1. A Study of Caching in Parallel File Systems Dissertation Proposal Brad Settlemyer

2. 2 Trends in Scientific Research Scientific inquiry is now information intensive –Astronomy, Biology, Chemistry, Climatology, Particle Physics – all utilize massive data sets Data sets under study are often very large –Genomics Databases (50 TB and growing) –Large Hadron Collider (15 PB/yr) Time spent manipulating data often exceeds time spent performing calculations –Checkpointing I/O demands are particularly problematic

3. 3 Typical Scientific Workflow 1.Acquire data Observational Data (sensor-based, telescope, etc.) Information Data (gene sequences, protein folding) 2.Stage/Reorganize data to fast file system Archive retrieval Filtering extraneous data 3.Process data (e.g. Feature Extraction) 4.Output results data 5.Reorganize data for visualization 6.Visualize Data

4. 4 Trends in Supercomputing CPU performance is increasing faster than disk performance –Multicore CPUs and increased intra-node parallelism Main memories are large –4GB cost < $100.00 Networks are fast and wide –>10Gb network and buses available Num Application Processes is increasing rapidly –RoadRunner > 128K concurrent processes achieving >1 Petaflop –BlueGene/P > 250K concurrent processes achieving >1 Petaflop

5. 5 I/O Bottleneck Application processes are able to construct I/O requests faster than the storage system can provide service Applications are unable to fully utilize the massive amounts of available computing power

6. 6 Parallel File Systems Addresses I/O bottleneck by providing simultaneous access to large number of disks Switched Network I/O Nodes CPU Nodes Process 0 PFS Server 0 Process 1Process 2Process 3 PFS Server 3 PFS Server 2 PFS Server 1

7. 7 PFS Data Distribution PFS Server 0 PFS Server 3 PFS Server 2 PFS Server 1 Strip A Strip B Strip C Strip D Strip E Strip F Logical File Data Physical Data Locations Strip A Strip E Strip B Strip F Strip DStrip C

8. 8 Parallel File Systems (cont.) Aggregate file system bandwidth requirements largely met –Large, aligned data requests can be rapidly transferred –Scalable to hundreds of client processes and improving Areas of inadequate performance –Metadata Operations (Create, Remove, Stat) –Small Files –Unaligned Accesses –Structured I/O

9. 9 Scientific Workflow Performance 1.Acquire or Simulate Data Primarily limited by physical bandwidth characteristics 2.Move or Reorganize Data for Processing Often metadata intensive 3.Data Analysis or Reconstruction Small, unaligned accesses perform poorly 4.Move/Reorganize Data for visualization May perform poorly (small, unaligned accesses) 5.Visualize Data Benefits from reorganization

10. 10 Alleviating the I/O bottleneck Avoid data reorganization costs –Additional work that does not modify results –Limits use of high level libraries Increase contiguity/granularity –Interconnects and parallel file systems are well tuned for large contiguous file accesses –Limits use of low latency messaging available between cores Improve locality –Avoid device accesses entirely –Difficult to achieve in user applications

11. 11 Benefits of Middleware Caching Improves locality –PVFS Acache and Ncache –Improve write-read and read-read accesses Small accesses –Can bundle small accesses into compound operation Alignment –Can compress accesses by performing aligned requests Transparent to application programmer

12. 12 Proposed Caching Techniques In order to improve the performance of small and unaligned file accesses, we propose middleware designed to enhance parallel file systems with the following: 1.Shared, Concurrent Access Caching 2.Progressive Page Granularity Caching 3.MPI File View Caching

13. 13 Shared Caching Single data cache per node –Leverages trend toward large numbers of cores –Improves contiguity of alternating request patterns Concurrent access –Single Reader/Writer –Page locking system

14. 14 File Write Example Logical File Process 0 I/O RequestsProcess 1 I/O Requests

15. 15 File Write Example Logical File Process 0 I/O RequestsProcess 1 I/O Requests

16. 16 File Write Example Logical File Process 0 I/O RequestsProcess 1 I/O Requests

17. 17 File Write Example Logical File Process 0 I/O RequestsProcess 1 I/O Requests

18. 18 File Write Example Logical File Process 0 I/O RequestsProcess 1 I/O Requests

19. 19 File Write w/ Cache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Page 0Cache Page 2Cache Page 1

20. 20 File Write w/ Cache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Page 0Cache Page 2Cache Page 1

21. 21 File Write w/ Cache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Page 0Cache Page 2Cache Page 1

22. 22 File Write w/ Cache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Page 0Cache Page 2Cache Page 1

23. 23 File Write w/ Cache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Page 0Cache Page 2Cache Page 1

24. 24 Progressive Page Caching Benefits of paged caching –Efficient for the file system –Reduces cache metadata overhead Issues with paged caching –Aligned pages may retrieve more data than otherwise required –Unaligned writes do not cache easily Read the remaining page fragment Do not update cache with small writes Progressive paged caching addresses issues while minimizing performance and metadata overhead

