Slideshow Transcript

1. Slide 1: Automatic Self-Tuning Architecture for Batch Scheduler on Large Scale Computing System sugree June 9, 2008 1
2. Slide 2: I am Sugree Phatanapherom from Kasetsart University. sugree June 9, 2008 2
3. Slide 3: This research is a co-work with Asst. Prof. Putchong Uthayopas. sugree June 9, 2008 3
4. Slide 4: Ready, steady, go. sugree June 9, 2008 4
5. Slide 5: What is batch scheduler? sugree June 9, 2008 5
6. Slide 6: Batch scheduler is responsible to schedule jobs to execute on resources at the right time. sugree June 9, 2008 6
7. Slide 7: Why do we need batch scheduler? sugree June 9, 2008 7
8. Slide 8: To utilize resources efficiently. sugree June 9, 2008 8
9. Slide 9: To finish all jobs as fast as possible. sugree June 9, 2008 9
10. Slide 10: To minimize power consumption. sugree June 9, 2008 10
11. Slide 11: In general, it is so called \"resource scheduling problem\". sugree June 9, 2008 11
12. Slide 12: Jobs, Resources and Time resources time June 9, 2008 12
13. Slide 13: In this research, main criteria is to minimize cost to run the resources. sugree June 9, 2008 13
14. Slide 14: Back to the past, most works focused on improving algorithms. sugree June 9, 2008 14
15. Slide 15: To simplify the problem, this research limits scope job characteristics to independent sequential jobs. sugree June 9, 2008 15
16. Slide 16: In short, a job contains the one and only one task. sugree June 9, 2008 16
17. Slide 17: In other words, job = task. sugree June 9, 2008 17
18. Slide 18: Scheduling Algorithms Scheduling On-line Batch RR OLB MET MCT MinMin MaxMin Sufferage CSufferage XSufferage CMinMin CMaxMin June 9, 2008 18
19. Slide 19: There are on-line and batch scheduling. sugree June 9, 2008 19
20. Slide 20: The most simple algorithm is \"Round Robin\". sugree June 9, 2008 20
21. Slide 21: \"Opportunistic Load Balancing\" assigns job to the next available machine. sugree June 9, 2008 21
22. Slide 22: \"Minimum Execution Time\" assigns job to the fastest machine. sugree June 9, 2008 22
23. Slide 23: \"Minimum Completion Time\" assigns job to the machine with minimum completion time for that job. sugree June 9, 2008 23
24. Slide 24: Next are batch scheduling algorithms. sugree June 9, 2008 24
25. Slide 25: \"MinMin\" assigns shortest job to the fastest machine. sugree June 9, 2008 25
26. Slide 26: \"MaxMin\" assign longest job to the fastest machine. sugree June 9, 2008 26
27. Slide 27: \"Sufferage\" is reassignable MaxMin. sugree June 9, 2008 27
28. Slide 28: \"XSufferage\" is Sufferage with data locality. sugree June 9, 2008 28
29. Slide 29: CMinMin, CMaxMin and CSufferage are derivative with costing. sugree June 9, 2008 29
30. Slide 30: How to verify? How to evaluate? sugree June 9, 2008 30
31. Slide 31: The answer is simulation. Why? sugree June 9, 2008 31
32. Slide 32: Closed. Controllable. Reproducible. sugree June 9, 2008 32
33. Slide 33: Simulation is assumption and modeling. sugree June 9, 2008 33
34. Slide 34: Grid is a meta-scheduler and underlying cluster schedulers managing hosts. sugree June 9, 2008 34
35. Slide 35: Grid Host Cluster Scheduler Host jobs Cluster Grid Scheduler Scheduler Cluster Scheduler June 9, 2008 35
36. Slide 36: Interconnection between scheduler and processors are dedicated. sugree June 9, 2008 36
37. Slide 37: Network Storage Scheduler Processor Processor Processor Processor June 9, 2008 37
38. Slide 38: Job consists of inputs, outputs and executable. sugree June 9, 2008 38
39. Slide 39: Job Output Input Executable Machine June 9, 2008 39
40. Slide 40: Operations are 2 steps; mapping and scheduling. sugree June 9, 2008 40
41. Slide 41: Mapping \"job\" to \"machine\". sugree June 9, 2008 41
42. Slide 42: Schedule \"job\" to the exact time. sugree June 9, 2008 42
43. Slide 43: In short, the result is generic priority index. sugree June 9, 2008 43
44. Slide 44: p ij= eij r j c ij g j d ij sugree June 9, 2008 44
