Application-level Application-level SchedulingSchedulingSathish S. VadhiyarSathish S. Vadhiyar

Credits / Sources: AppLeS web Credits / Sources: AppLeS web pages and paperspages and papers

IntroductionIntroduction

Everything about system is evaluated in Everything about system is evaluated in terms of its impact on the applicationterms of its impact on the applicationAppLeS – application-specific AppLeS – application-specific metacomputing scheduling agentmetacomputing scheduling agentEach application has its own AppLeSEach application has its own AppLeSAppLeS designs and implements an AppLeS designs and implements an adaptive application-specific scheduleadaptive application-specific scheduleApplication-centric scheduling customized Application-centric scheduling customized to reflect application resource usageto reflect application resource usage

Doctrines of AppLeSDoctrines of AppLeSBoth application-specific and system-specific Both application-specific and system-specific information are needed for good schedulesinformation are needed for good schedulesPerformance depends on the application’s Performance depends on the application’s own performance criteriaown performance criteriaThe distances between resources depend on The distances between resources depend on how the application uses themhow the application uses themDynamic information to assess system stateDynamic information to assess system statePredictions are accurate only within a Predictions are accurate only within a particular lifetimeparticular lifetimeA schedule is only as good as underlying A schedule is only as good as underlying predictionprediction

ArchitectureArchitecture

CoordinatorCoordinatorResource SelectorResource SelectorPlannerPlannerPerformance EstimatorPerformance EstimatorActuatorActuator

General AppLeS StrategyGeneral AppLeS Strategy

AppLeS with JacobiAppLeS with Jacobi

The problem: Appropriate partitioning strategy to balance processor efficiencies and communication overheads, i.e. deriving partitions to obtain resource performance

Deriving Partitions for JacobiDeriving Partitions for Jacobi

NotationsNotations

Per-processor execution timePer-processor execution time

The goalThe goal

Deriving Partitions for JacobiDeriving Partitions for Jacobi

Communication timeCommunication time

Soultion: system of linear equations by Gaussian Soultion: system of linear equations by Gaussian EliminationElimination

NWS in JacobiNWS in Jacobi

Resource Selection and SchedulingResource Selection and Scheduling

Resource Selection and SchedulingResource Selection and Scheduling

AppLeS Benefits - schedulingAppLeS Benefits - scheduling

AppLeS Benefits – partitioning and AppLeS Benefits – partitioning and memory usagememory usage

AppLeS Benefits – Partitioning and AppLeS Benefits – Partitioning and Memory UsageMemory Usage

ReferencesReferences

The AppLeS Project: A Status Report The AppLeS Project: A Status Report by by Fran BermanFran Berman and and Rich WolskiRich Wolski. . from Proceedings of the 8th NEC Research from Proceedings of the 8th NEC Research Symposium, Berlin, Germany, May 1997.Symposium, Berlin, Germany, May 1997.Application-Level Scheduling on Distributed Application-Level Scheduling on Distributed Heterogeneous Networks by Heterogeneous Networks by Fran Berman, Fran Berman, Richard Wolski, Silvia Figueira, Jennifer Schopf,Richard Wolski, Silvia Figueira, Jennifer Schopf, and and Gary ShaoGary Shao from Proceedings of Supercomputing 1996 from Proceedings of Supercomputing 1996