Adaptive selection of dynamic VM consolidation algorithm using neural network for cloud resource management

h i g h l i g h t s • Studies on multiple dynamic VM consolidation algorithms on cloud environment. • Discusses about trade-off between energy and SLA violation due to VM migration. • Proposes adaptive system which chooses best algorithm based on
of 8
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  Future Generation Computer Systems 87 (2018) 35–42 Contents lists available at ScienceDirect Future Generation Computer Systems  journal homepage: Adaptive selection of dynamic VM consolidation algorithm usingneural network for cloud resource management  Joseph Nathanael Witanto a , Hyotaek Lim b, * , Mohammed Atiquzzaman c a Department of Ubiquitous IT, Dongseo University, 617-716 Busan, South Korea b Division of Computer Engineering, Dongseo University, 617-716 Busan, South Korea c School of Computer Science, University of Oklahoma, Norman, OK 73019, United States h i g h l i g h t s •  Studies on multiple dynamic VM consolidation algorithms on cloud environment. •  Discusses about trade-off between energy and SLA violation due to VM migration. •  Proposes adaptive system which chooses best algorithm based on provider’s priority. •  Evaluates performance of proposed system and individual VM migration algorithms. a r t i c l e i n f o  Article history: Received 16 October 2017Received in revised form 28 February 2018Accepted 23 April 2018Available online 3 May 2018 Keywords: Cloud computingInfrastructure as a serviceResource managementDynamic consolidationVirtual machine migrationNeural network a b s t r a c t Cloudresourcemanagementbecomesmoreimportantwiththeincreasingusageofcloudresources.Withvarious cloud options available, cloud provider may have different priority in managing the resourcethrough resource scheduling and provisioning. Dynamic VM (Virtual Machine) consolidation algorithmis one of the techniques which can be used to reduce energy consumption through VM migration.HigherVMmigrationmayleadtolowerenergyconsumptionandhigherSLAviolation.Althoughpreviousresearch has successfully decreased energy consumption and SLA violation, cloud providers may needto manage trade-offs between energy and SLA violation through availability of priority in the system.This paper proposes neural network-based adaptive selection of VM consolidation algorithms whichadaptively chooses appropriate algorithm according to cloud provider’s goal priority and environmentparameters. Dataset generation and performance evaluation using simulations on real-world PlanetLabVMs workload trace showed that adaptive selector produced better average performance score thanindependent methods on various evaluation priority. © 2018 Elsevier B.V. All rights reserved. 1. Introduction Cloud computing is the technology to enable provisioning of resources (hardware and software) over the Internet. The pay-as-you-go pricing gives opportunity for cloud users to eliminate up-front cost [1]. Cloud resource management is important because it affects performance, functionality, and cost of a cloud system [2]. The integration of Internet of Things and cloud computing (Cloudof Things) for developing smart applications [3] will increase thenumber of cloud computing usage and the importance of cloudresource management.There are two types of resources to be managed, physical re-sources (CPU, memory, storage, workstation, network elements, *  Corresponding author. E-mail addresses: (J.N. Witanto), (H. Lim), (M. Atiquzzaman). sensors)andlogicalresources(OperatingSystem,energy,networkbandwidth, information security, protocols, APIs, and network de-lays) [4]. Since cloud involves large number of shared resources affected by external events, cloud resource management requirescomplex policies for optimization. For each cloud delivery model(Infrastructure as a Service, Platform as a Service, and Software asa Service), the resource management strategy also differ [2]. Important issues in cloud resource management include re-source provisioning and resource scheduling [4]. Resource provi- sioning is the allocation of a cloud provider’s resources to a cus-tomer. Inefficiency of resource provisioning leads to either over-provisioning or underprovisioning problem [4]. Since in the IaaS (InfrastructureasaService)modelservicesaredeployedasVirtualMachine (VM), the resource provisioning problem is reduced tohow to place the VM [5]. © 2018 Elsevier B.V. All rights reserved.  