Optimized Speculative Execution to Improve Performance of MapReduce …downloads.hindawi.com/journals/mpe/2017/2724531.pdf · 2019-07-30 · Optimized Speculative Execution to Improve

Research ArticleOptimized Speculative Execution to Improve Performance ofMapReduce Jobs on Virtualized Computing Environment

Lei Yang1 Yu Dai2 and Bin Zhang1

1College of Computer Science and Engineering Northeastern University Shenyang China2College of Software Northeastern University Shenyang China

Correspondence should be addressed to Yu Dai daiyswcneueducn

Received 26 April 2017 Revised 19 September 2017 Accepted 8 October 2017 Published 7 December 2017

Academic Editor Mauro Gaggero

Copyright copy 2017 Lei Yang 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

Recently virtualization has become more and more important in the cloud computing to support efficient flexible resourceprovisioning However the performance interference among virtualmachinesmay affect the efficiency of the resource provisioningIn a virtualized environment where multiple MapReduce applications are deployed the performance interference can also affectthe performance of the Map and Reduce tasks resulting in the performance degradation of the MapReduce jobs Then in orderto ensure the performance of the MapReduce jobs a framework for scheduling the MapReduce jobs with the consideration of theperformance interference among the virtual machines is proposed The core of the framework is to identify the straggler tasks in ajob and back up these tasks to make the backed up one overtake the original tasks in order to reduce the overall response time ofthe jobThen to identify the straggler task this paper uses a method for predicting the performance interference degree Amethodfor scheduling the backing-up tasks is presented To verify the effectiveness of our framework a set of experiments are done Theexperiments show that the proposed framework has better performance in the virtual cluster compared with the current speculativeexecution framework

1 Introduction

Recently the MapReduce [1 2] as a platform for massive dataanalysis has been widely adopted by most of companies forprocessing large body of data to correlate mine and extractvaluable features With the prevailing of the virtualized tech-niques the virtual clusters can provide much more flexiblemechanism for different applications sharing the commoncomputing resourcesThen currently lots ofMapReduce jobsare deployed in a virtual cluster However themodern virtualtechniques do not provide perfect performance isolationmechanism for example Xen [3] which may cause thevirtual machines to compete for the limited resource andresult in the performance interference among the virtualmachines Then how to ensure the performance of theMapReduce job in the virtual cluster becomes a key issue

Previous works focusing on the performance of theMapReduce job have indicated the performance degradationin the virtual clusters [4ndash7] Other researchers have found

that the performance interference [8ndash10] is one of the impor-tant factors causing such degradationThen a set of works inthe field of task scheduling were conducted [11ndash13] to ensurethe performance of the MapReduce applications Howevermost of them only focus on IO intensive applications andtry to find a uniform performance interference model topredict the performance degradation for different types ofthe applications In fact for different applications using auniform model to evaluate its performance may not alwayswork well

In this paper we present an optimized speculative execu-tion framework for MapReduce jobs which aims to improvethe performance of the jobs in the virtual clusters Thecontribution of the paper is as follows(1) In order to predict the performance degradation amethod for predicting the performance degree is proposedIn this method the linear regression model is used to reflectthe performance degree and the system workloads and a

HindawiMathematical Problems in EngineeringVolume 2017 Article ID 2724531 11 pageshttpsdoiorg10115520172724531

2 Mathematical Problems in Engineering

swarm particle algorithm is used for finding the coefficientsin the model(2) In order to find the stragglers the method forcomputing the remaining time of the task is presented withthe consideration of the performance interference degree(3) In order to back up the stragglers a schedulingalgorithm is proposed which assigns the tasks to the slot witha global optimization

The organization of the rest of the paper is as followsThenext part introduces the current works related to the MapRe-duce scheduling in the virtual cluster Section 3 overviewsour speculative execution framework Sections 4 and 5 showhow to predict the performance interference degree identifythe stragglers and schedule the tasks Section 6 presents theexperimental result to verify our methods Finally the paperis summarized in Section 7

2 Related Works

Currently lots of works in the field of performance analysisin the virtual cluster are conducted Reference [14] presents amethod for predicting the interthread cache conflicts basedon the hardware activity vector Reference [15] presentsa method to characterize the application performance inorder to predict the overheads caused by the virtualizationReference [16] uses an artificial neural network to predict theapplication performance References [17 18] analyze the net-work IO contention in the cloud environment Performanceinterference among the CPU-intensive applications has beendiscussed in [11] Reference [12] considers the performanceinterference of the disk IO intensive applications and pro-poses a model for predicting such interference Reference [8]analyzes the factors related to the performance interferenceand presents amethod for estimating it Reference [19] targetsthe problem of application scheduling in data centers withthe consideration of the heterogeneity and the interferenceAlthough some of the current works have noticed theperformance interference and the MapReduce applicationsrsquoperformance caused by such interference they only focuson IO intensive applications and try to find a uniformperformance interference model to evaluate the performancedegradation for different types of the applications In factfor different applications using a uniform model to evaluateits performance may not always work well as the resourceusage pattern can be very different Besides the methodproposed in [19] develops several microbenchmarks to deriveinterference sensitivity scores and uses a collaborative filter-ing method to induce the sensitivity score for a new arrivalapplication which needs the application to run against at least2 microbenchmarks for 1 minute to get its profile Then asthe method relies on the microbenchmarks for analyzing theinterference degree the diversity of the microbenchmarkswill affect the accuracy of the analysis Besides if the diversitynumber of the microbenchmarks is large the score matrixfor the collaborative filtering may be very sparse as the newapplication cannot run against many microbenchmarks for1 minute before inducing the interference sensitivity scoreThen the collaborative filtering method may not work well

as for the sparse matrix Although some methods have beenproposed to solve this problem the effect is not very goodMeanwhile in the field of the MapReduce job schedulingthe QoS may depend on not only the interference but alsothe factor of the data locality Then making the MapReducejob run against the microbenchmarks may not reflect itsactual performance and get its actual profile MapReducejob may need to read the data file remotely from themicrobenchmarks Then the runtime of the job under thissituation may be different from the runtime when the jobneed not read input data files remotely In this sense themethod proposed in [19] may not be used in the field ofMapReduce job scheduling

Many researchers have put their efforts in the field of taskscheduling inMapReduce Reference [20] proposes a capacityscheduler to guarantee the fairly share of the capacity of thecluster among different users To ensure the data locality [21]proposes a delay scheduler With this technique if the head-of-line job cannot launch a local task the scheduler can delayit and look at the subsequent jobWhen a job has been delayedfor more than the maximum delay time the scheduler willassign the jobrsquos nonlocal map tasks Reference [22] usesa linear regression method to model the relation betweenthe IO intensive applications Reference [23] uses nodestatus prediction to improve the data locality rate Reference[24] uses a matchmaking algorithm for scheduling not onlyconsidering the data locality but also respecting the clusterutilization Reference [25] introduces a Quincy schedulerto achieve data locality Several recent proposals such asresource-aware adaptive scheduling [26] and cost effectiveresource provisioning [27] have introduced resource-awarejob schedulers to the MapReduce framework Reference [28]mentions the problem of task assignment with the consid-eration of the data locality in cloud computing Reference[29] focuses on the scheduling with the consideration ofthe data locality to minimize the cost caused by accessingremote files Reference [30] proposes a scheduling algorithmto make the jobs meet the SLAs Reference [31] solves theproblem of job scheduling with the consideration of thefairness as well as the data locality Reference [19] proposesa method for application scheduling with the considerationof the interference and a greed algorithm is presented forfinding the optimal assignments However this method isonly for single application As for our problem we need tofind optimal assignments in each time interval for a set oftasks As stated above most of the current works assume theperfect performance isolation among virtualmachinesThenbased on such an assumption current works seldom considerthe performance interference As stated above some of theworks consider the performance interference for examplein [22] the scheduler optimizes the assignment with theconsideration of only one task or only one slot while it ishard to achieve the global optimization of minimizing theperformance interference For example when two slots arefree simultaneously and the first job in the wait queue hasthe acceptable interference degree with the two nodes inthis case one needs to determine which slot will be used toserve the job However current works do not highlight this

