119863119901

119895 if (119894119877119901

lt 119895119877119901

)

or ((119894119877119901

= 119895119877119901

) (119894120574119901

lt 119895120574119901

))

or ((119894119877119901

= 119895119877119901

) (119894120574119901

= 119895120574119901

)

(119894119863119901

gt 119895119863119901

))

(14)







Monitor

MOGA-LS


Placement policy

Host m

Host 1

Host 2



Resource pool









119901value of each










0obtains its 119877

119901value






follows

119894119863119901

=radic(


1

119891max1

minus 119891min1

)

2

+(


2

119891max2

minus119891min2

)

2

(15)






119877119901120574119901119863119901


120572 =

119895120574119901

119894120574119901

+ 119895120574119901

(16)









0




Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1

(a)


Rejected


Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1120574p = 3

S3

(b)












3


119897is the





119901of





119894maxminus1










Host1110

Host 9

70

Host1920


Host 710

Host1030

Host2560












119903





4 Evaluation






0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in













119903rep-


120578119903=

119899

sum

119894=1

119898

sum

119895=1

119905119895

119894= 119899 119894 isin 1 2 3 119899 119895 isin 1 2 3 119898

119903 isin 1 2 3 119904

(1)



120591119895

119894=

1


Host119895and rcrVM

119894le acrHost

119895

119894 isin 1 2 3 119899

0


Host119895and rcrVM

119894le acrHost

119895

119895 isin 1 2 3 119898

Invalidif VM


119895

(2)







0(119896)] 119903 isin








0(119896)] with


is defined by

Δ119875 = (120585119903

119896) minus (120585

119903

119896minus1) 119903 isin 1 2 3 119904 (3)



120575119875119903

=

119899

sum

119896=1

((120585119903

119896) minus (120585

119903

119896minus1)) 119903 isin 1 2 3 119904 (4)




defined as follows

119877119894119903

= 120572119877119894119903

CPU + 120573119877119894119903

MEM 119894 isin 1 2 3 119898

119903 isin 1 2 3 119904

120572 + 120573 = 1

(5)

where 119877119894119903

CPU and 119877119894119903



119864119894119903

=

119877119894119903

119879119894 119894 isin 1 2 3 119898 119903 isin 1 2 3 119904 (6)


119879119894

= 120572119879119894

CPU + 120573119879119894

MEM 119894 isin 1 2 3 119898 (7)



119864 (119883) =

sum119873

119894=1119883119894

119873

120590 = radic1

119873

119873

sum

119894=1

(119883119894minus 119864 (119883))

2

(8)


120590119903

=radic1

119898

119898

sum

119894=1

(

120572119877119894119903

CPU + 120573119877119894119903

MEM120572119879119894

CPU + 120573119879119894

MEMminus

sum119898

119896=1((120572119877119896119903

CPU + 120573119877119896119903

MEM) (120572119879119896

CPU + 120573119879119896

MEM))

119898

)

2

(9)


Min

120575119875119903

=

119899

sum

119896=1

((120585119903

119896) minus (120585

119903

119896minus1))

120590119903

= radic1

119898

119898

sum

119894=1

(

120572119877119894119903

CPU + 120573119877119894119903

MEM120572119879119894

CPU + 120573119879119894

MEMminus

sum119898

119896=1((120572119877119896119903

CPU + 120573119877119896119903

MEM) (120572119879119896

CPU + 120573119879119896

MEM))

119898

)

2

119903 isin 1 2 3 119904

st 120578119903=

119899

sum

119894=1

119898

sum

119895=1

120601119895

119894= 119899 119894 isin 1 2 3 119899 119895 isin 1 2 3 119898 119903 isin 1 2 3 119904

(10)





119899)


119899) (119906 ≺


119894(V) and

exist119895 isin 1 2 3 119896 st 119891119895(119906) lt 119891




(1199091 1199092 119909


notexist1198831015840

isin Ω st1198831015840 ≺ 119883


isin Ω 1198831015840

≺

119883

Pareto Front PF = 119865(119883) = (1198911(119883) 119891

2(119883) 119891

119896(119883)) | 119883 isin










1198941 119909119896

1198942 119909

119896

119894119895 119909

119896

119894119863)






119894and119883119905



119894and

119883119905+1


119883119905+1

119894= 120572119883119905

119894+ (1 minus 120572)119883

119905

119895

119883119905+1

119895= (1 minus 120572)119883

119905

119894+ 120572119883119905

119895

(11)



1 1198832 119883

119896 119883


new individual 1198831015840 = (1198831 1198832 119883

1015840

119896 119883

119899) If the gene



[119880119896

min 119880119896


formula

1198831015840

119896=

119883119896+ (119880119896


120576

119896)

if 119903 and (0 1) isin [0 05)

