Using SchedFlow for Performance Evaluation of Workflow Applications

1. Using SchedFlow for Performance Evaluation of Workflow Applications Elisa Heyman Gustavo Martínez Miquel Angel Senar Emilio Luque UniversitatAutònoma de Barcelona Elisa.Heymann@uab.es Barton P. Miller University of Wisconsin bart@cs.wisc.edu 1

2. 2 T1 T2 T3 T4 T5 T6 T7 Our Problem Scheduling Policies Workflow Engines

3. 3 T1 T2 T3 T4 T5 T6 T7 Our Solution Scheduling Policies SchedFlow Workflow Engines

4. 4 Outline • Introduction • SchedFlow • Experimental Study • Conclusions

5. 5 Introduction • For executing a workflow on a distributed environment, we need: • Scheduling policy integrated into a • Workflow engine • Reduce makespan • Factors • Workloadsize • Inaccuratecomputing and communication times • Machines appearing/disapperingdynamically

6. 6 Introduction • WithSchedFlow, weassessedtheinfluence of theworkloadonthemakespanconsidering: • Differentschedulingpolicies • Differentworkflowengines

7. SchedFlow T1 Scheduler T2 T3 Scheduler Scheduler Scheduler T4 T5 T6 Controller Observer T7 Adaptor Adaptor UserPolicy API logs queue Task manager Scheduler WorkflowEngine

8. SchedFlow T1 Theusersubmits a workflow T2 T3 Scheduler Scheduler Scheduler Scheduler T4 Controller Observer Adaptor Adaptor UserPolicy API logs queue Task manager Scheduler WorkflowEngine

9. SchedFlow TheScheduler uses thespecifiedschedulingpolicyontheavailableresourcesdiscoveredbytheObserver. T1 M1 Scheduler T2 M2 Scheduler Scheduler Scheduler T3 M3 Controller Observer T4 M4 Adaptor Adaptor logs queue Task manager Scheduler WorkflowEngine

10. SchedFlow Scheduler T2 M2 Scheduler Scheduler T1 Scheduler T3 M3 Controller Observer T4 M4 Adaptor Adaptor TheControllerreceivesthefirsttask-machine pairs logs queue Task manager M1 Scheduler WorkflowEngine

11. SchedFlow Scheduler T2 M2 Scheduler Scheduler Scheduler T3 M3 Controller Observer T4 M4 Adaptor Adaptor T1 TheControllertellstheadaptorwhichengineto use. Theadaptordealswithformatting and enqueuesthetask. logs queue Task manager M1 Scheduler WorkflowEngine

12. SchedFlow Scheduler T2 M2 Scheduler Scheduler Scheduler T3 M3 Controller Observer T4 M4 Adaptor Adaptor logs queue T1 Task manager M1 Scheduler WorkflowEngine

13. SchedFlow Scheduler T2 M2 Scheduler Scheduler Scheduler T3 M3 Controller Observer T4 M4 Adaptor Adaptor TheEnginesendsthetasktotheassigned machine. TheObservercheckstheEngine log forfinishedtasks. logs queue T1 Task manager M1 Scheduler WorkflowEngine

14. SchedFlow Scheduler T2 M2 Scheduler Scheduler Scheduler T3 M3 Controller Observer T4 M4 Adaptor Adaptor Whenthetaskfinishes, theObservernotifiestheScheduler. logs queue Task manager M1 Scheduler WorkflowEngine

15. SchedFlow Scheduler Scheduler Scheduler T2 T3 Scheduler Controller Observer T4 M4 Adaptor Adaptor TheSchedulerfindsthetasksthathavetheirdependenciessatisfied and sendsthemtotheController. logs queue Task manager M2 Scheduler WorkflowEngine M3

16. SchedFlow Scheduler Scheduler Scheduler Scheduler Controller Observer T4 M4 Adaptor Adaptor T2 T3 logs queue Task manager M2 Scheduler M3 WorkflowEngine

17. SchedFlow Scheduler Scheduler Scheduler Scheduler Controller Observer T4 M4 Adaptor Adaptor logs queue T2 T3 Task manager M2 Scheduler WorkflowEngine M3

18. SchedFlow T4 M4 Scheduler Scheduler Scheduler Scheduler Controller Observer Adaptor Adaptor logs queue T2 M2 Task manager M2 Scheduler T3 M3 WorkflowEngine M3

19. SchedFlow T4 M4 Scheduler Scheduler Scheduler Scheduler Controller Observer Adaptor Adaptor T2 finishes OK. M3 isclaimed. logs queue M2 Task manager M2 Scheduler T3 WorkflowEngine M3