Mathematical Problems in Engineering 3

issue and in fact it needs to make a decision with a globaloptimization

As for the performance of the MapReduce in the het-erogeneous environment [32] presents a LATE methodto improve the performance of MapReduce applicationsthrough speculative execution Reference [33] proposes amethod for optimizing the speculative execution by con-sidering the computing power to optimize the method forestimating the remaining time Reference [34] proposes aschedulingmethod especially for the heterogeneous environ-ment This algorithm according to the historical executionprogress of the task dynamically estimates the execution timeto determine whether to start a backup task for the task withlow progress rate However the above literature does notconsider the factor of the performance interference amongvirtualized computing resource on the problem of identifyingthe stragglers when estimating the remaining time Besideswhen assigning the backup task to the slot current worksdo not consider the performance interference which maycause the future straggler again Besides this current workonly waits for the straggler without a prediction in order tomake the backup decision earlyThen the effectiveness of themethod may be affected also

For the limitations of the above works this paper pro-poses an optimized speculative execution framework forMapReduce jobs on the virtualized computing resourcesTheframework considers the interference Then an interferenceprediction is employed and according to the predictionthe framework will compute the remaining time of the taskto predict the stragglers and assign the backup task to anappropriate node

3 Framework Overview

Figure 1 shows the optimized speculative execution frame-work for MapReduce jobs This framework is mainly forthe MapReduce applications running in a virtual cluster Inthe cluster there are a set of physical servers We imaginethat each of the physical servers has the same virtualizedenvironment Each physical server can allocate its resource tomultiple virtual machines The virtual machine can host theapplicationThevirtual cluster serves theHadoop frameworkThe Hadoop framework has one master node and multipleslave nodes The master node is deployed on a dedicatedphysical host For each of the slave nodes it will be deployedon a VM In the master node there are 4 major componentsStraggler Identification Module Backup Module Heart BeatReceiver andPerformance InterferenceModelingampPredictionStraggler Identification Module is to compute the remainingtime of the task in order to identify the straggler BackupModule is to assign the straggler tasks to the slotsHeart BeatReceiver is to collect the running states of the servers andthe tasks by receiving the heart beat information from theslave nodes Performance Interference Modeling amp Predictionis to train or retrain the performance interference model forpredicting

In the Sections 4 and 5 the major components in ourframework will be discussed

4 Methods for PredictingPerformance Interference

41 Modeling the Performance Interference In a virtual clus-ter the application 119886119901119901 deployed on a virtual machine (VM)will consume the resource of this VM Due to the contentionof the limited shared resource the resource usage of theVMs consolidated on the same physical host may affectothersrsquo access to the shared resource Then the performancedegradation of the applications on the VMs may be causedTo mitigate such degradation one of the important issues isto predict the extent to which the applicationrsquos performanceis affected by the contention of the shared resource Bythis when the predicted result shows a bad degradation wecan place this application on the other VM to mitigate theperformance degradation In the following for simplicitythe ldquoforeground VMrdquo is used to signify the VM whichserves the application app to be deployed while the otherVMs consolidated with the ldquoforeground VMrdquo are called theldquobackground VMsrdquo

As stated above the contention of the shared resourcemay cause the performance interference of the VMs to beconsolidated on the same physical server Then the resourceusage pattern of the ldquobackground VMrdquo may affect the per-formance of the ldquoforeground VMrdquoWith the difference of theresource usage of the background VM the performance ofthe foreground one will be different That is to say the extentto which the foreground VMrsquos performance is affected by thebackground one is different Then the term ldquoperformanceinterference degreerdquo is used for signifying this extent

Definition 1 (performance interference degree) We use (1) toshow the performance interference degree

PID (FWBW)= Perf (FWBW) minus Perf (FWIdle)

Perf (FWIdle) (1)

where we use system-level workloads to reflect the resourceusage pattern of a VM The system-level workloads consid-ered in this paper are shown in Table 1 FW and BW are theworkloads of the foreground and background VMs respec-tivelyThe performance of the application on FWmay includeresponse time and throughputWe use Perf(FWBW) to sig-nify such performance when the background VMrsquos workloadis BW Here Idle is especially for the background VM whenno application has been deployed on it

Since the contention of the shared resource can causethe performance degradation the interference degree of theforeground VM will have a relation with the resource usagepattern of the backgroundVMWe also do some experimentsto show this relation as Tables 2 and 3 show

Tables 2 and 3 show that with the background VMserving different types of applications the response time ofthe foreground one is different Here when the backgroundVM serves different types of applications it means that theresource usage pattern of the background VM is differentwhich also causes the difference of the performance of the


VMresourcemonitor

Physical resource monitorPhysical resource monitorPhysical resource monitor

Physical hostVMresourcemonitor

resourcemonitor

VMPhysical host

resourcemonitor

VMresourcemonitor

VMPhysical host

resourcemonitor

VM

SlotSlot Slot Slot Slot Slot

Historical data ofperformance interference

Performanceinterference modeltraining

Performanceinterferencemodel

Matchmakingworkload patterns

Predictingperformanceinterference

Selected performance interference model

Performance interferencemodeling amp prediction

Heart beat receiver

VM status

Taskassignment

Masternode

Backup module

Performanceinterference degree

Task prole

Heart beatinformation

Slavenodes

Straggleridentication

Stragglers


Task prole

middot middot middot


middot middot middotmiddot middot middot

Figure 1 Optimized speculative execution framework for MapReduce jobs

Table 1 System-level workload considered in this paper

System-levelworkload Meaning

cpuutil Average CPU utilizationmemutil Average memory utilizationrps Average number of read operations per secondwps Average number of write operations per secondawait Average waiting time of the IO operationssvctm Average time spent for the request in the disk device

foreground VM For example with the background VMrsquossystem-levelworkloads varying the application cat has differ-ent response timeThen we use the system-level workloads toreflect the interference degree as (2) shows

PID (FWBW) = 1198860 + 1198861 times cpuutilBW + 1198862timesmemutilBW + 1198863 times rpsBW + 1198864times wpsBW + 1198865 times awaitBW + 1198866times svctmBW(2)

where 1198860 1198861 1198862 1198863 1198864 1198865 and 1198866 are coefficients

By using (2) the interference degree can be known ifthe coefficients are known Then we need to estimate thecoefficients Imagine that the estimated coefficients are 1198861015840011988610158401 11988610158402 11988610158403 11988610158404 11988610158405 and 11988610158406 Then according to (2) the modelfor estimating the performance interference degree can be asfollows

