Adaptive Techniques for Homebased Software DSMs

Transcription

1 Adaptive Techniques for Homebased Software DSMs Lauro Whately, Raquel Pinto, Muralidharan Rangarajan, Liviu Iftode, Ricardo Bianchini, Claudio L. Amorim COPPE / UFRJ Rutgers University

2 Contents Motivation Adaptive Protocols HAP Experimental Results Conclusions

3 Motivation Problems with home-based protocols Location of the home Sharing between two non-home processes The choice of the home node Coherence Solution Diffs creation Single writer pages Access faults Adapt location of home and coherence protocol according to behavior

4 Adaptive Protocols Can adapt between single and multiple writer Can adapt between invalidate and update coherence Can be very successful at reducing communication, coherence and memory overheads in traditional LRC

5 Adaptive Protocols Benefits to home-based protocol Can dynamically assign the home node according to the sharing pattern Can use update-based coherence for migratory data inside critical section and producer/consumers data

6 HAP Propose: Home-based Adaptive Protocol HAP = HLRC + ADSM-like adaptiveness

7 HAP Page Sharing Patterns and Actions Falsely-shared Multiple writers Twinning and diffing No home migration Migratory Single writer Try to avoid twinning and diffing Home migrates to next writer Producer/Consumer(s) Single-writer Avoid twinning and diffing Home moves to producer Update consumers

8 HAP Single Writer Protected by locks Migratory pages associated with a lock variable Home migrates on acquire operations Protected by barriers Producer/Consumer(s) Consumers vector Sends updates with write-notices Migratory Migrates the home to the first requester/single writer

9 HAP Pattern Detection Detection done by home node and at barrier point Home node receives modifications and lock id Nodes at barrier receive write-notices Single-writer page goes back to multiple-writer if a modification is received

10 HAP Page States and Transitions

11 Experimental Results Environment : 8-nodes cluster 650 MHz Pentium III 512 MB RAM 256 KB Cache L2 Linux VIA Giganet : one-way latency = 8.2 s bandwidth = 10 MB/s Workload Appls. IS SOR FFT Problem Size 2 16 keys, 300 iter. 256x5120, 100 iter elements Synchronization Locks, barriers Barriers Barriers

12 Experimental Results IS Execution time Breakdown

13 Experimental Results IS Execution Statistics (average over all nodes) HLRC H_MI HAP Messages (k) Data (kb) Access Faults Page Requests Diffs

14 Experimental Results SOR Execution time breakdown

15 Experimental Results SOR Execution Statistics (average over all nodes) HLRC H_PC HAP Messages (k) Data (kb) Access Faults (k) Page Requests Diffs

16 Experimental Results FFT Execution time breakdown

17 Experimental Results FFT Execution Statistics (average over all nodes) HLRC H_PC H_MO HAP Messages (k) Data (kb) Access Faults (k) Page Requests (k) Diffs

18 Experimental Results Discussion Sucessful at improving performance for IS (19%) Potential for improvements with other patterns: SOR communication overhead decreased by 83% number of pages requested reduced by 86% Overhead : Detection of MIGo pages Unnecessary PC updates

19 Conclusion HAP Detects migratory, producer/consumer and multiplewriter pages Adaptation : Dynamic adaptation between multiple and single-writer coherence protocols Dynamic adaptation between invalidation and update-based coherence Home migration of single-writer pages to writing node Preliminary implementation of HAP performs well for certain applications, but requires modification for others