Dror G. Feitelson – författare

Reliable performance evaluations require the use of representative workloads. This is no easy task since modern computer systems and their workloads are complex, with many interrelated attributes and complicated structures. Experts often use sophisticated mathematics to analyze and describe workload models, making these models difficult for practitioners to grasp. This book aims to close this gap by emphasizing the intuition and the reasoning behind the definitions and derivations related to the workload models. It provides numerous examples from real production systems, with hundreds of graphs. Using this book, readers will be able to analyze collected workload data and clean it if necessary, derive statistical models that include skewed marginal distributions and correlations, and consider the need for generative models and feedback from the system. The descriptive statistics techniques covered are also useful for other domains.

Thisvolumecontainsthepaperspresentedatthe8thWorkshoponJobSched- ingStrategiesforParallelProcessing,whichwasheldinconjunctionwith HPDC11andGGF5inEdinburgh,UK,onJuly24,2002. Thepapershave beenthroughacompletereviewprocess,withthefullversionbeingreadand evaluatedby?vetosevenmembersoftheprogramcommittee. Wewouldliketo takethisopportunitytothanktheprogramcommittee,AndreaArpaci-Dusseau, WalfredoCirne,AllenDowney,WolfgangGentzsch,AllanGottlieb,MoeJette, RichardLagerstrom,JensMache,CathyMcCann,ReaganMoore,BillNitzberg, MarkSquillante,andJohnTowns,foranexcellentjob. Thanksarealsodueto theauthorsfortheirsubmissions,presentations,and?nalrevisionsforthisv- ume. Finally,wewouldliketothanktheMITLaboratoryforComputerScience andtheSchoolofComputerScienceandEngineeringattheHebrewUniversity fortheuseoftheirfacilitiesinthepreparationoftheseproceedings. Thisyearsawanemphasisontwomainthemes. The?rstwastheclassical MPPschedulingarea. Themainfocusinthisareawasonback?lling,including severaladvancedvariationsonthebasicscheme. Itisalsonoteworthythatseveral papersdiscussedtheuseofadaptivenessinjobscheduling.Thesecondmajor themewasschedulinginthecontextofgridcomputing,whichisemergingasan areaofmuchactivityandrapidprogress. Thesearecomplementedbyaninvited paperprovidinganoverviewoftheschedulingandresourcemanagementareaof theGlobalGridForum(GGF)e?ort. Thiswastheeighthannualworkshopinthisseries,whichre?ectsthec- tinuedinterestinthisarea. Theproceedingsofpreviousworkshopsareavailable fromSpringer-VerlagasLNCSvolumes949,1162,1291,1459,1659,1911,and 2221(andsince1998theyhavealsobeenavailableonline). Wehopeyou?ndthesepapersinterestinganduseful. September2002 DrorFeitelson LarryRudolph UweSchwiegelshohn Table of Contents ASelf-TuningJobSchedulerFamilywithDynamicPolicySwitching ...1 AchimStreit PreemptionBasedBack?ll...24 QuinnO. Snell,MarkJ. Clement,andDavidB. Jackson JobSchedulingfortheBlueGene/LSystem...38 ElieKrevat,Jos'eG. Castan"Thisvolumecontainsthepaperspresentedatthe8thWorkshoponJobSched- ingStrategiesforParallelProcessing,whichwasheldinconjunctionwith HPDC11andGGF5inEdinburgh,UK,onJuly24,2002. Thepapershave beenthroughacompletereviewprocess,withthefullversionbeingreadand evaluatedby?vetosevenmembersoftheprogramcommittee.Wewouldliketo takethisopportunitytothanktheprogramcommittee,AndreaArpaci-Dusseau, WalfredoCirne,AllenDowney,WolfgangGentzsch,AllanGottlieb,MoeJette, RichardLagerstrom,JensMache,CathyMcCann,ReaganMoore,BillNitzberg, MarkSquillante,andJohnTowns,foranexcellentjob. Thanksarealsodueto theauthorsfortheirsubmissions,presentations,and?nalrevisionsforthisv- ume. Finally,wewouldliketothanktheMITLaboratoryforComputerScience andtheSchoolofComputerScienceandEngineeringattheHebrewUniversity fortheuseoftheirfacilitiesinthepreparationoftheseproceedings. Thisyearsawanemphasisontwomainthemes. The?rstwastheclassical MPPschedulingarea. Themainfocusinthisareawasonback?lling,including severaladvancedvariationsonthebasicscheme. Itisalsonoteworthythatseveral papersdiscussedtheuseofadaptivenessinjobscheduling. Thesecondmajor themewasschedulinginthecontextofgridcomputing,whichisemergingasan areaofmuchactivityandrapidprogress. Thesearecomplementedbyaninvited paperprovidinganoverviewoftheschedulingandresourcemanagementareaof theGlobalGridForum(GGF)e?ort. Thiswastheeighthannualworkshopinthisseries,whichre?ectsthec- tinuedinterestinthisarea.Theproceedingsofpreviousworkshopsareavailable fromSpringer-VerlagasLNCSvolumes949,1162,1291,1459,1659,1911,and 2221(andsince1998theyhavealsobeenavailableonline). Wehopeyou?ndthesepapersinterestinganduseful. September2002 DrorFeitelson LarryRudolph UweSchwiegelshohn Table of Contents ASelf-TuningJobSchedulerFamilywithDynamicPolicySwitching ...1 AchimStreit PreemptionBasedBack?ll...24 QuinnO. Snell,MarkJ. Clement,andDavidB. Jackson JobSchedulingfortheBlueGene/LSystem...38 ElieKrevat,Jos'eG. Castan"os,andJos'eE. Moreira SelectiveReservationStrategiesforBack?llJobScheduling...55 SrividyaSrinivasan,RajkumarKettimuthu,VijaySubramani, andPonnuswamySadayappan Multiple-QueueBack?llingSchedulingwithPrioritiesandReservations forParallelSystems...72 BarryG. LawsonandEvgeniaSmirni SchedulingJobsonParallelSystemsUsingaRelaxedBack?llStrategy...88 WilliamA. Ward,Jr. ,CarrieL. Mahood,andJohnE. West TheImpactofMoreAccurateRequestedRuntimes onProductionJobSchedulingPerformance...103 Su-HuiChiang,AndreaArpaci-Dusseau,andMaryK. Vernon EconomicSchedulinginGridComputing...128 CarstenErnemann,VolkerHamscher,andRaminYahyapour SNAP:AProtocolforNegotiatingServiceLevelAgreements andCoordinatingResourceManagementinDistributedSystems...153 KarlCzajkowski,IanFoster,CarlKesselman,VolkerSander, andStevenTuecke LocalversusGlobalSchedulerswithProcessorCo-allocation inMulticlusterSystems...184 AncaI. D. BucurandDickH. J. Epema PracticalHeterogeneousPlaceholderScheduling inOverlayMetacomputers:EarlyExperiences...205 ChristopherPinchak,PaulLu,andMarkGoldenberg CurrentActivitiesintheSchedulingandResourceManagementArea oftheGlobalGridForum...229 BillNitzbergandJenniferM. Schopf Author Index...237 ASelf-TuningJobSchedulerFamily withDynamic PolicySwitching Achim Streit 2 PC - Paderborn Center for Parallel Computing, Paderborn University 33102 Paderborn, Germany streit@upb. de http://www. upb. de/pc2 Abstract. The performance of job scheduling policies strongly depends on the properties of the incoming jobs. If the job characteristics often change, the scheduling policy should follow these changes. For this p- pose the dynP job scheduler family has been developed. The idea is to dynamically switch the scheduling policy during runtime.In a basic v- sion the policy switching is controlled by two parameters. The basic concept of theself-tuning dynP scheduler is to compute virtual schedules for each policy in every scheduling step. That policy is chosen which generates the 'best' schedule. The performance of the self-tuning dynP scheduler no longer depends on a adequate setting of the input parameters.