PID (FWBW) = 11988610158400 + 11988610158401 times cpuutilBW + 11988610158402timesmemutilBW + 11988610158403 times rpsBW + 11988610158404times wpsBW + 11988610158405 times awaitBW + 11988610158406times svctmBW

(3)

Then when the background VMrsquos workloads are fedinto the above equation we can estimate the performanceinterference degree To estimate the coefficients we need tocompute the error between the predicted interference degreeand the actual one according to the observed data record

Then the problem of finding the combination of thecoefficients can be mapped to a problem according to theset of observed data (pid1 cpuutil1 memeutil1 rps1 wps1await1 svctm1) (pid119899 cpuutil119899 memeutil119899 rps119899 wps119899


Table 2 Response time of the application with the idle domain

App cpuutil memutil rps wps await Svctm (s) Response time (s)Bizp2 091 0007 278099 0999 0775 0345 84265cat 0001 0044 335158 433111 137643 158 27801Super PI 098 0001 0245 1547 12222 939 99107Iozone 0264 0036 37025 39295 15199 993 108724Ccrypt 0762 0421 17277 16868 5162 213 216611Gzip 0912 0053 29572 1183 844 118 21895

Table 3 Response time of the application with the background VM varying

Bzip2 cat Super PI Iozone Ccrypt GzipBzip2 93887 14828 89787 146988 14282 144168cat 30996 58892 40762 40762 45104 50026Super PI 101141 99981 100892 12199 10521 10354Iozone 17533 20527 109756 19873 11026 11364Ccrypt 24503 29655 23350 31173 25791 25403Gzip 23075 39399 22709 40374 24510 24002

await119899 svctm119899) to make the overall error the minimumwhich can be seen in

Error = 119899sum119894=1

[pid119894 minus (11988610158400 + 11988610158401 times cpuutil119894 + 11988610158402 timesmemutil119894

+ 11988610158403 times rps119894 + 11988610158404 times wps119894 + 11988610158405 times await119894 + 11988610158406times svctm119894)]2

(4)

The above problem can be seen as a problem of findingthe optimal combination of the coefficients in order to makethe error between the predicted interference degree andthe actual one the minimum In this paper for solving theproblem efficiently we use a swarm particle algorithm

When using swarm particle algorithm to solve suchproblem the first task is to define the particle For thisproblem the particle 119894 in the swarm can be defined as 119901119894 =[1198860119894 1198861119894 1198862119894 1198863119894 1198866119894] Here 119886119895119894 signifies the location of theparticle 119894 in the direction 119895 The number of particles in aswarm is signified as119898The particle119901119894 will update its locationin the direction 119895 with a speed V119895119894 The particle will computethe speed according to the best location pBest the particleis experiencing and the best location 119892119861119890119904119905 the swarm isexperiencing The best location means the location whichis the closest one to the optimal solution which usually isexpressed as the fitness function As for our problem thefitness function should evaluate how the swarm is close to theoptimal solution Then according to formula (4) the fitnessfunction of a swarm can be defined as follows

fitness (119901119894) = 119899sumV=1

[pidV minus (1198860119894 + 1198861119894 times cpuutilV + 1198862119894timesmemutilV + 1198863119894 times rpsV + 1198864119894 times wpsV + 1198865119894 times awaitV

+ 1198866119894 times svctmV)]2 (5)

Then we can use formula (6) to update the speed of theparticle 119901119894 in the direction 119895 and compute the location of theparticle in the same direction as formula (7)

V119895119894 (119905 + 1) = 119908 times V119895119894 (119905) + 1198881 times 1199031 (119901119861119890119904119905119895119894 minus 119909119895119894 (119905))+ 1198882 times 1199032 (119892119861119890119904119905119895119894 minus 119909119895119894 (119905)) (6)

119909119895119894 (119905 + 1) = 119909119895119894 (119905) + V119895119894 (119905 + 1) (7)

where V119895119894(119905+1) signifies the speed in the direction 119895 in the (119896+1)th iterations 119909119895119894(119905+1) signifies the location in the direction119895 in the (119896 + 1)th iterations 1199031(119911) and 1199032(119911) are 2 functionswhich return a random number between 0 and 1 1198881 and 1198882 arethe constants and 119908 is the weight which can be computed asformula (8) according to [35] In our experiment the size ofthe swarm is 30 the iteration number is 1000 and 1198881 = 1198882 = 2

119908 = 119908max minus 119908max minus 119908min119905119905max (8)

where 119908max and 119908min are the maximum and minimumweights 119905 is the current iteration number and 119905max is themaximum iteration number

Then the PSO algorithm can find the optimal combina-tion of the coefficients of each attribute Algorithm 1 presentsthe detailed algorithm

The method which uses regression model for estimatingthe performance interference degree can work well whenthere are historical data for training the coefficients Howeveras for the problem of MapReduce job scheduling suchhistorical datamay not always be availableThis is because thenew arriving jobs may not have the historical data about therunning status togetherwith the consolidatedVM in the samephysical hostThen in this case the historical data for trainingmay not be available For this situation we will discuss thecorresponding method in the following


Procedure PSOInitialize particle 119894 by giving velocity and positionInitialize pbest and 119892119861119890119904119905for each particle 119894 do

compute pBest and 119892119861119890119904119905Update the speed and location of 119894 by pBest and 119892119861119890119904119905

end forWhilemaximum iterationsEnd procedure

Algorithm 1 PSO algorithm to find the optimal combination of the weights

42 Inferring the Performance Interference Degree For twoapplications if their resource usage patterns are similar withthe same background VM their extents of the performancedegradation may be similar Then when one of the appli-cations is new and little historical data can be used fortraining its performance interference degree model we canpredict its performance interference by looking at anotheronersquos model Based on this idea we will discuss our methodin the following

Imagine that the performance interference degreemodelscan be kept and stored Then all the models can be a set119867 =PID(FW1)PID(FW2) PID(FW119899) Here FW119894 ofeach item PID(FW119894) in 119867 is called the workload patternThen if we do not have enough historical data for trainingapplication 119860rsquos performance interference model we can usean available and appropriate model in119867 for prediction

Letwp be theworkload pattern of the virtualmachine vmTo find an appropriate equation in 119867 is to find the equationwhose workload pattern is the most similar to wp

Then in the following we will show how to compute thesimilarity degree

For comparing the similarities we will use an Euclideandistance For two VMs vm119894 and vm119895 the similarity degreebetween their workload patterns can be computed as follows

119889 (wp119894wp119895) = ((cpuuilvm119894 minus cpuutilvm119895)2cpuuilvm119894 times cpuutilvm119895

+ (memuilvm119894 minusmemutilvm119895)2memuilvm119894 timesmemutilvm119895

+ (wpsvm119894 minus wpsvm119895)2wpsvm119894 times wpsvm119895

+ (awaitvm119894 minus awaitvm119895)2awaitvm119894 times awaitvm119895

+ (svctmvm119894 minus svctmvm119895)2svctmvm119894 times svctmvm119895

)minus12

(9)

Then we can use (9) to find the workload patterns whichare similar to the workload pattern of the VM to be predictedIn this paper if the similarity is beyond the predefinedthreshold it means the two workload patterns are similar