119883119896minus (119883119896minus 119880119896

min) (1 minus 120574(1minus(119905119879))

120576

119896)

if 119903 and (0 1) isin [05 1]

(12)




120576

119896) ranges


119896and 120576









119901values



119901


Cuboid

f2

f1

i minus 1

Dp

i

i + 1







119901serves












119901value is


defined as follows

119877119901= 1 +

119905

sum

119895=1

120574119883119895

(13)


1198831 1205741198832 120574



1and


119901value of









119901values





f2

f1

A

B

rP = 1

rP = 2

rP = 3

rP = 4

RP = 2

RP = 5




119901


as follows

119894≺119877119901120574119901119863119901

119895 if (119894119877119901

lt 119895119877119901

)

or ((119894119877119901

= 119895119877119901

) (119894120574119901

lt 119895120574119901

))

or ((119894119877119901

= 119895119877119901

) (119894120574119901

= 119895120574119901

)

(119894119863119901

gt 119895119863119901

))

(14)







Monitor

MOGA-LS


Placement policy

Host m

Host 1

Host 2



Resource pool









119901value of each










0obtains its 119877

119901value






follows

119894119863119901

=radic(


1

119891max1

minus 119891min1

)

2

+(


2

119891max2

minus119891min2

)

2

(15)






119877119901120574119901119863119901


120572 =

119895120574119901

119894120574119901

+ 119895120574119901

(16)









0




Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1

(a)


Rejected


Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1120574p = 3

S3

(b)












3


119897is the





119901of





119894maxminus1










Host1110

Host 9

70

Host1920


Host 710

Host1030

Host2560












119903





4 Evaluation






0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





120575119875119903

=

119899

sum

119896=1

((120585119903

119896) minus (120585

119903

119896minus1)) 119903 isin 1 2 3 119904 (4)




defined as follows

119877119894119903

= 120572119877119894119903

CPU + 120573119877119894119903

MEM 119894 isin 1 2 3 119898

119903 isin 1 2 3 119904

120572 + 120573 = 1

(5)

where 119877119894119903

CPU and 119877119894119903



119864119894119903

=

119877119894119903

119879119894 119894 isin 1 2 3 119898 119903 isin 1 2 3 119904 (6)


119879119894

= 120572119879119894

CPU + 120573119879119894

MEM 119894 isin 1 2 3 119898 (7)



119864 (119883) =

sum119873

119894=1119883119894

119873

120590 = radic1

119873

119873

sum

119894=1

(119883119894minus 119864 (119883))

2

(8)


120590119903

=radic1

119898

119898

sum

119894=1

(

120572119877119894119903

CPU + 120573119877119894119903

MEM120572119879119894

CPU + 120573119879119894

MEMminus

sum119898

119896=1((120572119877119896119903

CPU + 120573119877119896119903

MEM) (120572119879119896

CPU + 120573119879119896

MEM))

119898

)

2

(9)


Min

120575119875119903

=

119899

sum

119896=1

((120585119903

119896) minus (120585

119903

119896minus1))

120590119903

= radic1

119898

119898

sum

119894=1

(

120572119877119894119903

CPU + 120573119877119894119903

MEM120572119879119894

CPU + 120573119879119894

MEMminus

sum119898

119896=1((120572119877119896119903

CPU + 120573119877119896119903

MEM) (120572119879119896

CPU + 120573119879119896

MEM))

119898

)

2

119903 isin 1 2 3 119904

st 120578119903=

119899

sum

119894=1

119898

sum

119895=1

120601119895

119894= 119899 119894 isin 1 2 3 119899 119895 isin 1 2 3 119898 119903 isin 1 2 3 119904

(10)





119899)


119899) (119906 ≺


119894(V) and

exist119895 isin 1 2 3 119896 st 119891119895(119906) lt 119891




(1199091 1199092 119909


notexist1198831015840

isin Ω st1198831015840 ≺ 119883


isin Ω 1198831015840

≺

119883

Pareto Front PF = 119865(119883) = (1198911(119883) 119891

2(119883) 119891

119896(119883)) | 119883 isin










1198941 119909119896

1198942 119909

119896

119894119895 119909

119896

119894119863)






119894and119883119905



119894and

119883119905+1


119883119905+1

119894= 120572119883119905

119894+ (1 minus 120572)119883

119905

119895

119883119905+1

119895= (1 minus 120572)119883

119905

119894+ 120572119883119905

119895

(11)



1 1198832 119883

119896 119883


new individual 1198831015840 = (1198831 1198832 119883

1015840

119896 119883

119899) If the gene



[119880119896

min 119880119896


formula

1198831015840

119896=

119883119896+ (119880119896


120576

119896)

if 119903 and (0 1) isin [0 05)

119883119896minus (119883119896minus 119880119896

min) (1 minus 120574(1minus(119905119879))

