Opportune Job Shredding: An Efficient Approach for Scheduling Parameter Sweep Applications

1. Opportune Job Shredding:An Efficient Approach for Scheduling Parameter Sweep Applications Rohan Kurian, Pavan Balaji, P. Sadayappan The Ohio State University

2. Parameter Sweep Applications • An important class of applications • Set of independent tasks • MCell Application • 3D simulations for sub-cellular architecture/physiology • GTOMO (Parallel Tomography) Application • Multiple view-point simulation • Systems exist for scheduling on the Grid • Cluster-based Scheduling?

3. Application Level Schedulers • Manage the scheduling of applications • Break the application to appropriate chunks • APST (AppLeS Parameter Sweep Template) • NIMROD • Greedy approach to schedule PSA chunks

4. Presentation Roadmap • Job Scheduling in Clusters • Multi-Site Job Scheduling • PSA Scheduling Strategies • Multi-Site Scheduling of PSAs • Performance Evaluation • Conclusions

5. Job Scheduling in Clusters • Mapping arriving jobs to available resources • Multiple Schemes for Scheduling • First Come First Serve (FCFS) • Conservative Scheduling • Aggressive or EASY Scheduling • Fair-Share Constraints • A user can not have more than ‘N’ queued jobs • Submitting the multiple chunks of a PSA job • Violation of Fair-Share constraints • Combine chunks to form a single parallel job

6. Formation of PSAs in Clusters Small Independent Tasks Parallel Parameter Sweep Application

7. Presentation Roadmap • Job Scheduling in Clusters • Multi-Site Job Scheduling • PSA Scheduling Strategies • Multi-Site Scheduling of PSAs • Performance Evaluation • Conclusions

8. Multi-Site Job Scheduling • Multiple Simultaneous Requests • Job submitted to multiple sites • Started on the earliest cluster • Existing schemes have limitations • Heterogeneous Clusters • Different Scheduling Schemes

9. Jobs Jobs Jobs Meta Scheduler Meta Scheduler Meta Scheduler Local Scheduler Local Scheduler Local Scheduler Multiple-simultaneous-requests Site 1 Site 2 Site 3

10. Presentation Roadmap • Job Scheduling in Clusters • Multi-Site Job Scheduling • PSA Scheduling Strategies • Multi-Site Scheduling of PSAs • Performance Evaluation • Conclusions

11. PSA Scheduling Strategies • Flooding based Job Shredding • Submit all chunks in the PSA at once • Greedy approach • Improves User and System metrics • Doesn’t ensure fairness to Non-PSA jobs • Opportune Job Shredding • Uses an additional Application-Level Scheduler • Monitors the current schedule of the system • If no normal backfill is possible • Allow PSA jobs to shred and backfill

12. Presentation Roadmap • Job Scheduling in Clusters • Multi-Site Job Scheduling • PSA Scheduling Strategies • Multi-Site Scheduling of PSAs • Performance Evaluation • Conclusions

13. Multi-Site Scheduling for PSAs • Two-level Application Level Schedulers • No constraints on sites • Allowed to have different speeds • Allowed to have different scheduling policies • Similar to “Multiple Simultaneous Requests” • Simultaneous requests only for PSAs

14. Multi-Site Scheduling for PSAs Meta Application-Level Scheduler Site 1 App-Level Scheduler App-Level Scheduler Site 2 Job Queue Local Scheduler Job Queue Local Scheduler App-Level Scheduler Job Queue Local Scheduler Site 3

15. Presentation Roadmap • Job Scheduling in Clusters • Multi-Site Job Scheduling • PSA Scheduling Strategies • Multi-Site Scheduling of PSAs • Performance Evaluation • Conclusions

16. Performance Metrics • Response Time • Completion Time – Submit Time • Slowdown • Response Time / Runtime • Loss of Capacity (LOC) • LOC = min {(waiting jobs procs), idle procs} • T = Time for which this state lasts • LOC = LOC x T

17. Evaluation Scheme • Simulation based Approach • CTC trace from Feitelson’s archive • EASY backfilling used • For multi-site evaluation • CTC traces from 3 different months • Processing speeds in the ratio 2:1:3

18. Flooding Based Job Shredding • Up to 60% improvement for PSA Jobs • Up to 90% worse performance for Non-PSA Jobs

19. Flooding: Job Category wise breakup • Narrow Short Non-PSA jobs suffer most • Loss of back-filling opportunities is the main reason

20. Flooding: Loss of Capacity • Up to 75% improvement in the Loss of Capacity

21. Opportune Job Shredding • Up to 70% improvement for PSA Jobs • Less than 2% worsening in performance for Non-PSA Jobs

22. Opportune: Job Category wise breakup • No category of Non-PSA jobs suffers more than 7%

23. Opportune: Loss of Capacity • Up to 12% improvement in the Loss of Capacity

24. Opportune (Multi-Site) • Up to 95% improvement for PSA Jobs • No significant loss of performance for Non-PSA jobs

25. Opportune (Multi-Site):Response Time • Up to 75% improvement for PSA Jobs • No significant loss of performance for Non-PSA jobs

26. Opportune (Multi-Site):Slowdown • Up to 95% improvement for PSA Jobs • No significant loss of performance for Non-PSA jobs

27. Opportune (Multi-Site):Loss of Capacity • Up to 45% improvement in the Loss of Capacity

28. Concluding Remarks • Opportune Job Shredding • Efficient Scheduling of PSAs • Single Site and Multi-Site versions • Significant improvement for PSA jobs • Ensures that Non-PSA jobs are not affected • Plan to integrate this with Prod. Schedulers

29. Thank You!