Then for a workload pattern wp by comparing the similaritydegrees we may find multiple workload patterns satisfyingthe predefined threshold requirement Then we can use thefollowing equation to generate a combined equation By usingsuch combined equation we can estimate the performanceinterference degree for the VM which has no historical datafor training the model

PID (FWBW) = sum119894

( 119889119894sum119895 119889119895) times PID (FW119894BW) (10)

where for the VM which is used to predict the performanceFW is used for signifying its workload Imagine the workloadpatterns satisfying the threshold requirements form the set119877 PID(FW119894BW) is the interference model correspondingto the 119894th workload pattern in 119877 119889119894 is the similarity degreebetween FW and FW119894

Then by using the above methods the performanceinterference model can be generated By using the modelwe can estimate the performance interference degree of anapplication For a MapReduce job it may contain a set oftasks The resource usage patterns of these tasks are alwayssimilar [36] And there are also many research works forpredicting the resource demand of the MapReduce jobsThen using this information the performance interferencedegree between the tasks to be assigned (no matter whetherthe corresponding job is newly submitted or runs for a while)and the VMs on the candidate physical host can be predicted

5 Methods for Identifying Straggler andBacking-Up in Virtualized Environment

In our framework the task trackers will send the heartbeat information which includes the resource status ofthe VMs Taking the task profile the status of VMs andthe physical host as inputs the module of PerformanceInterference Modeling amp Prediction will return a value toevaluate the interferenceThen in every interval the StragglerIdentification Module will predict the remaining time of eachrunning task in the next time interval according to the heartbeat information from the slave node and the performanceinterference degree provided by the Performance InterferenceModelingamp PredictionThe backupmodulewill back up a newtask for the straggler by assigning a new slot to it


In the speculative execution the task which will finishfarthest into the future will be backed up since the backed uptask will have a greatest opportunity to overtake the originalone and reduce the overall response time of the jobThen thecore of identifying a straggler is to estimate whether the taskhas a bad progress rate that is to say compared with othertasks in a job it has a longer remaining time to be finishedThen in the following we will introduce how to estimate theremaining time of the task in order to identify the stragglers

Imagine we have a job 119895 = 1199051 1199052 119905119899which contains aset of tasks Then we will introduce how to find the stragglertasks in the job Imagine that the number of the allocatedmapslots for this job is 119904119898 and the number of the allocated reduceslots for this job is 119904119903 Imagine that the number of the maptasks in this job to be executed is 119899119898 and the number of theallocated reduce slots for this job to be executed is 119899119903 Theoverall remaining time of the job is a sum of the remainingtime of the map phase and the reduce phase The remainingtime of either the map phase or the reduce phase dependson the slowest task Then the remaining time of 119905119894 can becomputed as (5)

According to (5) 119905119898predict119894 is the predicted completiontime of the current running map task 119894 which can becomputed as (11)119905119903predict119894 is the predicted completion time of

the current running reduce task 119894 which can be computed as(12)119905119898119894 is the execution time of map task 119894 119905119903119894 is the executiontime of reduce task 119894 119905119898max and 119905119898avg are the maximum andaverage completion time respectively of all the map taskswhich have been executed completely and 119905119903max and 119905119903avg arethe maximum and average completion time respectively ofall the reduce tasks which have been executed completely

119905119898predict119894 = 119905119898119894 times PIDpredictslot(119894)

PIDavgslot(119894)

(11)


PIDavgslot(119894)

(12)

where slot(119894) is the function to return the slot where thetask 119894 is deployed on PIDpredict

slot(119894) is the predicted performanceinterference degree among the slot slot(119894) and the other slotsconsolidated on the same physical server in the next timeinterval andPIDavg

slot(119894) is the average performance interferencedegree among the slot slot(119894) and the other slots consolidatedon the same physical server in the last interval from thebeginning of the execution to the current time

119879 =

if 119899119898mod 119904119898 = 0 and 119899119903mod 119904119903 = 0119905119898avg times lfloor119899119898119904119898 rfloor +max max

119894119905119898predict119894 minus 119905119898119894 119905119898max + 119905119903avg times lfloor119899119903119904119903 rfloor +max max

119894119905119903predict119894 minus 119905119903119894 119905119903max

if 119899119898mod 119904119898 = 0 and 119899119903mod 119904119903 = 0119905119898avg times lfloor119899119898119904119898 rfloor +max

119894119905119898predict119894 minus 119905119898119894 + 119905119903avg times lfloor119899119903119904119903 rfloor +max

119894119905119903predict119894 minus 119905119903119894


119894119905119898predict119894 minus 119905119898119894 + 119905119903avg times lfloor119899119903119904119903 rfloor +max max



119894119905119898predict119894 minus 119905119898119894 119905119898max + 119905119903avg times lfloor119899119903119904119903 rfloor +max

119894119905119903predict119894 minus 119905119903119894

(13)

Then based on (13) the remaining time of the job canbe predicted If there exists a running task whose predictedcompletion time makes the remaining time bigger than therequired one this task will be the straggler

Then after identifying the stragglers a backup task for thestragglers needs to be initiated by assigning a slot for this taskSince from every time interval the Straggler IdentificationModule will predict the stragglers in the next time intervalthere may be a set of straggler tasks to be backed up Thisproblem can be seen as a problem of scheduling this setof tasks in a virtualized computing environment As theperformance interference is an important factor which mayaffect the execution of the tasks when scheduling the taskto a slot with high time interference degree with others thetask may become a new straggler in the future again which

may result in the bad performance of the job Then whendealing with the problem of how to back up the stragglersthe performance interference degree needs to be consideredalso Previous works [37] schedule the tasks to the slotif the predicted interference degree is not higher than apredefined threshold 119879 otherwise the task will wait for theavailable node with the required interference degree or willbe assigned to a slot when the task is waiting for a long timeIn these works the scheduler optimizes the assignment withthe consideration of only one task or only one slot while itis hard to achieve the global optimization of minimizing theperformance interference For example when two slots arefree simultaneously and the first task in thewait queue has theacceptable interference degree with the two nodes which slotis used to place the task in will affect the following assigning


Input the set SL of slots to be free in the next interval the queue 119876 of tasks to be assignedOutput assignment plan APBegin(1)While 119876 is not empty do(2) Begin(3) 119898119894119899 = 10000(4) For each slot in SL do(5) Begin(6) If sloticapacity gt= Qelement[i]demand then(7) PID = GetPID(Qelement[i] slotjBackground)(8) Ifmin gt PID then(9) begin(10) min = PID(11) AP candidate[i]=slotj(12) end(13) End(14) Ifmin lt threshold then(15) AP[i] = AP candidate[i](16) end(17) Return AP

End

Algorithm 2 Backing up the stragglers with a global optimization

plan That is to say a decision with a global optimizationneeds to be made

This paper presents a scheduling strategy with a globaloptimization as mentioned in Algorithm 2 In each intervalthe backupmodule will collect the status of the tasks runningin the slots and estimate which slots will be free in the nextinterval by computing the remaining time of the task Thenin each interval the backup module will assign a set of tasksto the set of free slots for the next interval with the globaloptimization of minimizing the performance interferencedegree of each task Optimally finding the solution to theabove problem is anNP-complete problemThen we proposea greedy algorithm for solving this problem with betterefficiency Firstly the algorithm will place the task on the slotwith least interference degree Then for the remaining slotsto be free in the next interval redo the first step until all theslots are assigned with a task