120576

119896)

if 119903 and (0 1) isin [05 1]

(12)




120576

119896) ranges


119896and 120576









119901values



119901


Cuboid

f2

f1

i minus 1

Dp

i

i + 1







119901serves












119901value is


defined as follows

119877119901= 1 +

119905

sum

119895=1

120574119883119895

(13)


1198831 1205741198832 120574



1and


119901value of









119901values





f2

f1

A

B

rP = 1

rP = 2

rP = 3

rP = 4

RP = 2

RP = 5




119901


as follows

119894≺119877119901120574119901119863119901

119895 if (119894119877119901

lt 119895119877119901

)

or ((119894119877119901

= 119895119877119901

) (119894120574119901

lt 119895120574119901

))

or ((119894119877119901

= 119895119877119901

) (119894120574119901

= 119895120574119901

)

(119894119863119901

gt 119895119863119901

))

(14)







Monitor

MOGA-LS


Placement policy

Host m

Host 1

Host 2



Resource pool









119901value of each










0obtains its 119877

119901value






follows

119894119863119901

=radic(


1

119891max1

minus 119891min1

)

2

+(


2

119891max2

minus119891min2

)

2

(15)






119877119901120574119901119863119901


120572 =

119895120574119901

119894120574119901

+ 119895120574119901

(16)









0




Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1

(a)


Rejected


Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1120574p = 3

S3

(b)












3


119897is the





119901of





119894maxminus1










Host1110

Host 9

70

Host1920


Host 710

Host1030

Host2560












119903





4 Evaluation






0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





119899)


119899) (119906 ≺


119894(V) and

exist119895 isin 1 2 3 119896 st 119891119895(119906) lt 119891




(1199091 1199092 119909


notexist1198831015840

isin Ω st1198831015840 ≺ 119883


isin Ω 1198831015840

≺

119883

Pareto Front PF = 119865(119883) = (1198911(119883) 119891

2(119883) 119891

119896(119883)) | 119883 isin










1198941 119909119896

1198942 119909

119896

119894119895 119909

119896

119894119863)






119894and119883119905



119894and

119883119905+1


119883119905+1

119894= 120572119883119905

119894+ (1 minus 120572)119883

119905

119895

119883119905+1

119895= (1 minus 120572)119883

119905

119894+ 120572119883119905

119895

(11)



1 1198832 119883

119896 119883


new individual 1198831015840 = (1198831 1198832 119883

1015840

119896 119883

119899) If the gene



[119880119896

min 119880119896


formula

1198831015840

119896=

119883119896+ (119880119896


120576

119896)

if 119903 and (0 1) isin [0 05)

119883119896minus (119883119896minus 119880119896

min) (1 minus 120574(1minus(119905119879))

120576

119896)

if 119903 and (0 1) isin [05 1]

(12)




120576

119896) ranges


119896and 120576









119901values



119901


Cuboid

f2

f1

i minus 1

Dp

i

i + 1







119901serves












119901value is


defined as follows

119877119901= 1 +

119905

sum

119895=1

120574119883119895

(13)


1198831 1205741198832 120574



1and


119901value of









119901values





f2

f1

A

B

rP = 1

rP = 2

rP = 3

rP = 4

RP = 2

RP = 5




119901


as follows

119894≺119877119901120574119901119863119901

119895 if (119894119877119901

lt 119895119877119901

)

or ((119894119877119901

= 119895119877119901

) (119894120574119901

lt 119895120574119901

))

or ((119894119877119901

= 119895119877119901

) (119894120574119901

= 119895120574119901

)

(119894119863119901

gt 119895119863119901

))

(14)







Monitor

MOGA-LS


Placement policy

Host m

Host 1

Host 2



Resource pool









119901value of each










0obtains its 119877

119901value






follows

119894119863119901

=radic(


1

119891max1

minus 119891min1

)

2

+(


2

119891max2

minus119891min2

)

2

(15)






119877119901120574119901119863119901


120572 =

119895120574119901

119894120574119901

+ 119895120574119901

(16)









0




Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1

(a)


Rejected


Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1120574p = 3

S3

(b)












3


119897is the





119901of





119894maxminus1










Host1110

Host 9

70

Host1920


Host 710

Host1030

Host2560












119903





4 Evaluation






0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




[119880119896

min 119880119896


formula

1198831015840

119896=

119883119896+ (119880119896


120576

119896)

if 119903 and (0 1) isin [0 05)

119883119896minus (119883119896minus 119880119896

min) (1 minus 120574(1minus(119905119879))

120576

119896)

if 119903 and (0 1) isin [05 1]

(12)




120576

119896) ranges


119896and 120576









119901values



119901


Cuboid

f2

f1

i minus 1

Dp

i

i + 1







119901serves












119901value is


defined as follows

119877119901= 1 +

119905

sum

119895=1

120574119883119895

(13)


