Job Scheduling Strategies for Parallel Processing: 15th International Workshop, JSSPP 2010, Atlanta, GA, USA, April 23, 2010, Revised Selected Papers PDF
Preview Job Scheduling Strategies for Parallel Processing: 15th International Workshop, JSSPP 2010, Atlanta, GA, USA, April 23, 2010, Revised Selected Papers
Lecture Notes in Computer Science 6253 CommencedPublicationin1973 FoundingandFormerSeriesEditors: GerhardGoos,JurisHartmanis,andJanvanLeeuwen EditorialBoard DavidHutchison LancasterUniversity,UK TakeoKanade CarnegieMellonUniversity,Pittsburgh,PA,USA JosefKittler UniversityofSurrey,Guildford,UK JonM.Kleinberg CornellUniversity,Ithaca,NY,USA AlfredKobsa UniversityofCalifornia,Irvine,CA,USA FriedemannMattern ETHZurich,Switzerland JohnC.Mitchell StanfordUniversity,CA,USA MoniNaor WeizmannInstituteofScience,Rehovot,Israel OscarNierstrasz UniversityofBern,Switzerland C.PanduRangan IndianInstituteofTechnology,Madras,India BernhardSteffen TUDortmundUniversity,Germany MadhuSudan MicrosoftResearch,Cambridge,MA,USA DemetriTerzopoulos UniversityofCalifornia,LosAngeles,CA,USA DougTygar UniversityofCalifornia,Berkeley,CA,USA GerhardWeikum MaxPlanckInstituteforInformatics,Saarbruecken,Germany Eitan Frachtenberg Uwe Schwiegelshohn (Eds.) Job Scheduling Strategies for Parallel Processing 15th International Workshop, JSSPP 2010 Atlanta, GA, USA, April 23, 2010 Revised Selected Papers 1 3 VolumeEditors EitanFrachtenberg Facebook,475BrannanSt. SanFrancisco,CA,94107,USA E-mail:[email protected] UweSchwiegelshohn RoboticsResearchInstitute SectionInformationTechnology TUDortmundUniversity Otto-Hahn-Str.8 44227Dortmund,Germany E-mail:[email protected] LibraryofCongressControlNumber:2010937281 CRSubjectClassification(1998):C.2.4,D.4,C.2,D.2,H.3,I.6 LNCSSublibrary:SL1–TheoreticalComputerScienceandGeneralIssues ISSN 0302-9743 ISBN-10 3-642-16504-4SpringerBerlinHeidelbergNewYork ISBN-13 978-3-642-16504-7SpringerBerlinHeidelbergNewYork Thisworkissubjecttocopyright.Allrightsarereserved,whetherthewholeorpartofthematerialis concerned,specificallytherightsoftranslation,reprinting,re-useofillustrations,recitation,broadcasting, reproductiononmicrofilmsorinanyotherway,andstorageindatabanks.Duplicationofthispublication orpartsthereofispermittedonlyundertheprovisionsoftheGermanCopyrightLawofSeptember9,1965, initscurrentversion,andpermissionforusemustalwaysbeobtainedfromSpringer.Violationsareliable toprosecutionundertheGermanCopyrightLaw. springer.com ©Springer-VerlagBerlinHeidelberg2010 PrintedinGermany Typesetting:Camera-readybyauthor,dataconversionbyScientificPublishingServices,Chennai,India Printedonacid-freepaper 06/3180 Preface Thisvolumecontainsthepaperspresentedatthe15thworkshoponJobSchedul- ing Strategies for Parallel Processing that was held in Atlanta (GA), USA, on April 23, 2010 in conjunction with the IEEE International Parallel Processing Symposium 2010. This year 18 papers were submitted to the workshop. All submitted papers went through a complete review process, with the full version being read and evaluated by an average of four reviewers. We would like to especially thank the programcommittee members and additionalreferees for their willingness to participate in this effort and their excellent, detailed reviews: Henri Casanova, Peter A. Chronz, Walfredo Cirne, Julita Corbalan, Arash Deshmeh,DickEpema,DrorG.Feitelson,AllanGottlieb,RajkumarKettimuthu, Virginia Lo, Kuan Lu, Vicent Matossian, Jose E. Moreira,Bill Nitzberg, Elizeu Santos-Neto,AngelaC.Sodan,MarkS.Squillante,DanTsafrir,PhilippWieder, and Ramin Yahyapour. Thepapersinthis volumeshowaprolificgrowthinthe areasofapplicability forparallelscheduling.Togetherwiththemorecommonschedulingaspects(such asclusterandGridscheduling,workloadanalysis,metrics,qualityofservice,and task scheduling), these papers increasingly discuss more recent problems and applications, such as virtualized environments, many-core processors, DNA se- quencing,andHadoop.This volume alsoincludes a paper thatsummarizes Dan Tsafrir’sworkonunderstandingtheroleofuserestimatesinjobschedulingeval- uations.Hisinsights,whichwerepresentedinthis workshop’skeynote,arequite instructive and lead to the conclusion that accurate user estimates are indeed betterforefficientscheduling.Althoughthisconclusionmaysoundintuitive,itis actuallycontradictorytopreviousstudiesthatfoundinaccurateestimatestoim- proveschedulerperformance.Followinghis analysis,Danalsosuggestspractical ways to deal with estimate inaccuracy for realistic job scheduler evaluations. One of the stated goals of the JSSPP workshop is to explore the applica- tion of traditional scheduling topics to novel scenarios. A good example of such topics is task and graph scheduling, which until last year was largely outside the scope of JSSPP. Recent technologies are reviving interest in this topic, as was discussed last year in the context of