36  J.N. Witanto et al. / Future Generation Computer Systems 87 (2018) 35–42 Resource scheduling in cloud computing assigns the preciseand accurate task to CPU, network, and storage. There are vari-ous evaluation parameters for resource scheduling used in recentstudies including utilization, energy consumption, and SLA (Ser-vice Level Agreement) violation [6]. Although there are variousevaluation parameters for a system, cloud provider may need tosetdifferentprioritydependonthecurrentsituation.Forexample,analysis on IaaS cloud storage consumer behavior in 2011 showsthatservicefeeandstabilityarecriticalfactorsforcloudcomputingadoption [7]. One of the way to increase resource utilization and reduceenergy consumption is through dynamic VM consolidation (doingVM live migration on underloaded hosts and switching idle hoststo sleep mode). The process includes detection of overload orunderloaded hosts, selection of VMs to migrate, and selection of hosts for migration destination [8]. Energy-aware allocation heuristics for data center resourcesutilizing static threshold for dynamic consolidation of VMs areproposed in [9]. The developed version, adaptive heuristics for dynamic VM consolidation based on past VM utilization was pro-posedin[8].VMworkloadpredictionsystembasedondeepneural network (Deep Belief Network) was proposed in [10]. Host over- load and underload detection utilizing long-term prediction andprobability distribution was proposed in [11]. Greedy-based VM consolidation approach using prediction of future PM utilizationis proposed in [12].Although the new algorithms have successfully decreased en-ergy consumption and SLA violation, generally the SLA violationincreases as the energy consumption is decreased (due to theincreasingnumberofVMmigrationsneeded).Eachcloudprovidermay require different priority between energy consumption andSLA violation.The objective of this work is to improve management per-formance for dynamic VM consolidation problems through man-aging trade-off between energy and SLA violation minimization.Different with previous works, our approach is not to create newalgorithm but incorporating several existing algorithms.Our approach is based on the observation that certain algo-rithms are better for energy minimization and certain algorithmsare better for SLA minimization. This approach will allow easy ad-ditionorremovalofalgorithmsinthesystem.Theproposedsysteminvolves training of neural networks to adaptively choose over-loading detection algorithm, underloading detection algorithm,and host selection algorithm. We evaluate the system by usingCloudSimtoolkitandtheworkloaddatafromPlanetLabVMs.Basedon the performance evaluation, the proposed system can adap-tively choose appropriate algorithms for dynamic VM consolida-tion based on system’s environment conditions which fulfill thecloud provider’s target priority.Cloud provider’s priority in this paper refers to how importantan objective metric is to the provider in the scale of 0 to 1. EnergyconsumptionandSLAviolationaremetricsusedforourimplemen-tation. Cloud provider may set high priority on SLA violation andlow priority on energy consumption for critical data centers andthe opposite for non-critical data centers. SLA violation is calcu-lated the same way as in [8] which is also described in Section 3. The metric contains percentage of time when active host haveexperienced 100% CPU utilization and performance degradation of VMs due to migration.The main contributions of this paper are as follows:1. Evaluation of some algorithms for dynamic VM consolida-tion problem.2. Proposal of neural network-based adaptive selector whichcan automatically choose dynamic VM consolidation algo-rithmbasedonenvironmentconditionandcloudprovider’spriority.3. Performance evaluation and analysis of proposed adaptiveselector using simulation on real-world workload traces.The remainder of the paper is organized as follows. Section 2begins with related work on dynamic VM consolidation. Section 3dealswithproposedmethodologyincludingneuralnetworkstruc-ture. Section 4 describes the performance evaluation through ex-perimental results and analysis. Finally we conclude with futuredirections in Section 5. 