6 Simulation Results

We evaluate our framework in a 24-node virtual cluster Thecluster has 6 physical servers one is for the mast node Theconfiguration of each server is as follows the memory is4G disk amount is 250G and the version of CPU is i3 Oneach physical server 4 virtual machines are deployed EachVM is created using Xen hypervisor and has 4VCPU and1GBmemoryWe configured each virtual machine with 1 slotwhich can be a map slot or a reduce slot In the whole virtualcluster we allocate 16 map slots and 8 reduce slots

We evaluate the framework using 10 MapReduce appli-cations seen in Table 4 These applications are widely usedfor evaluating the performance of MapReduce frameworkin the previous research works [21 32 38 39] To verifythe effectiveness of our works the experiments will be

Table 4 Test Applications

NameMajorresourceused

Introduction

TeraSort IO Sort the input data into a total orderTeraGen IO Generate and write data into systemGrep IO Extract matching regular expressionWordCount IO Count words in the input filePiEst CPU Estimate PiBayes CPU Construct Bayes classifiersMatrix CPU Matrix add and multiplicationgzip mixed Compress text filesBzip2 mixed Compress text filespovray mixed A frame rendering tool for 3-D graphics

carried out for some comparisons between our scheduler andothermain competitors which also consider the performanceinterference in the scheduling

In this section we evaluate whether our method is effec-tive in estimating the interference degree We will compareit with the model discussed in previous works [12] whichuses a uniform model for evaluating all the applicationsIn our experiment the predicted and actual performanceinterference degrees are considered Figure 2 shows theprediction error for each type of jobs using different models

From Figure 2 we can see that the current method led toan average of 29 error rate while our method can achievethe average rate of 15 This is because our method trainsthe model with the consideration of no historical data aboutperformance interference while the current method relies


CPU intensive IO intensive Mixed

Current methodOur method

0

10

20

30

40

Pred

ictio

n er

ror (

)

Figure 2 Comparison of prediction errors

Actual remaining timePredicted remaining time

0

200

400

600

800

Rem

aini

ng ti

me

200 400 600 8000Time

Figure 3 Comparison of predicted remaining time and actual one

on establishing a uniform model to evaluate all the types ofapplications which will sacrifice the prediction accuracy

In the following part the experiments will be done toshow whether our method is effective in predicting theremaining time in every time interval

From Figure 3 we can see that the current method led toan average of 20 This is because our method considers theperformance interference in the estimation of the remainingtime while the current method in [32] only takes an averageprogress rate for the estimation

In the following the experiments will show the effec-tiveness of our method in speculative execution The per-formance of the backup module is also affected by the datalocality Then to emphasize the performance interferenceonly we conduct the experiment in an intranet environmentwhere when accessing the data it does not need to readthe data remotely which minimizes the effect caused by thedata locality as much as possible We select the applicationsof Matrix and TeraGen which need no input and we alsoselect the applications of TeraSort and Gzip which need toread data We set the numbers of map tasks in the Matrix

0

1

2

Matrix TeraGen TeraSort Gzip

Current speculative execution

Nor

mal

ized

com

plet

ion

Pure fair scheduler

Our speculative executionframework

Capacity scheduler withoutspeculative execution

time

Figure 4 Comparison of the normalized completion times underthe light workload of the background

01234567

Matrix TeraGen TeraSort GzipNor

mal

ized

com

plet

ion

time

Current speculative executionPure fair scheduler



Figure 5 Comparison of the normalized completion times underthe heavy workload of the background

job TeraGen job TeraSort job and Gzip job which are 1510 10 and 5 respectively Every 15 seconds a batch of jobswhich contains 3Matrix jobs 3 TeraGen jobs 5 TeraSort jobsand 2 Gzip jobs will be submitted in the virtual cluster Theaverage normalized completion time is used for evaluation Inour method we model the relation between the performanceinterference degree and the background workload Then inthe experiment we will show the effectiveness of our sched-uler under the different status of the background workloadWe will adjust the background workload in this way that welet different jobs run on the virtualized slave node in orderto adjust the cpu memory and other system load to simulatethe variations of the background workload Figures 4 and 5show the result when using different schedulers in the masternode

From Figures 4 and 5 when the workload of the back-ground is heavy for example with the high CPU and mem-ory utilization all the applications suffer the performancedegradation severely when using the FairScheduler [37] andCapacityScheduler [20] Even under the situation with thelight workload of the background the speculative executionhas the better performance than the FairScheduler andCapacityScheduler The reason is that speculative executioncan identify the stragglers and speed up the speed of the


application Besides our speculative execution outperformsthe current speculative execution This is because ours findsthe stragglers by prediction while the current one findsthem by waiting for the degradation Besides the backing-up module in our framework also considers the performanceinterference when assigning the slots which may reduce thefuture risk of the degradation caused by the performanceinterference However we also notice that when the back-ground workload is light the performance of the differentschedulers is not too different This is because with the lightbackground workload the application suffers not too badperformance as a result of the interference among virtualizedslave nodes However in reality maintaining a light back-ground workload is usually not an easy task especially withthe consideration of the cost of the hardware and the systemutilization

7 Conclusions

This paper presents an optimized speculative executionframework for MapReduce jobs which aims to improve theperformance of the jobs on the virtual cluster Firstly weanalyze the factors related to the performance degradationin the virtual cluster and present a method for modelinghow the factors affect the degradation Secondly we developan algorithm that works with the performance interferenceprediction to identify the stragglers and assign the tasks

In this work when predicting the remaining time of theMapReduce job only the performance interference factor isconsidered In fact there are other factors such as the faultratio of the physical server which can also affect the accuracyof estimating the remaining time Then in the future workswe will optimize our method in predicting the remainingtime of the MapReduce jobs

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

Thisworkwas supported in part by theNational KeyTechnol-ogy RampD Program of the Ministry of Science and Technol-ogy (2015BAH09F02 and 2015BAH47F03) National NaturalScience Foundation of China (60903008 and 61073062) andthe Fundamental Research Funds for the Central Universities(N130417002 and N130404011)

References

[1] J Dean and S Ghemawat ldquoMapReduce simplified data pro-cessing on large clustersrdquo in Proceedings of the Symposium onOperating SystemsDesign and Implementation pp 137ndash150 NewYork NY USA 2004

[2] B R Chang N T Nguyen B Vo andH Hsu ldquoAdvanced CloudComputing and Novel ApplicationsrdquoMathematical Problems inEngineering vol 2015 pp 1-2 2015

[3] ldquoXen Virtual Machine Monitorrdquo httpwwwxenorg

[4] S Ibrahim H Jin L Lu L Qi S Wu and X Shi ldquoEvaluatingMapReduce on Virtual Machines The Hadoop Caserdquo in Pro-ceedings of the International Conference on Cloud Computingvol 1-4 of Lecture Notes in Computer Science pp 519ndash528Springer Berlin Germany 2009

[5] B He S M Yang and Z Guo Y ldquoWave Computing in theCloudrdquo in Proceedings of the Usenix Workshop on Hot Topics inOperating Systems Monte Verita Switzerland 2009

[6] S Ibrahim H Jin L Lu S Wu B He and L Qi ldquoLEENLocalityfairness-aware key partitioning for MapReduce in thecloudrdquo in Proceedings of the 2nd IEEE International Conferenceon Cloud Computing Technology and Science CloudCom 2010pp 17ndash24 USA December 2010