workflow jobs in a Grid environment (in a paper by Gong et al.), and this year in the context of the emerging multi- core/multi-threadedprocessors.Xia,Prasanna,andLiapplyideasfromdynamic load balancing and hierarchical thread grouping to the contemporary architec- ture of the Sun Niagara processor with its 64 hardware threads. A different thread-level scheduling aspect for many-core architectures is introduced in the paper by Yamada and Kusakabe. Here, a modification to the operating sys- tem is suggested to permit multi-threaded applications to aggregate time and space resources for improved cache efficiency. In another example of applying VI Preface traditional scheduling to newer problems, the paper by Saule, Bozdag, and Catalyurek shows how a contemporaryDNA sequencing workloadcan benefit – in terms of reduced slowdown– from the application of earliest-deadline-firstto moldable-job scheduling. Two other papers combine traditional scheduling techniques with modern technologicalchallengesandcapabilities.For the former,the paper by Klusa´ˇcek andRudova´?addressesa deficiency in many ofthe workloadtracesused for job scheduling simulations: the lack of complete system and workload information, includingmachinecharacteristicsandfailures.Addingthesedatatoatrace(syn- theticallyorwithactualcollecteddata,inthecaseofMetaCentrum’strace)can significantly alter the results of a scheduler evaluation.For the latter, the paper by Verboven, Vanmechelen, and Broeckhoveintroduces a scheduling scheme for virtualmachines, where the workloadhas mixed high and low quality-of-service (QoS)requirements.Theschedulertakesadvantageofthe relativeeasethatvir- tualizationoffersforjobpreemptiontoletoverbookedlow-priorityjobsfillinfor underutilized resources,without interrupting high-priority jobs. Overbooking is alsoexploredinthepaperbyBirkenheuer,Brinkmann,andKarl,extendingtheir work from the previous workshop. Here, the authors integrate a statistical risk assessmentmoduletotheirGrid/Cloudschedulerthatusesautomatedrun-time predictions to overbook backfilled “gaps”, while still minimizing the economic penalties to the system owner from missed deadlines. The economics of Grid scheduling and QoS were in fact a major theme in this year’s papers, and played a minor part in several other papers. The paper by Sandholm and Lai explores a dynamic resource allocation scheduler for the popular Hadoop environment, in which users receive a share of computational resources that is proportional to their bid. On a related vein, in Ding’s con- tribution, a greedy double-auction mechanism to dynamically price resources is simulated,backedbyatheoreticalanalysis.Xiong’spaperonthe otherhandas- sumes static pricing per QoS level, and proposes a resource allocation approach to minimize the total cost to the application, again offering a theoretical treat- ment of the optimization problem.Similarly, in the paper by Takefusa,Nakada, Kudoh,andTanaka,alinear-programmingmethodisusedtoco-allocateprocess- ing and networking resources to Grid applications, based on an actual system’s problem statement. And last but not least, in the metascheduler proposed by Fo¨lling, Grimme, Lepping, and Papaspyrou, different sites can “lease;; under- utilized resources to heavily loaded sites that request them. A few years ago, we described the transition of the classic job scheduling paradigm for parallel processing based on the proliferation of new technolo- gies like many-core architectures and Grid or Cloud computing. Today, we can observethat these technologicaladvances produce significantchangesin the us- age patterns. Due to many-core architectures, the exploitation of parallelism is not restricted any more to a few high performance applications. Similar to the situation during the last decade of the previous century when every application exploitedsuperscalarprocessors,applicationsarenowexpectedtomakeefficient useofmultiple cores.This ofcourseresultsinnew challengesforjobscheduling. Preface VII Furthermore, Grid and Cloud computing provides affordable processing power tomanyusers,leadingtonewapplicationsifthereareefficientresourcemanage- mentsystems.Theseresourcemanagementsystemswillconsistofmultiplelayers relatedto users,resourceproviders,anddomainmanagers.However,the alloca- tion of tasks to these layers is still under investigation. Therefore, we strongly believe that research in the field of this workshop will remain interesting and challenging in the years to come. The proceedings of previous workshopsare availablefrom Springer as LNCS volumes 949, 1162, 1291, 1459, 1659, 1911, 2221, 2537, 2862, 3277, 3834, 4376, 4942, and 5798. Since 1995 these volumes have also been available online. June 2010 Eitan Frachtenberg Uwe Schwiegelshohn Organization Workshop Organizers Eitan Frachtenberg Facebook Uwe Schwiegelshohn TU Dortmund University Program Committee Henri Casanova University of Hawaii at Manoa Walfredo Cirne Google Julita Corbalan Technical University of Catalunya Dick Epema Delft University of Technology Dror Feitelson The Hebrew University Allan Gottlieb New York University Rajkumar Kettimuthu Argonne National Lab Virginia Lo University of Oregon Jose Moreira IBM Thomas J. Watson Research Center Bill Nitzberg Altair Engineering, Inc. Angela Sodan University of Windsor Mark Squillante IBM Thomas J. Watson Research Center Dan Tsafrir Technion Ramin Yahyapour TU Dortmund University Table of Contents Resource Provisioning in SLA-Based Cluster Computing .............. 