2. Related work Beloglazov and Buyya [9] proposed MM (Minimization of mi- grations) algorithm which selects the minimum number of VMsneeded to migrate from a host to lower the CPU utilization belowthe upper utilization threshold if the upper threshold is violated.If the CPU utilization of a host falls below the lower threshold, allVMs have to be migrated to another host in order to lower powerconsumption.ModifiedBestFitDecreasingAlgorithmwasusedfornew host selection. VMs are sorted in decreasing order of theircurrent CPU utilizations and allocated to host that provides theleast increase of power consumption due to new allocation [9].However, since the resource usage pattern in IaaS environmentsare not static, fixed utilization thresholds are not very efficient.The next research paper by Beloglazov and Buyya [8] proposedadaptive heuristics for dynamic VM consolidations based on his-torical analysis of resource usage. Based on the evaluation usingsimulation on PlanetLab VMs, local regression algorithm for over-loaded host detection combined with MMT (Minimum MigrationTime)forselectingVMstobemigratedoutperformsotherdynamicVM consolidation algorithms in regard to SLA violation level andnumberofVMmigrations.Underloadedhostwaschosenbyselect-ingactivehostwithminimumutilization[8].Althoughenergyand SLA violation is better compared to the previous research [9], thealgorithms show various result (high energy and no SLA violationto low energy and high SLA violation). Our proposed system adap-tively chooses algorithm according to cloud provider’s priority.Comparison of VM placement algorithms was conducted byMann and Szabo [13] to find optimal allocation of VMs in data centerwithheterogeneousPMsize.Intheexperiment,Guazzone’sheuristicalgorithm[14]andShi’sAbsoluteCapacityalgorithm[15] generally gave the best results with SLA violation and energyconsumptionasthemetrics.Intheexperiment,eachalgorithmhasitsownadvantages(e.g.Guazzone’salgorithmshowslowerenergyconsumption but higher SLA violation). Therefore, our proposedsystem may help to choose between the methods according to theneed. Some other related VM placement algorithms can be seen inTable 1.A deep learning prediction approach was proposed to predictthe workload of virtual machines [10]. Based on deep belief net- work(DBN)composedofmultiplelayersofRBM(RestrictedBoltz-mann Machine) and regression layer, the model received inputof all VM’s past history (dimension of number of VMs x numberof timesteps used). The output is prediction of VMs’ workload inthe future. However, the author did not try to incorporate theprediction with other parts of VM consolidation algorithm. Ourproposed system may be used to test parts of VM consolidationalgorithms and combine it with other parts (e.g. VM placementalgorithm) to choose the best combination.Overloaded and underloaded host detection for VM live migra-tion using long-term utilization prediction and probability distri-bution was proposed in [11]. The action is determined by compar-ing current utilization and next timesteps’ utilization (using Gaus-sian Processes) with the upper and lower threshold in probabilis-tic manner. Greedy-based VM consolidation approach proposedin[12]formulatestheVMconsolidationasamulti-objectivevector   J.N. Witanto et al. / Future Generation Computer Systems 87 (2018) 35–42  37  Table 1 VM placement algorithms.Author Method GoalAldhalaan et al. [16] Heuristic algorithm to find optimal VM allocation based on communication strength between VMs.Optimize cloud provider’s revenue.Nejad et al. [17] Truthful greedy mechanisms for dynamic virtual machine provisioning and allocation problem for auction-based cloud computing.Optimize cloud provider’s revenue.Xu et al. [18] Heuristic (improved multi-objective particle swarm optimization) virtual machine placement.Optimize resource utilization and reduce turnaround times due tomigration.Bartok et al. [19] Branch-and-bound approach (using search tree) that exploits problem-specific knowledge for virtual machine placement.Minimize cost from energy consumption and migration overhead.Verma et al. [20] Predict resource demand for service tenants which is expected to increase and allocate VMs using best-fit heuristic approach.Optimize prediction time and host’s resource usage.Ahvar et al. [21] Use A* algorithm and Fuzzy Sets to find optimal VM placement based on