[7] Z Peng D Cui J Zuo and W Lin ldquoResearch on CloudComputing Resources Provisioning Based on ReinforcementLearningrdquo Mathematical Problems in Engineering vol 2015Article ID 916418 2015

[8] Y Koh R Knauerhase P Brett M Bowman Z Wen andC Pu ldquoAn analysis of performance interference effects invirtual environmentsrdquo in Proceedings of the ISPASS 2007 IEEEInternational Symposium on Performance Analysis of Systemsand Software pp 200ndash209 USA April 2007

[9] S Ibrahim H Jin L Lu B He and S Wu ldquoAdaptive diskIO scheduling for MapReduce in virtualized environmentrdquoin Proceedings of the 40th International Conference on ParallelProcessing ICPP 2011 pp 335ndash344 Taiwan September 2011

[10] X Zhang E Tune R Hagmann R Jnagal V Gokhale and JWilkes ldquoCPI2 CPU performance isolation for shared computeclustersrdquo in Proceedings of the 8th ACMEuropean Conference onComputer Systems EuroSys 2013 pp 379ndash391 Czech RepublicApril 2013

[11] R Nathuji A Kansal and A Ghaffarkhah ldquoQ-clouds Manag-ing performance interference effects for QoS-aware cloudsrdquo inProceedings of the 5th ACM EuroSys Conference on ComputerSystems EuroSys 2010 pp 237ndash250 France April 2010

[12] R C Chiang and H H Huang ldquoTRACON Interference-aware scheduling for data-intensive applications in virtualizedenvironmentsrdquo in Proceedings of the 2011 International Confer-ence for High Performance Computing Networking Storage andAnalysis SC11 USA November 2011

[13] P Lama and X Zhou ldquoNINEPIN Non-invasive and energyefficient performance isolation in virtualized serversrdquo in Pro-ceedings of the 42nd Annual IEEEIFIP International Conferenceon Dependable Systems and Networks DSN 2012 USA June2012

[14] A Settle J Kihm A Janiszewski and D Connors ldquoArchitec-tural support for enhanced SMT job schedulingrdquo in Proceedingsof the Proceedings 13th International Conference on ParallelArchitecture andCompilation Techniques 2004 PACT 2004 pp63ndash73 Antibes Juan-les-Pins France

[15] T Wood L Cherkasova K Ozonat and P Shenoy ldquoProfilingand Modeling Resource Usage of Virtualized Applicationsrdquoin Middleware 2008 vol 5346 of Lecture Notes in ComputerScience pp 366ndash387 Springer Berlin Heidelberg Berlin Hei-delberg 2008

[16] S Kundu R Rangaswami K Dutta and M Zhao ldquoAppli-cation performance modeling in a virtualized environmentrdquoin Proceedings of the 2010 IEEE 16th International Symposiumon High Performance Computer Architecture (HPCA) pp 1ndash10Bangalore India January 2010

[17] Y Mei L Liu X Pu and S Sivathanu ldquoPerformance measure-ments and analysis of network IO applications in virtualized


cloudrdquo in Proceedings of the IEEE 3rd International Conferenceon Cloud Computing pp 59ndash66 Miami Fla USA July 2010

[18] X Pu L Liu Y Mei S Sivathanu Y Koh and C Pu ldquoUnder-standing performance interference of IO workload in virtu-alized cloud environmentsrdquo in Proceedings of the 3rd IEEEInternational Conference on Cloud Computing CLOUD 2010pp 51ndash58 USA July 2010

[19] C Delimitrou andC Kozyrakis ldquoParagon QoS-Aware schedul-ing for heterogeneous datacentersrdquoACMSIGPLANNotices vol48 no 4 pp 77ndash88 2013

[20] Yahoo inc Capacity scheduler 2011 httpdeveloperyahoocomblogshadoopposts201102capacity-scheduler

[21] M Zaharia D Borthakur J Sen Sarma K Elmeleegy SShenker and I Stoica ldquoDelay scheduling a simple techniquefor achieving locality and fairness in cluster schedulingrdquo inProceedings of the 5th ACM EuroSys Conference on ComputerSystems (EuroSys rsquo10) pp 265ndash278 April 2010

[22] X Bu J Rao and C Xu ldquoInterference and locality-aware taskscheduling for MapReduce applications in virtual clustersrdquo inProceedings of the the 22nd international symposium p 227 NewYork New York USA June 2013

[23] X Zhang Z Zhong S Feng B Tu and J Fan ldquoImprovingData Locality of MapReduce by scheduling in homogeneouscomputing environmentsrdquo in Proceedings of the 9th IEEEInternational Symposium on Parallel and Distributed Processingwith Applications ISPA 2011 pp 120ndash126 Republic of KoreaMay 2011

[24] C He Y Lu and D Swanson ldquoMatchmaking A new MapRe-duce scheduling techniquerdquo in Proceedings of the 2011 3rd IEEEInternational Conference on Cloud Computing Technology andScience CloudCom 2011 pp 40ndash47 Greece December 2011

[25] M Isard V Prabhakaran J Currey UWieder K Talwar andAGoldberg ldquoQuincy Fair scheduling for distributed computingclustersrdquo in Proceedings of the 22nd ACM SIGOPS Symposiumon Operating Systems Principles SOSPrsquo09 pp 261ndash276 USAOctober 2009

[26] J Polo C Castillo D Carrera et al ldquoResource-Aware AdaptiveScheduling for MapReduce Clustersrdquo in Middleware 2011 vol7049 of Lecture Notes in Computer Science pp 187ndash207 SpringerBerlin Heidelberg Berlin Heidelberg 2011

[27] B Palanisamy A Singh and L Liu ldquoCost-Effective ResourceProvisioning for MapReduce in a Cloudrdquo IEEE Transactions onParallel and Distributed Systems vol 26 no 5 pp 1265ndash12792015

[28] X Fu Y Cang X Zhu and S Deng ldquoScheduling method ofdata-intensive applications in cloud computing environmentsrdquoMathematical Problems in Engineering vol 2015 Article ID605439 2015

[29] X Ma X Fan J Liu H Jiang and K Peng ldquoVLocalityRevisiting Data Locality forMapReduce in Virtualized CloudsrdquoIEEE Network vol 31 no 1 pp 28ndash35 2017

[30] N Lim S Majumdar and P Ashwood-Smith ldquoMRCP-RM ATechnique for Resource Allocation and Scheduling of MapRe-duce Jobs with Deadlinesrdquo IEEE Transactions on Parallel andDistributed Systems vol 28 no 5 pp 1375ndash1389 2017

[31] S Tang B-S Lee and B He ldquoDynamicMR a dynamic slotallocation optimization framework for mapreduce clustersrdquoIEEE Transactions on Cloud Computing vol 2 no 3 pp 333ndash347 2014

[32] M Zaharia A Konwinski and A Joseph ldquoImproving mapre-duce performance in heterogeneous environmentsrdquo in Proceed-ings of the Usenix Symposium on Opearting Systems Design andImplementation pp 29ndash42 San Diego Ca USA 2008

[33] H Jung and H Nakazato ldquoDynamic scheduling for speculativeexecution to improve MapReduce performance in heteroge-neous environmentrdquo in Proceedings of the 2014 IEEE 34thInternational Conference on Distributed Computing SystemsWorkshops ICDCSW 2014 pp 119ndash124 Spain July 2014