This volume contains the papers presented at the sixth workshop on Job Sched- ing Strategies for Parallel Processing, which was held in conjunction with the IPDPS 2000 Conference in Cancun, Mexico, on 1 May 2000. The papers have been through a complete refereeing process, with the full version being read and evaluated by ?ve to seven members of the program committee. We would like to take this opportunity to thank the program committee, Andrea Arpaci-Dusseau, Fran Berman, Steve Chapin, Allen Downey, Allan Gottlieb, Atsushi Hori, Phil Krueger, Richard Lagerstrom, Virginia Lo, Reagan Moore, Bill Nitzberg, Uwe Schwiegelshohn, and Mark Squillante, for an excellent job. Thanks are also due to the authors for their submissions, presentations, and ?nal revisions for this volume. Finally, we would like to thank the MIT Laboratory for Computer S- ence and the Computer Science Institute at the Hebrew University for the use of their facilities in the preparation of these proceedings. This was the sixth annual workshop in this series, which re?ects the continued interest in this ?eld. The previous ?ve were held in conjunction with IPPS'95 through IPPS/SPDP'99.Their proceedings are available from Springer-Verlag as volumes 949, 1162, 1291, 1459, and 1659 of the Lecture Notes in Computer Science series. The last two are also available on-line from Springer LINK.