geographically varying energy prices and carbon emission rates forresourcesOptimize overall cost and carbon emission in distributed clouds.Li et al. [22] Multiple objective optimization framework for each VM resource across hosts and different resources within same host.Achieve load balance of multiple resources.Lin et al. [23] Use mathematical modeling of peak similarity to measure similarity of  VMs’ peak workload and avoid VMs with high correlation for betterVM placement and consolidation.Optimize resource utilization. bin packing problem in order to reduce energy consumption, min-imize VM migrations, and avoid SLA violations. VMs are migratedfromPMsthatarecurrentlyoverloadedorwillbeoverloadedinthefuture. New PMs are selected based on the current and future re-source utilization of PM and VM in order to avoid unnecessary VMmigrations.AlthoughtheenergyandSLAviolationfor[11]and[12] arebetterthanpreviousworks,theSLAviolationisstillhigherthanno-migration policy. Our proposed system can incorporate newalgorithms to manage the trade-off needed by cloud provider.In complement to previous works, we propose a neural-network based adaptive selector to choose the algorithms to be usedfor dynamic VM consolidation at a particular time. The purposeis to manage trade-off of algorithms in terms of energy and SLAviolation minimization. Different with previous works which domanagement independently, we incorporate several algorithms.The system adapts according to the environment parameters, andthe cloud provider’s priority on energy usage and SLA violationminimization.IdeatousemultipleVMconsolidationalgorithmsthatofferdif-ferent trade-offs between speed and quality was proposed in [24].The system used threshold parameter of current environment todecide whether to choose fast heuristic algorithm or slow (sup-posedlyhavebetterperformance)integerlinearprogram.Differentwith the previous work, we proposed end-to-end neural networkto decide consolidation algorithm based on provider’s priorityand environment variable while previous work used thresholdparameter to manage trade-off between optimization quality andreaction time. 3. Proposed method As stated in Section 2, an adaptive VM consolidation algorithm is proposed in this paper. This section will cover the mechanismof VM consolidation and proposed method. System design willinclude adaptive selector structure, detailed process of datasetcreation, adaptive selector training, and evaluation. 3.1. System model The system model inn Fig. 1 shows the relationship betweenalgorithmadaptiveselectorandotherentities.Thesystemreceivesresource requirement from cloud client which is managed withresource provisioning algorithm which also manages dynamic VMconsolidation. Based on the monitored resource utilization sum-mary and provider’s evaluation priority (described further in 3.2),method for VM Consolidation will be periodically chosen to theresource provisioning algorithm. Fig. 1.  Proposed system model. 3.2. Adaptive selector  Fig. 2 shows the structure of adaptive selector. Neural networkwithinput,hidden,andoutputlayersareused.Someenvironmentparameters (from resource utilization summary) and evaluationparameters (from provider) are used in the input layer to selectdynamic VM consolidation algorithm combinations. The environ-ment parameters used are current active host percentage, averageresource utilization, and workload distribution balance.Activehostpercentageiscalculatedbydividingactivehostdur-ingtheselectionprocesswithtotalhostsavailable.Host’sresourceutilization and workload distribution imbalance are calculated  38  J.N. Witanto et al. / Future Generation Computer Systems 87 (2018) 35–42 Fig. 2.  Adaptive selector (Neural network). fromaverageandstandarddeviationCPUutilizationofactivehoststhroughout the time.Total monetary cost in a data center can be calculated as sumof fees to external cloud providers, SLA violation fees, and cost of consumed power [25]. Assuming optimization model for indepen- dentcloudprovider,theevaluationparameters(provider’sprioritywithrangeof0.0–1.0)beingusedareenergyconsumptionandSLAviolation.SLA violation for evaluation is the same as in [8]. The SLA Violation or SLAV (Eq. (1)) is calculated using two metrics:1. SLA violation Time per Active Host (SLATAH): Percentage of time when active host have experienced 100% CPU utiliza-tion (Eq. (2)).2. Performance Degradation due to Migrations (PDM) OverallperformancedegradationofVMsduetomigrations(Eq.