20. SchedFlow T4 M4 Scheduler Scheduler Scheduler Scheduler Controller Observer Adaptor Adaptor TheObserverdetectstheproblem and T3 is removed from M3 and dynamcallyrescheduled. logs queue M2 Task manager M2 Scheduler T3 M3 WorkflowEngine M3

21. SchedFlow T4 M4 Scheduler Scheduler T3 Scheduler Scheduler T3 isrescheduled. TheObserverdoesnotinclude M3 as anavailableresource. Controller Observer Adaptor Adaptor logs queue Task manager M2 Scheduler WorkflowEngine M3

22. SchedFlow T4 M4 Scheduler Scheduler M2 T3 Scheduler Scheduler Controller Observer Adaptor Adaptor logs queue Task manager M2 Scheduler WorkflowEngine M3

23. SchedFlow T4 M4 Scheduler Scheduler Scheduler T3 Scheduler Controller Observer Adaptor Adaptor logs queue Task manager M2 Scheduler WorkflowEngine M3

24. SchedFlow T4 M4 Scheduler Scheduler Scheduler Scheduler Controller Observer Adaptor Adaptor T3 logs queue Task manager M2 Scheduler WorkflowEngine M3

25. SchedFlow T4 M4 Scheduler Scheduler Scheduler Scheduler Controller Observer Adaptor Adaptor logs queue T3 Task manager M2 Scheduler WorkflowEngine M3

26. SchedFlow T4 M4 Scheduler Scheduler Scheduler Scheduler Controller Observer Adaptor Adaptor logs queue T3 Task manager M2 Scheduler WorkflowEngine M3

27. SchedFlow T4 M4 Scheduler Scheduler Scheduler Scheduler Controller Observer Adaptor Adaptor T3 finishes OK. TheObservernotifiestheScheduler, and itreleases T4. logs queue Task manager M2 Scheduler WorkflowEngine M3

28. SchedFlow Scheduler Scheduler Scheduler T4 Scheduler Controller Observer Adaptor Adaptor logs queue Task manager Scheduler M4 WorkflowEngine

29. SchedFlow Scheduler Scheduler Scheduler Scheduler Controller Observer Adaptor Adaptor T4 logs queue Task manager Scheduler M4 WorkflowEngine

30. SchedFlow Scheduler Scheduler Scheduler Scheduler Controller Observer Adaptor Adaptor logs queue T4 Task manager Scheduler M4 WorkflowEngine

31. SchedFlow Scheduler Scheduler Scheduler Scheduler Controller Observer Adaptor Adaptor When T4 finishestheObserver computes themakespan. logs queue Task manager T4 Scheduler M4 WorkflowEngine

32. 32 Experimental Study • Execution environment: • 140 machines • Workflow applications: • Montage (53 tasks) • LIGO (81 tasks) • Workflow engines: • Condor-DAGMan 7.0 • Taverna 1.4.8 • Karajan 4_0_a1

33. 33 Experimental Study • Schedulingpolicies: • Default • Min-min • HEFT • BMCT

34. 34 Experimental Study • Input workload: • 400 MB • 1024 MB • Westudiedtheeffect of theschedulingpolicies. • Wemeasuredapplicationmakespan • Real executions

35. 35 Results • MantageranonTaverna, DAGMan, Karajan • 400 MB input workload • 120 executions • Default schedulingpolicy

36. 36 Results • SameexperimentsbutusingSchedFlow • Min-min, HEFT, BMCT • Rescheduling

37. 37 Results • MantageranonTaverna, DAGMan, Karajan • 1024 MB input workload • 120 executions • Default schedulingpolicy

38. 38 Results • SameexperimentsbutusingSchedFlow • Min-min, HEFT, BMCT • Rescheduling

39. 39 Results • LIGO ranonTaverna, DAGMan, Karajan • 400 MB input workload • 120 executions • Default schedulingpolicy

40. 40 Results • SameexperimentsbutusingSchedFlow • Min-min, HEFT, BMCT • Rescheduling

41. 41 Results • LIGO ranonTaverna, DAGMan, Karajan • 1024 MB input workload • 120 executions • Default schedulingpolicy

42. 42 Results • SameexperimentsbutusingSchedFlow • Min-min, HEFT, BMCT • Rescheduling

43. 43 Conclusions • No single scheduling policy is the best for all scenarios • SchedFlow allows us to obtain better performance providing: • Flexibility regarding scheduling policies • Support for rescheduling • Integration with Workflow Engines

44. Using SchedFlow for Performance Evaluation of Workflow Applications Elisa Heyman Gustavo Martínez Miquel Angel Senar Emilio Luque UniversitatAutònoma de Barcelona Elisa.Heymann@uab.es Barton P. Miller University of Wisconsin bart@cs.wisc.edu 44