This book constitutes the thoroughly refereed post-proceedings of the 7th International Workshop on Job Scheduling Strategies for Parallel Processing, JSSPP 2001, held in Cambridge, MA, USA, in June 2001.The 11 revised full papers presented were carefully selected and improved during two rounds of reviewing and revision, and present state-of-the-art results in the area.

This book constitutes the thoroughly refereed post-proceedings of the 7th International Workshop on Job Scheduling Strategies for Parallel Processing, JSSPP 2001, held in Cambridge, MA, USA, in June 2001.The 11 revised full papers presented were carefully selected and improved during two rounds of reviewing and revision, and present state-of-the-art results in the area.

This volume contains the papers selected after a very careful refereeing process for presentation during the Workshop on Job Scheduling Stategies for Parallel Processing, held in Santa Barbara, California, as a prelude to the IPPS '95 conference in April 1995.The 19 full papers presented demonstrate that parallel job scheduling takes on a crucial role as multi-user parallel supercomputers become more widespread. All aspects of job scheduling for parallel systems are covered, from the perspectives of academic research, industrial design of parallel systems, as well as user needs. Of particular interest, also for nonexpert readers, is the introductory paper "Parallel Job Scheduling: Issues and Approaches" by the volume editors.

This book constitutes the strictly refereed post-workshop proceedings of the International Workshop on Job Scheduling Strategies for Parallel Processing, held in conjunction with IPPS '96 symposium in Honolulu, Hawaii, in April 1996.The book presents 15 thoroughly revised full papers accepted for inclusion on the basis of the reports of at least five program committee members. The volume is a highly competent contribution to advancing the state-of-the-art in the area of job scheduling for parallel supercomputers. Among the topics addressed are job scheduler, workload evolution, gang scheduling, multiprocessor scheduling, parallel processor allocation, and distributed memory environments.

This book constitutes the strictly refereed post-workshop proceedings of the 1997 IPPS Workshop on Job Scheduling Strategies for Parallel Processing held in Geneva, Switzerland, in April 1997, as a satelite meeting of the IEEE/CS International Parallel Processing Symposium.The 12 revised full papers presented were carefully reviewed and revised for inclusion in the book. Also included is a detailed introduction surveying the state of the art in the area. Among the topics covered are processor allocation, parallel scheduling, massively parallel processing, shared-memory architectures, gang scheduling, etc.

This book constitutes the thoroughly refereed post-workshop proceedings of the 4th International Workshop on Job Scheduling Strategies for Parallel Processing held during IPPS/SPDP'98, in Orlando, Florida, USA, in March 1998. The 13 revised full papers presented have gone through an iterated reviewing process and give a report on the state of the art in the area.

This volume contains the papers presented at the f th workshop on Job SchedulingStrategiesforParallelProcessing,whichwasheldinconjunctionwith the IPPS/SPDP'99conference in San Juan, Puerto Rico, on April 16, 1999.The papers have been through a complete refereeing process, with the full version beingreadandevaluatedbyv etosevenmembersoftheprogramcommittee.We would like to take this opportunity to thank the program committee, Andrea Arpaci-Dusseau, Stephen Booth, Allen Downey, Allan Gottlieb, Atsushi Hori, PhilKrueger,RichardLagerstrom,MironLivny,VirginiaLo,ReaganMoore,Bill Nitzberg,UweSchwiegelshohn,KenSevcik,MarkSquillante,andJohnZahorjan, for an excellent job. Thanks are also due to the authors for their submissions, presentations,and nal revisionsfor this volume. Finally,we wouldlike to thank the MIT Laboratory for Computer Science and the Computer Science Institute at the Hebrew Universityfor the use of their facilities in the preparationof these proceedings. Thiswasthe fth annualworkshopinthis series,whichre?ectsthe continued interest in this eld. The previous four were held in conjunction with IPPS'95 through IPPS/SPDP'98.Their proceedings are available from Springer-Verlag as volumes 949, 1162, 1291, and 1459 of the Lecture Notes in Computer Science series. Sinceour rstworkshop,parallelprocessinghas evolvedtothe pointwhereit is no longer synonymous with scienti c computing on massively parallel sup- computers. In fact, enterprise computing on one hand and metasystems on the other hand often overshadow the original uses of parallel processing. This shift has underscored the importance of job scheduling in multi-user parallelsystems. Correspondingly, we had a session in the workshop devoted to job scheduling on standalonesystems, emphasizing gang scheduling, and another on scheduling for meta-systems. A third session continued the trend from previous workshops of discussing evaluation methodology and workloads. Aninnovationthisyearwasapaneldiscussiononthepossiblestandardization ofaworkloadbenchmarkthatwillservefortheevaluationofdi erentschedulers.