(3)). SLAV   = SLATAH  × PDM   (1) SLATAH   = 1 N  N  ∑ i = 1 T  si Tai (2) PDM   = 1 M  M  ∑  j = 1 C  dj Crj (3)where  N   is the number of hosts;  T  si  is the total time during whichthehost i hasexperiencedSLA(100%CPUutilization); T  ai  isthetotalactive time of host  i ;  M   is the number of VMs;  C  dj  is the estimateofVM  j performancedegradation(10%ofCPUutilizationduringallmigrations of VM  j );  C  rj  is the total CPU capacity requested by VM  j  [8].According to [9], problem of dynamic VM consolidation can be split into:1. Detection of overloaded host detection, selection of VM inoverloaded host, and placement of selected VMs in newhosts to avoid performance degradation.2. Detection of underloaded host and placement of all VMs innew hosts to conserve energy.Theoutputclassesofadaptiveselectorarecombinationofsomedynamic VM consolidation algorithms described in Section 2. No- migration and local regression [8] are used for overloaded host detection.No-migrationandminimumCPUutilization[8]areused forunderloadedhostdetection.Guazzone’s[14]andShi’sAbsolute Capacity [15] algorithms are used for host selection and VM place-ment. For VM selection algorithm, Minimum Migration Time [8] is used. Fig. 3.  Proposed system process. 3.3. System process Fig. 3 shows the system process for the adaptive selector. Exe-cution steps of the system are as follows:1. Generaterawdatasetbysimulatingthemethodsforseveraltimesteps. Workload data that was used in evaluation canbeseenin4.1.Eachrowwillcontaintheinitialenvironment parameters and normalized evaluation result of all meth-ods. The result (energy and SLA violation) for each row arenormalized by converting the values from worst-best to therange of 0.0–1.0 (Eq. (4)). Lower energy and SLA violationresults are better. nresult  ij  = 1 − result  ij − minresult   j maxresult   j − minresult   j (4)where  i  indicates the method’s index;  j  indicates the pa-rameter’s index (energy or SLA violation);  nresult  ij  is thenormalized value of evaluation result  result  ij ;  minresult   j  and maxresult   j  are the minimum and maximum evaluation re-sult of   j  parameter across all methods at a specific row.2. Generate full dataset. Full dataset requires environmentparameters, evaluation parameters’ priority, and chosenmethod (neural network training requires data of input andoutput layer) as illustrated in Fig. 2. To create training and testingdataset,randomevaluationprioritybetween0.0–1.0willbegenerated.Method’sperformancescoreiscalculatedwith the sum of multiplication between evaluation priorityandnormalizedevaluationresultfromrawdataset(Eq.(5)). score i  = E  ∑  j = 1  priority  j × nresult  ij  (5)where  score i  is performance score of method  i ;  E   is numberof evaluation parameters;  priority  j  is priority value of eval-uation parameter  j ;  nresult  ij  is the normalized evaluationresult from previous step. Random evaluation priority willbeusedininputlayerandmethodwithhighestperformancescore will be used as the output class of neural network.   J.N. Witanto et al. / Future Generation Computer Systems 87 (2018) 35–42  39  Table 2 Planetlab workload.Date NN dataset generation (sample first 5 h) Performance score evaluation03/03/2011 yes yes06/03/2011 yes yes09/03/2011 yes yes22/03/2011 yes yes25/03/2011 yes yes03/04/2011 no yes09/04/2011 no yes11/04/2011 no yes12/04/2011 no yes20/04/2011 no yes 3. Train neural network using training dataset.4. Evaluate adaptive selector with trained neural network weight:(a) Neural network accuracy on training and testing dataset.(b) Performance score comparison with srcinal methods(without adaptive selector) on various evaluation pri-ority. 4. Performance evaluation In this section, we describe conducted experiment of our pro-posed system, including libraries and dataset that were used forthe experiment. The experiment, result, and analysis are shown toevaluate feasibility of proposed system compared to independent,non-adaptive methods. 