1198831 1205741198832 120574



1and


119901value of









119901values





f2

f1

A

B

rP = 1

rP = 2

rP = 3

rP = 4

RP = 2

RP = 5




119901


as follows

119894≺119877119901120574119901119863119901

119895 if (119894119877119901

lt 119895119877119901

)

or ((119894119877119901

= 119895119877119901

) (119894120574119901

lt 119895120574119901

))

or ((119894119877119901

= 119895119877119901

) (119894120574119901

= 119895120574119901

)

(119894119863119901

gt 119895119863119901

))

(14)







Monitor

MOGA-LS


Placement policy

Host m

Host 1

Host 2



Resource pool









119901value of each










0obtains its 119877

119901value






follows

119894119863119901

=radic(


1

119891max1

minus 119891min1

)

2

+(


2

119891max2

minus119891min2

)

2

(15)






119877119901120574119901119863119901


120572 =

119895120574119901

119894120574119901

+ 119895120574119901

(16)









0




Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1

(a)


Rejected


Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1120574p = 3

S3

(b)












3


119897is the





119901of





119894maxminus1










Host1110

Host 9

70

Host1920


Host 710

Host1030

Host2560












119903





4 Evaluation






0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in













119901value is


defined as follows

119877119901= 1 +

119905

sum

119895=1

120574119883119895

(13)


1198831 1205741198832 120574



1and


119901value of









119901values





f2

f1

A

B

rP = 1

rP = 2

rP = 3

rP = 4

RP = 2

RP = 5




119901


as follows

119894≺119877119901120574119901119863119901

119895 if (119894119877119901

lt 119895119877119901

)

or ((119894119877119901

= 119895119877119901

) (119894120574119901

lt 119895120574119901

))

or ((119894119877119901

= 119895119877119901

) (119894120574119901

= 119895120574119901

)

(119894119863119901

gt 119895119863119901

))

(14)







Monitor

MOGA-LS


Placement policy

Host m

Host 1

Host 2



Resource pool









119901value of each










0obtains its 119877

119901value






follows

119894119863119901

=radic(


1

119891max1

minus 119891min1

)

2

+(


2

119891max2

minus119891min2

)

2

(15)






119877119901120574119901119863119901


120572 =

119895120574119901

119894120574119901

+ 119895120574119901

(16)









0




Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1

(a)


Rejected


Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1120574p = 3

S3

(b)












3


119897is the





119901of





119894maxminus1










Host1110

Host 9

70

Host1920


Host 710

Host1030

Host2560












119903





4 Evaluation






0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





Monitor

MOGA-LS


Placement policy

Host m

Host 1

Host 2



Resource pool









119901value of each










0obtains its 119877

119901value






follows

119894119863119901

=radic(


1

119891max1

minus 119891min1

)

2

+(


2

119891max2

minus119891min2

)

2

(15)






119877119901120574119901119863119901


120572 =

119895120574119901

119894120574119901

+ 119895120574119901

(16)









0




Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1

(a)


Rejected


Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1120574p = 3

S3

(b)












3


119897is the





119901of





119894maxminus1










Host1110

Host 9

70

Host1920


Host 710

Host1030

Host2560












119903





4 Evaluation






0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1

(a)


Rejected


Pt

Qt

Rt

120574p = 1

S1

120574p = 2

S2

Pt+1120574p = 3

S3

(b)












3


119897is the





119901of





119894maxminus1










Host1110

Host 9

70

Host1920


Host 710

Host1030

Host2560












119903





4 Evaluation






0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




Host1110

Host 9

70

Host1920


Host 710

Host1030

Host2560












119903





4 Evaluation






0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






119903





4 Evaluation






0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




0 1 2 4 530

50010001500200025003000350040004500500055006000

Pow

er co

nsum

ptio

n (K

Wh)

Time (weeks)

Random-migrationDLB

MOGA-LSStdPSO


010

015

020

025

030

035

040

045

050



Stan

dard

dev

iatio

n fo

r loa

d ba

lanc

ing






10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




10 20 30 40 50 60 70 80 900 100 1100

5

10

15

20

25

30

35

40

45


The n

umbe

r of f

ailu

res i

n V

M m

igra

tion

even

ts


StdPSO




0

5

10

15

20

25

30

35

40

45

20 40 60 80

The i

ncre

men

tal c

ost d

ue to

SLA

vio

latio

n (

)


StdPSOMOGA-LS










05

101520253035



Incr

ease

in co

st (

)







Acknowledgments



References





























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in









05

101520253035



Incr

ease

in co

st (

)







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 A Location Selection Policy of Live …downloads.hindawi.com/journals/tswj/2013/492615.pdfResearch Article A Location Selection Policy of Live Virtual Machine Migration