[34] K Kc and K Anyanwu ldquoScheduling hadoop jobs to meet dead-linesrdquo in Proceedings of the 2nd IEEE International Conferenceon Cloud Computing Technology and Science CloudCom 2010pp 388ndash392 USA December 2010

[35] Y Shi and R C Eberhart ldquoFuzzy adaptive particle swarmoptimizationrdquo in Proceedings of the Congress on EvolutionaryComputation vol 1 pp 101ndash106 IEEE Seoul Republic of Korea2001

[36] A Ganapathi Y Chen A Fox R Katz and D PattersonldquoStatistics-driven workloadmodeling for the cloudrdquo in Proceed-ings of the 2010 IEEE 26th International Conference on DataEngineering Workshops ICDEW 2010 pp 87ndash92 USA March2010

[37] A Beloglazov and R Buyya ldquoOptimal online deterministicalgorithms and adaptive heuristics for energy and performanceefficient dynamic consolidation of virtual machines in Clouddata centersrdquo Concurrency and Computation Practice andExperience vol 24 no 13 pp 1397ndash1420 2012

[38] B Palanisamy A Singh L Liu and B Jain ldquoPurlieus Locality-aware resource allocation for mapreduce in a cloudrdquo in Pro-ceedings of the 2011 International Conference for High Perfor-mance Computing Networking Storage andAnalysis SC11 USANovember 2011

[39] G Ananthanarayanan S Agarwal S Kandula et al ldquoScarlettCopingwith skewed content popularity inMapReduce clustersrdquoin Proceedings of the 6th ACM EuroSys Conference on ComputerSystems EuroSys 2011 pp 287ndash300 Austria April 2011

Submit your manuscripts athttpswwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MathematicsJournal of


Mathematical Problems in Engineering

Hindawi Publishing Corporationhttpwwwhindawicom

Differential EquationsInternational Journal of

Volume 2014

Applied MathematicsJournal of


Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of


Mathematical PhysicsAdvances in

Complex AnalysisJournal of


OptimizationJournal of


CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of


Operations ResearchAdvances in

Journal of


Function Spaces

Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of Mathematics and Mathematical Sciences


The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Algebra

Discrete Dynamics in Nature and Society



Decision SciencesAdvances in

Journal of


Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Stochastic AnalysisInternational Journal of


swarm particle algorithm is used for finding the coefficientsin the model(2) In order to find the stragglers the method forcomputing the remaining time of the task is presented withthe consideration of the performance interference degree(3) In order to back up the stragglers a schedulingalgorithm is proposed which assigns the tasks to the slot witha global optimization

The organization of the rest of the paper is as followsThenext part introduces the current works related to the MapRe-duce scheduling in the virtual cluster Section 3 overviewsour speculative execution framework Sections 4 and 5 showhow to predict the performance interference degree identifythe stragglers and schedule the tasks Section 6 presents theexperimental result to verify our methods Finally the paperis summarized in Section 7

2 Related Works

Currently lots of works in the field of performance analysisin the virtual cluster are conducted Reference [14] presents amethod for predicting the interthread cache conflicts basedon the hardware activity vector Reference [15] presentsa method to characterize the application performance inorder to predict the overheads caused by the virtualizationReference [16] uses an artificial neural network to predict theapplication performance References [17 18] analyze the net-work IO contention in the cloud environment Performanceinterference among the CPU-intensive applications has beendiscussed in [11] Reference [12] considers the performanceinterference of the disk IO intensive applications and pro-poses a model for predicting such interference Reference [8]analyzes the factors related to the performance interferenceand presents amethod for estimating it Reference [19] targetsthe problem of application scheduling in data centers withthe consideration of the heterogeneity and the interferenceAlthough some of the current works have noticed theperformance interference and the MapReduce applicationsrsquoperformance caused by such interference they only focuson IO intensive applications and try to find a uniformperformance interference model to evaluate the performancedegradation for different types of the applications In factfor different applications using a uniform model to evaluateits performance may not always work well as the resourceusage pattern can be very different Besides the methodproposed in [19] develops several microbenchmarks to deriveinterference sensitivity scores and uses a collaborative filter-ing method to induce the sensitivity score for a new arrivalapplication which needs the application to run against at least2 microbenchmarks for 1 minute to get its profile Then asthe method relies on the microbenchmarks for analyzing theinterference degree the diversity of the microbenchmarkswill affect the accuracy of the analysis Besides if the diversitynumber of the microbenchmarks is large the score matrixfor the collaborative filtering may be very sparse as the newapplication cannot run against many microbenchmarks for1 minute before inducing the interference sensitivity scoreThen the collaborative filtering method may not work well

as for the sparse matrix Although some methods have beenproposed to solve this problem the effect is not very goodMeanwhile in the field of the MapReduce job schedulingthe QoS may depend on not only the interference but alsothe factor of the data locality Then making the MapReducejob run against the microbenchmarks may not reflect itsactual performance and get its actual profile MapReducejob may need to read the data file remotely from themicrobenchmarks Then the runtime of the job under thissituation may be different from the runtime when the jobneed not read input data files remotely In this sense themethod proposed in [19] may not be used in the field ofMapReduce job scheduling

Many researchers have put their efforts in the field of taskscheduling inMapReduce Reference [20] proposes a capacityscheduler to guarantee the fairly share of the capacity of thecluster among different users To ensure the data locality [21]proposes a delay scheduler With this technique if the head-of-line job cannot launch a local task the scheduler can delayit and look at the subsequent jobWhen a job has been delayedfor more than the maximum delay time the scheduler willassign the jobrsquos nonlocal map tasks Reference [22] usesa linear regression method to model the relation betweenthe IO intensive applications Reference [23] uses nodestatus prediction to improve the data locality rate Reference[24] uses a matchmaking algorithm for scheduling not onlyconsidering the data locality but also respecting the clusterutilization Reference [25] introduces a Quincy schedulerto achieve data locality Several recent proposals such asresource-aware adaptive scheduling [26] and cost effectiveresource provisioning [27] have introduced resource-awarejob schedulers to the MapReduce framework Reference [28]mentions the problem of task assignment with the consid-eration of the data locality in cloud computing Reference[29] focuses on the scheduling with the consideration ofthe data locality to minimize the cost caused by accessingremote files Reference [30] proposes a scheduling algorithmto make the jobs meet the SLAs Reference [31] solves theproblem of job scheduling with the consideration of thefairness as well as the data locality Reference [19] proposesa method for application scheduling with the considerationof the interference and a greed algorithm is presented forfinding the optimal assignments However this method isonly for single application As for our problem we need tofind optimal assignments in each time interval for a set oftasks As stated above most of the current works assume theperfect performance isolation among virtualmachinesThenbased on such an assumption current works seldom considerthe performance interference As stated above some of theworks consider the performance interference for examplein [22] the scheduler optimizes the assignment with theconsideration of only one task or only one slot while it ishard to achieve the global optimization of minimizing theperformance interference For example when two slots arefree simultaneously and the first job in the wait queue hasthe acceptable interference degree with the two nodes inthis case one needs to determine which slot will be used toserve the job However current works do not highlight this













Perf (FWIdle) (1)