25. 25 Unaligned Access Caches Accesses are independent and not on page boundaries Requires increased cache overhead How to organize unaligned data –List I/O Tree –Binary Space Partition Tree

26. 26 Paged Cache Organization Logical File

27. 27 BSP Tree Cache Organization 12 1 4 20 5 Logical File 11 8

28. 28 List I/O Tree Cache Organization 10,2 0,1 2,2 Logical File 5,3

29. 29 Progressive Page Organization Logical File 2,21,3 0,1 2,2

30. 30 View Cache MPI provides a more descriptive facility for describing file I/O –Collective I/O –MPI provides file views for describing file subregions Use file views as a mechanism for coalescing reads and writes during collective I/O How to take the union of multiple views. –Use a heuristic approach to detect structured I/O

31. 31 Collective Read Example Logical File Process 0 I/O RequestsProcess 1 I/O Requests

32. 32 Collective Read Example Logical File Process 0 I/O RequestsProcess 1 I/O Requests

33. 33 Collective Read Example Logical File Process 0 I/O RequestsProcess 1 I/O Requests

34. 34 Collective Read w/ Cache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Block 0Cache Block 2Cache Block 1

35. 35 Collective Read w/ Cache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Block 0Cache Block 2Cache Block 1

36. 36 Collective Read w/Cache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Block 0Cache Block 2Cache Block 1

37. 37 Collective Read w/ Cache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Block 0Cache Block 2Cache Block 1

38. 38 Collective Read w/ Cache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Block 0Cache Block 2Cache Block 1

39. 39 Collective Read w/ ViewCache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Block 0Cache Block 2Cache Block 1

40. 40 Collective Read w/ ViewCache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Block 0Cache Block 2Cache Block 1

41. 41 Collective Read w/ ViewCache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Block 0Cache Block 2Cache Block 1

42. 42 Collective Read w/ ViewCache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Block 0Cache Block 2Cache Block 1

43. 43 Collective Read w/ ViewCache Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Block 0Cache Block 2Cache Block 1

44. 44 Study Methodology Simulation-based study –HECIOS Closely modelled on PVFS2 and Linux 40,000 sloc Leverages OMNeT++, INET Framework –Cache Organizations Core Sharing Aligned Page access Unaligned page access

45. 45 HECIOS Overview HECIOS System Architecture

46. 46 HECIOS Overview (cont.) HECIOS Main Window

47. 47 HECIOS Overview (cont.) HECIOS Simulation Top View

48. 48 HECIOS Overview (cont.) HECIOS Simulation Detailed View

49. 49 Contributions 1.HECIOS, the High End Computing I/O Simulator developed and made available under open source license. 2.Flash I/O and BT-IO traced at large scale and traces now publicly available 3.Rigorous study of caching factors in parallel file system 4.Novel cache designs for unaligned file access and MPI view coalescing

50. 50 The End Thank You For Your Time! Questions? Brad Settlemyer bradles@clemson.edu

51. 51 Dissertation Schedule August – Complete trace parser enhancements. Shared cache impl. Complete trace collection. September – Aligned cache sharing study. October – Unaligned cache sharing study. November – SigMetrics deadline. View coalescing cache. December – Finalize experiments. Finish writing thesis. Defend thesis.

52. 52 PVFS Scalability Read and Write Bandwidth Curves for PVFS

53. 53 Shared Caching (cont.) Logical File Process 0 I/O RequestsProcess 1 I/O Requests Cache Page 0Cache Page 2Cache Page 1

54. 54 Bandwidth Effects Write Bandwidth on Adenine (MB/sec) Num Clients PVFS w/ 8 IONodes PVFS w/ Replication 16 IONodes Percent Performance 110.39.895.1% 428.228.7101.8% 843.439.891.5% 1643.440.392.9% 3250.138.276.2%

55. 55 Experimental Data Distribution PFS Server 0 PFS Server 3 PFS Server 2 PFS Server 1 Strip A Strip B Strip C Strip D Strip E Strip F Logical File Data Physical Data Locations Strip A Strip E Strip B Strip F Strip DStrip C Strip A Strip E Strip DStrip CStrip B Strip F

56. 56 Discussion (cont.) PFS Server 0 PFS Server 3 PFS Server 2 PFS Server 1 Strip A Strip B Strip C Strip D Strip E Strip F Logical File Data Physical Data Locations Strip A Strip E Strip B Strip F Strip DStrip C Strip AStrip DStrip C Strip F Strip B Strip E

57. 57 Switched Network I/O Nodes CPU Nodes Process 0 PFS Server 0 Process 1Process 2Process 3 PFS Server 3 PFS Server 2 PFS Server 1