4.1. Implementation CloudSim toolkit [26] was used to simulate the VM consolida-tion algorithms. The experiment setup is similar to Beloglasov andBuyya’s experiment [8] including the usage of PlanetLab VMs [27]. PlanetLab data is provided as a part of the CoMon Project andcontains the CPU utilization of more than a thousand VMs fromservers located at more than 500 places around the world. Theinterval of utilization measurements is 5 min.Instead of Modified Best Fit Decreasing Algorithm [8] which chooses host with minimum energy consumption directly, Guaz-zone’s [14] and Shi’s Absolute Capacity [15] algorithms were used. DeepLearning4j library is used for running neural network inadaptive selector. One hidden layer with 10 nodes is used. Xavierweight initialization [28] and ReLU activation function [29] are used.In total there are four output classes of the adaptive selector:1. No migration2. Migration only for underloading3. Migration for underloading and overloading (local regres-sion for detection and Guazzone for VM placement)4. Migration for underloading and overloading (local regres-sion for detection and Shi-AC for VM placement).To generate dataset for neural network training and testing,CloudSim is used to simulate different methods (algorithm com-binations) on PlanetLab workload. PlanetLab workload consists of workload at 10 different days and first 5 days are used to generatedataset as described in Table 2. Dataset will be split into training and testing dataset. After neural network training and testing, allworkloads (10 days) will be used to compare performance scorewith individual methods.For these 5 workloads, the first 5 h are used for NN datasetgeneration. Every simulation hour, raw dataset is created (cor-responds to Fig. 3 step 1). Since all combinations are generated,  Table 3 VM consolidation algorithms.Method DescriptionMethod 1 No migrationMethod 2 Underload-only migrationMethod 3 Local regression — GuazzoneMethod 4 Local regression — Shi ACAdaptive Proposed adaptive selector the raw dataset contains 341 records per each day (1 record from1st hour, 4 records from 2nd hour, 16 records from 3rd hour,64 records from 4th hour, and 256 records from 5th hour). Foreach record 250 combinations of random evaluation priority areused and 90% of them are randomly chosen for training dataset.Therefore, training dataset contains 383,652 records and testingdataset contains 42,598 records.The training dataset is used to train neural network usingDeepLearning4j library. To tackle imbalance of dataset (certainoutput classes have more record than others) weighted arrays forcalculating loss functions are used. Training is done using batchsize of 100 for 1 epoch. 4.2. Result and analysis DeepLearning4J library is used to train and test neural net-work to predict method with best performance (first rank) pertimestep(onehourinterval)usinggenerateddataset.Theaccuracyis calculated by dividing the correct number of predictions bytotalnumberofpredictions.Accuracyofneuralnetworktopredictmethod with best performance (first rank) per timestep is 74.92%on training dataset and 75.05% on testing dataset.The next results are evaluation results of adaptive selectorusing CloudSim in 10 workloads using various evaluation prioritybetween0.0and1.0.Highenergypriority(pE)meanstheexpectedenergy result is low. High SLA priority (pSLA) means the expectedSLA violation is low.Table 3 shows methods that will be tested independently andourproposedadaptivealgorithm.Thecalculationofmethod’sper-formance score is as described in step 2 of Section 3.3 (sum of multiplicationbetweenpriorityandnormalizedevaluationresult).Fig.4showsaverageenergyutilizationperdayforeachmethod.Method 1 (with no migration) uses much more energy than othermethods.Adaptive(pE > pSLA)comesatsecondplacewithslightlyworseperformancethanmethod2(lowestenergyusage).Adaptive(pE ≤ pSLA) uses higher energy than method 2–4, but still muchlower than method 1.Fig. 5 shows average SLA violation per day for each method.Method 1(with no migration) produces no SLA violation. Adaptive(pE  ≤  pSLA) shows better performance than other migration-enabled methods. Adaptive (pE  >  pSLA) shows slightly betterperformance than method 2 (highest SLA violation).Table 4 compares performance of each method in terms of en-ergy and SLA violation. Whereas method 1–2 has extreme results
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks

We need your sign to support Project to invent "SMART AND CONTROLLABLE REFLECTIVE BALLOONS" to cover the Sun and Save Our Earth.

More details...

Sign Now!

We are very appreciated for your Prompt Action!