VMresourcemonitor



resourcemonitor

VMPhysical host

resourcemonitor

VMresourcemonitor

VMPhysical host

resourcemonitor

VM









Heart beat receiver

VM status

Taskassignment

Masternode

Backup module


Task prole


Slavenodes


Stragglers


Task prole













(3)









Error = 119899sum119894=1



(4)



fitness (119901119894) = 119899sumV=1


+ 1198866119894 times svctmV)]2 (5)



119909119895119894 (119905 + 1) = 119909119895119894 (119905) + V119895119894 (119905 + 1) (7)





















)minus12

(9)















PIDavgslot(119894)

(11)


PIDavgslot(119894)

(12)




119879 =






119894119905119903predict119894 minus 119905119903119894






119894119905119903predict119894 minus 119905119903119894

(13)






End









Introduction








0

10

20

30

40

Pred

ictio

n er

ror (

)



0

200

400

600

800

Rem

aini

ng ti

me

200 400 600 8000Time






0

1

2



Nor

mal

ized

com

plet

ion

Pure fair scheduler



time


01234567


mal

ized

com

plet

ion

time









7 Conclusions





Acknowledgments


References

















































Volume 2014




Journal of











Journal of


Function Spaces






Algebra





Journal of
















Perf (FWIdle) (1)





VMresourcemonitor



resourcemonitor

VMPhysical host

resourcemonitor

VMresourcemonitor

VMPhysical host

resourcemonitor

VM









Heart beat receiver

VM status

Taskassignment

Masternode

Backup module


Task prole


Slavenodes


Stragglers


Task prole













(3)









Error = 119899sum119894=1



(4)



fitness (119901119894) = 119899sumV=1


+ 1198866119894 times svctmV)]2 (5)



119909119895119894 (119905 + 1) = 119909119895119894 (119905) + V119895119894 (119905 + 1) (7)





















)minus12

(9)















PIDavgslot(119894)

(11)


PIDavgslot(119894)

(12)




119879 =






119894119905119903predict119894 minus 119905119903119894






119894119905119903predict119894 minus 119905119903119894

(13)






End









Introduction








0

10

20

30

40

Pred

ictio

n er

ror (

)



0

200

400

600

800

Rem

aini

ng ti

me

200 400 600 8000Time






0

1

2



Nor

mal

ized

com

plet

ion

Pure fair scheduler



time


01234567


mal

ized

com

plet

ion

time









7 Conclusions





Acknowledgments


References

















































Volume 2014




Journal of











Journal of


Function Spaces






Algebra





Journal of





VMresourcemonitor



resourcemonitor

VMPhysical host

resourcemonitor

VMresourcemonitor

VMPhysical host

resourcemonitor

VM









Heart beat receiver

VM status

Taskassignment

Masternode

Backup module


Task prole


Slavenodes


Stragglers


Task prole













(3)









Error = 119899sum119894=1



(4)



fitness (119901119894) = 119899sumV=1


+ 1198866119894 times svctmV)]2 (5)



119909119895119894 (119905 + 1) = 119909119895119894 (119905) + V119895119894 (119905 + 1) (7)





















)minus12

(9)















PIDavgslot(119894)

(11)


PIDavgslot(119894)

(12)




119879 =






119894119905119903predict119894 minus 119905119903119894






119894119905119903predict119894 minus 119905119903119894

(13)






End









Introduction








0

10

20

30

40

Pred

ictio

n er

ror (

)



0

200

400

600

800

Rem

aini

ng ti

me

200 400 600 8000Time






0

1

2



Nor

mal

ized

com

plet

ion

Pure fair scheduler



time


01234567


mal

ized

com

plet

ion

time









7 Conclusions





Acknowledgments


References

















































Volume 2014




Journal of











Journal of


Function Spaces






Algebra





Journal of










Error = 119899sum119894=1



(4)



fitness (119901119894) = 119899sumV=1


+ 1198866119894 times svctmV)]2 (5)



119909119895119894 (119905 + 1) = 119909119895119894 (119905) + V119895119894 (119905 + 1) (7)





















)minus12

(9)















PIDavgslot(119894)

(11)


PIDavgslot(119894)

(12)




119879 =






119894119905119903predict119894 minus 119905119903119894






119894119905119903predict119894 minus 119905119903119894

(13)






End









Introduction








0

10

20

30

40

Pred

ictio

n er

ror (

)



0

200

400

600

800

Rem

aini

ng ti

me

200 400 600 8000Time






0

1

2



Nor

mal

ized

com

plet

ion

Pure fair scheduler



time


01234567


mal

ized

com

plet

ion

time









7 Conclusions





Acknowledgments


References

















































Volume 2014




Journal of











Journal of


Function Spaces






Algebra





Journal of



















)minus12

(9)















PIDavgslot(119894)

(11)


PIDavgslot(119894)

(12)




119879 =






119894119905119903predict119894 minus 119905119903119894






119894119905119903predict119894 minus 119905119903119894

(13)






End









Introduction








0

10

20

30

40

Pred

ictio

n er

ror (

)



0

200

400

600

800

Rem

aini

ng ti

me

200 400 600 8000Time






0

1

2



Nor

mal

ized

com

plet

ion

Pure fair scheduler



time


01234567


mal

ized

com

plet

ion

time









7 Conclusions





Acknowledgments


References

















































Volume 2014




Journal of











Journal of


Function Spaces






Algebra





Journal of










PIDavgslot(119894)

(11)


PIDavgslot(119894)

(12)




119879 =






119894119905119903predict119894 minus 119905119903119894






119894119905119903predict119894 minus 119905119903119894

(13)






End









Introduction








0

10

20

30

40

Pred

ictio

n er

ror (

)



0

200

400

600

800

Rem

aini

ng ti

me

200 400 600 8000Time






0

1

2



Nor

mal

ized

com

plet

ion

Pure fair scheduler



time


01234567


mal

ized

com

plet

ion

time









7 Conclusions





Acknowledgments


References

















































Volume 2014




Journal of











Journal of


Function Spaces






Algebra





Journal of






End









Introduction








0

10

20

30

40

Pred

ictio

n er

ror (

)



0

200

400

600

800

Rem

aini

ng ti

me

200 400 600 8000Time






0

1

2



Nor

mal

ized

com

plet

ion

Pure fair scheduler



time


01234567


mal

ized

com

plet

ion

time









7 Conclusions





Acknowledgments


References

















































Volume 2014




Journal of











Journal of


Function Spaces






Algebra





Journal of







0

10

20

30

40

Pred

ictio

n er

ror (

)



0

200

400

600

800

Rem

aini

ng ti

me

200 400 600 8000Time






0

1

2



Nor

mal

ized

com

plet

ion

Pure fair scheduler



time


01234567


mal

ized

com

plet

ion

time









7 Conclusions





Acknowledgments


References

















































Volume 2014




Journal of











Journal of


Function Spaces






Algebra





Journal of






7 Conclusions





Acknowledgments


References

















































Volume 2014




Journal of











Journal of


Function Spaces






Algebra





Journal of



































Volume 2014




Journal of











Journal of


Function Spaces






Algebra





Journal of











Volume 2014




Journal of











Journal of


Function Spaces






Algebra





Journal of




Documents

Optimized Speculative Execution to Improve Performance of MapReduce …downloads.hindawi.com/journals/mpe/2017/2724531.pdf · 2019-07-30 · Optimized Speculative Execution to Improve