1. Increasing the Energy Efficiency of TLS Systems Using Intermediate Checkpointing Salman Khan 1, Nikolas Ioannou 2, Polychronis Xekalakis 3 and Marcelo Cintra 2 1 University of Manchester 2 University of Edinburgh 3 Intel Labs Barcelona - UPC

2. HiPC 20112 Introduction  Power efficiency, complexity and time-to-market reasons lead to CMPs  Problem: –No benefits for sequential applications –Even for mostly parallel applications Amdahl’s Law limits performance gains with many cores  Solution: Thread Level Speculation(TLS) –But performance through TLS costs in energy Can we reduce the wastefulness of re-execution due to misspeculation without losing performance?

3. 3 Key Contributions  Propose checkpointing to improve efficiency of speculative execution  Evaluate dependence prediction techniques to guide checkpoint placement  Our approach results in an energy saving of up to 14%, with 7% on average over normal TLS execution, with no significant effect on speedup. HiPC 2011

4. 4 Outline  Introduction  Checkpointing  Dependence Predictors  Checkpointing Policy  Experimental Setup and Results  Conclusions HiPC 2011

5. Thread Level Speculation 5HiPC 2011

6. Thread Level Speculation with Checkpointing 6HiPC 2011

7. 7 Outline  Introduction  Checkpointing  Dependence Predictors  Checkpointing Policy  Experimental Setup and Results  Conclusions HiPC 2011

8. Placing Checkpoints  Stride  Dependence Prediction –Address based –Program Counter Based –Hybrid HiPC 20118

9. Dependence Prediction HiPC 20119

10. Hybrid Dependence Predictor HiPC 201110

11. 11 Outline  Introduction  Checkpointing  Dependence Predictors  Checkpointing Policy  Experimental Setup and Results  Conclusions HiPC 2011

12. Placing Checkpoints  Limited number of checkpoints  Placing a checkpoint has a cost  Checkpointing on every positive prediction results in too many checkpoints HiPC 201112

13. 13 Outline  Introduction  Checkpointing  Dependence Predictors  Checkpointing Policy  Experimental Setup and Results  Conclusions HPCA 2010

14. Setup  Simulator, Compiler and Benchmarks: –SESC (http://sesc.sourceforge.net/)http://sesc.sourceforge.net/ –POSH (Liu et al. PPoPP ‘06) –Spec 2000 Int.  Architecture: –Four way CMP, 4-Issue cores –16KB L1 Data (multi-versioned) and Instruction Caches –1MB unified L2 Caches –Cycles from Violation to Kill/Restart: 12 –Cycles to Spawn: 12 HiPC 201114

15. Measuring Dependence Prediction HiPC 201115

16. ICS 200916

17. HiPC 201117  Wasted Instructions: Unnecessarily squashed instructions.

18. HiPC 201118

19. HiPC 201119

20. 20 Outline  Introduction  Checkpointing  Dependence Predictors  Checkpointing Policy  Experimental Setup and Results  Conclusions HPCA 2010

21. Conclusions  Effective checkpointing improves the efficiency of TLS  Placing checkpoints by stride is not sufficient to reduce waste significantly  Checkpointing using dependence predication obtains energy saving of up to 14%, with 7% on average over normal TLS execution, with no significant effect on speedup. HiPC 201121

22. Read the paper for…  Complete results  Microarchitectural issues that arise from checkpointing running tasks  Modified squash/restart mechanism that is needed to avoid performance degradation from checkpointing HiPC 201122