Memory State Compressors for Gigascale Checkpoint/Restore

1. Memory State Compressors for Gigascale Checkpoint/Restore www.eecg.toronto.edu/aenao Andreas Moshovos moshovos@eecg.toronto.edu

2. Gigascale Checkpoint/Restore • Several Potential Uses: • Debugging • Runtime Checking • Reliability • Gigascale Speculation Instruction Stream checkpoint Restore trigger Many instructions

3. Key Issues & This Study • Track and Restore Memory State • I/O? • This Work: Memory State Compression • Goals: • Minimize On-Chip Resources • Minimize Performance Impact • Contributions: • Used Value Prediction to simplify compression hardware • Fast, Simple and Inexpensive • Benefits whether used alone or not

4. Outline • Gigascale Checkpoint/Restore • Compressor Architecture: Challenges • Value-Prediction-Based Compressors • Evaluation

5. Our Approach to Gigascale CR (GCR) • Checkpoint: • blocks that were written into • Current Memory State + Checkpoint = Previous Memory State Checkpoints: Can be large (Mbytes) and we may want many 3 1 checkpoint begins Restore trigger 2 Checkpoint memory block on first write Restore all checkpointedmemory blocks 4 5

6. Checkpoint Storage Requirements 1G 32M 1M Max. Checkpoint Size in Bytes 32K 1K Checkpoint Interval in Instructions

7. Architecture of a GCR Compressor out-buffer in-buffer L1 Data Cache Main Memory Alignment Network Compressor Resources & Performance Size Size • Previous work: Compressor = Dictionary-Based • Relatively Slow, Complex Alignment, order 10K of Transistors • 64K In-Buffer  ~3.7% Avg. Slowdown

8. Our Compression Architecture • Standalone: • ~Compression, - Resources • In Combination: • -Resources (in-buffer), +Compression, +Performance out-buffer in-buffer DictionaryCompressor L1 Data Cache Alignment Network Main Memory VP Compressor Simple Alignment VP stage Optional

9. Value-Predictor-Based Compression Input stream Output stream mispredicted 0 value Value Predictor value predicted 1

10. Example 0 VP 0 0 22 VP 0 22 TIME 22 VP 1

11. Block VP-Based Compressor Input stream Output stream • Shown is Last-Outcome Predictor • Studied Others (four combinations per word) Header (one word) address VP 0 1 1 word 0 VP mispredicted words Cache block word 1 VP value value word 15 VP Half-word alignment single entry predictors

12. Evaluation • Compression Rates • Compared with LZW • Performance • As a function of in-buffer size

13. Methodology • Simplescalar v3 • SPEC CPU 2000 with reference inputs • Ignore first checkpoint to avoid artificially skewing the results • Simulated up to: • 80Billion instructions (compression rates) • 5Billion instructions (performance) • 8-way OOO Superscalar • 64K L1D, L1I, 1M UL2

14. Compression Rate vs. LZW better 256M Instructions Checkpoint Interval

15. Performance Degradation • LZW + 64K buffer = ~3.7% slowdown • LZW + LO + 1K buffer = 1.6% slowdown better

16. Summary • Memory State Compression for Gigascale CR • Many Potential Applications • Used Simple Value-Prediction Compressors • Few Resources • Low Complexity • Fast Performance • Can be Used Alone • Can be Combined with Dictionary-based Compressors • Reduced on-chip buffering • Better Performance • Main memory compression?