45. Slide 45: Time execution time period before deadline eij d ij time rj Di ready time deadline June 9, 2008 45
46. Slide 46: Cost cost cij cost gj cumulative cost June 9, 2008 46
47. Slide 47: Experimented based on GAMESS job log in ThaiGrid to assume a small and a big system and named them, KUGrid and ThaiGrid, respectively. sugree June 9, 2008 47
48. Slide 48: Makespan and cost are observed. sugree June 9, 2008 48
49. Slide 49: Makespan is the period of time from when the first job submitted to the last job finished. sugree June 9, 2008 49
50. Slide 50: Price-Performance 30000 Cost-Time Ratio ($/h) 25000 20000 15000 10000 5000 0 KU Grid Thai Grid rr olb mct met minmin maxmin sufferage cminmin cmaxmin csufferage June 9, 2008 50
51. Slide 51: Cost 4500 Thousands 4000 3500 3000 Cost ($) 2500 2000 1500 1000 500 0 KU Grid Thai Grid rr olb mct met minmin maxmin sufferage cminmin cmaxmin csufferage June 9, 2008 51
52. Slide 52: Makespan 139.150 139.100 Makespan (hours) 139.050 139.000 138.950 138.900 138.850 138.800 KU Grid Thai Grid rr olb mct met minmin maxmin sufferage cminmin cmaxmin csufferage June 9, 2008 52
53. Slide 53: Looks great! Any problems? Yes! sugree June 9, 2008 53
54. Slide 54: Priority index contains 5 factors. What are the right values? sugree June 9, 2008 54
55. Slide 55: What are the factors of those factors? sugree June 9, 2008 55
56. Slide 56: There are so many dependencies. Job characteristics. Resource characteristics. User characteristics. sugree June 9, 2008 56
57. Slide 57: This problem is so called \"Multi-variate Optimization\". sugree June 9, 2008 57
58. Slide 58: Plus, a bit more complex with evaluation in simulator. sugree June 9, 2008 58
59. Slide 59: How to solve? sugree June 9, 2008 59
60. Slide 60: Optimization Architecture Monitoring Accounting Batch System System Scheduler Simulator Simulator Simulator Optimizer Simulator June 9, 2008 60
61. Slide 61: Optimization Algorithm? sugree June 9, 2008 61
62. Slide 62: Particle Swarm Optimization is selected as the first one to try. sugree June 9, 2008 62
63. Slide 63: The position of each particle in n-dimension plane represents solution. sugree June 9, 2008 63
64. Slide 64: PSO is social influence in various scopes. sugree June 9, 2008 64
65. Slide 65: Local, neighbor and global. sugree June 9, 2008 65
66. Slide 66: Usually, one trust oneself, friends and the world, respectively. The level of trust. sugree June 9, 2008 66
67. Slide 67: PSO June 9, 2008 67
68. Slide 68: How to fully automate self-tuning process? sugree June 9, 2008 68
69. Slide 69: Historical data are the key. sugree June 9, 2008 69
70. Slide 70: The quality of solution depends on optimizer. sugree June 9, 2008 70
71. Slide 71: Running optimizer longer may return better solution. sugree June 9, 2008 71
72. Slide 72: Precision of using historical data depends on data period and amount of data. sugree June 9, 2008 72
73. Slide 73: How to use historical data? Log replay or estimation. sugree June 9, 2008 73
74. Slide 74: How to maximize solution quality to near optimal? sugree June 9, 2008 74
75. Slide 75: Just run more simulations using the whole grid system to optimize itself at night! sugree June 9, 2008 75
76. Slide 76: Results? Please accept my apologize. They are not published yet. sugree June 9, 2008 76
77. Slide 77: Conclusion. sugree June 9, 2008 77
78. Slide 78: Flexible algorithms introduce more adjustable factors. sugree June 9, 2008 78
79. Slide 79: The factors are vary from time to time. sugree June 9, 2008 79
80. Slide 80: In other view, these algorithms are improved by external optimization periodically. sugree June 9, 2008 80
81. Slide 81: Particle swarm optimization is selected to solve multi-variate optimization. sugree June 9, 2008 81
82. Slide 82: Improve scheduler by scheduler itself. sugree June 9, 2008 82
83. Slide 83: Any questions? sugree June 9, 2008 83