1 Kaiqi Xiong and Sang Suh AnAdvanceReservation-BasedCo-allocationAlgorithmforDistributed Computers and Network Bandwidth on QoS-GuaranteedGrids ........ 16 Atsuko Takefusa, Hidemoto Nakada, Tomohiro Kudoh, and Yoshio Tanaka A Greedy Double Auction Mechanism for Grid Resource Allocation .... 35 Ding Ding, Siwei Luo, and Zhan Gao Risk Aware Overbooking for Commercial Grids ...................... 51 Georg Birkenheuer, Andr´e Brinkmann, and Holger Karl The Gain of Resource Delegation in Distributed Computing Environments ................................................... 77 Alexander Fo¨lling, Christian Grimme, Joachim Lepping, and Alexander Papaspyrou A Moldable Online Scheduling Algorithm and Its Application to ParallelShort Sequence Mapping .................................. 93 Erik Saule, Doruk Bozda˘g, and Umit V. Catalyurek Dynamic ProportionalShare Scheduling in Hadoop................... 110 Thomas Sandholm and Kevin Lai The Importance of Complete Data Sets for Job Scheduling Simulations ..................................................... 132 Dalibor Klusa´ˇcek and Hana Rudov´a HierarchicalSchedulingofDAGStructuredComputations onManycore Processorswith Dynamic Thread Grouping ......................... 154 Yinglong Xia, Viktor K. Prasanna, and James Li Multiplexing Low and High QoS Workloads in Virtual Environments ... 175 Sam Verboven, Kurt Vanmechelen, and Jan Broeckhove Proposal and Evaluation of APIs for Utilizing Inter-Core Time AggregationScheduler............................................ 191 Satoshi Yamada and Shigeru Kusakabe Using Inaccurate Estimates Accurately ............................. 208 Dan Tsafrir Author Index.................................................. 223 Resource Provisioning in SLA-Based Cluster Computing Kaiqi Xiong and Sang Suh Department of Computer Science, Texas A&MUniversity, Commerce, TX 75429, USA kaiqi [email protected] Abstract. Cluster computing is excellent for parallel computation. It has become increasingly popular. In cluster computing, a service level agreement (SLA) is a set of quality of services (QoS) and a fee agreed between a customerand an application serviceprovider. Itplaysan im- portantroleinane-businessapplication.Anapplicationserviceprovider uses a set of cluster computing resources to support e-business appli- cations subject to an SLA. In this paper, the QoS includes percentile response time and cluster utilization. We present an approach for re- sourceprovisioninginsuchanenvironmentthatminimizesthetotalcost of cluster computing resources used by an application service provider forane-businessapplicationthatoftenrequiresparallelcomputationfor high service performance, availability, and reliability while satisfying a QoS and a fee negotiated between a customer and the application ser- vice provider. Simulation experiments demonstrate the applicability of theapproach. Keywords:Clustercomputing,schedulingtheory,resourceprovisioning, service level agreement, and percentile response time. 1 Introduction In computer science,scheduling theory is concernedwith the optimal allocation ofscarceresourcessuchasservers,processorsandnetworklinkstocomputerser- viceactivitiesovertime,withtheobjectiveofoptimizingoneorseveralcomputer performance measures (e.g., see Levner [13]). Cluster computing is excellent for parallel computation. It has become increasingly popular. The management of computing service resources is fundamental to cluster computing. The increas- ing pervasiveness of network connectivity and the proliferationof on demand e- businessapplicationsandservicesinpublicdomains,corporatenetworks,aswell as home environmentsgive rise to the need for the designof appropriateservice management solutions in cluster computing. Accurately predicting e-business application and scientific computation performance based on service statistics and a customer’s perceived quality allows an application service provider (sim- ply called a service provider) not only to assure quality of services but also to avoid over provisioning to meet a service level agreement (SLA). E.FrachtenbergandU.Schwiegelshohn(Eds.):JSSPP2010,LNCS6253,pp.1–15,2010. (cid:2)c Springer-VerlagBerlinHeidelberg2010
Description:The list of books you might like
