Upgrade to Pro — share decks privately, control downloads, hide ads and more …

MAGEEC

Sponsored · Ship Features Fearlessly Turn features on and off without deploys. Used by thousands of Ruby developers. →

 MAGEEC

At the GNU Tools Cauldron in California, Jeremy Bennett and I introduced the MAGEEC project to the GCC community and began discussions on how we aim to reduce software energy usage.

Simon Cook

July 11, 2013

Transcript

35 slides
  1. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license What is MAGEEC? Today we optimize for speed or space
  2. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license What is MAGEEC? Today we optimize for speed or space What if we could optimize for energy usage?
  3. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license How We Got Here Research into modeling energy usage
  4. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license How We Got Here Research into modeling energy usage Energy measurement
  5. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license How We Got Here Research into modeling energy usage Research into feedback directed optimization Energy measurement
  6. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license How We Got Here Research into modeling energy usage Research into feedback directed optimization Energy measurement
  7. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license How We Got Here Research into modeling energy usage Research into feedback directed optimization Energy measurement
  8. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license How We Got Here Research into modeling energy usage Research into feedback directed optimization Energy measurement
  9. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Literature Review • Energy measuring and modeling – The software drained my battery. Kerstin Eder & Jeremy Bennett, NMI Yearbook 2012, www.embecosm.com/resources/articles/#EAR12. • MILEPOST GCC - Feedback directed optimization – ctuning.org/milepost-gcc • Measurement of compiler energy usage – Identifying Compiler Options to Minimize Energy Consumption for Embedded Platforms. James Pallister, Simon Hollis, Jeremy Bennett arxiv.org/abs/1303.6485 • MAGEEC – mageec.org
  10. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license What's New? Objective is energy optimization
  11. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license What's New? Objective is energy optimization Energy measured not modeled
  12. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Energy Measurement • Based on ST Discovery board – ARM Cortex A8 with ADC daughter board • Burst sample rate 6Msample/s – 192kB on-board RAM buffer – Short code samples e.g. superoptimizer • Sustained sample rate 2Msample/s – Pre-processed and Streamed off-board – Used for MAGEEC • The board in action: – mageec.org/wiki/Power_Sensing_Board
  13. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license What's New? Objective is energy optimization Generic framework: GCC and LLVM initially Energy measured not modeled
  14. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license What's New? Objective is energy optimization Generic framework: GCC and LLVM initially Energy measured not modeled Working system, not research prototype
  15. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Our Plan • Implement MILEPOST concepts in a generic way. • Train and evaluate based on real hardware energy measurements and existing passes. • Write and evaluate optimization passes specifcally for energy efficiency (Jörn Rennecke).
  16. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner Plugin I/F ML I/F MAGEEC MAGEEC
  17. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner * may be gen_features()
  18. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() * may be gen_features() init
  19. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() init() * may be gen_features() version / target init
  20. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() init() init() * may be gen_features() version / target version / target init
  21. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() init() init() gen_features() * may be gen_features() version / target version / target init
  22. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() init() init() gen_features() decision() * may be gen_features() version / target version / target program features init running loop
  23. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() init() init() gen_features() decision() run_pass() * may be gen_features() version / target version / target program features pass list init running loop
  24. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() init() init() gen_features() decision() run_pass() stats_gen()* * may be gen_features() version / target version / target program features pass list init running loop
  25. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() init() init() gen_features() decision() run_pass() stats_gen()* next_pass() mod_stats()* * may be gen_features() version / target version / target program features pass list init running loop
  26. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() init() init() gen_features() decision() run_pass() stats_gen()* next_pass() mod_stats()* decision() * may be gen_features() version / target version / target program features new features pass list init running loop
  27. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() init() init() gen_features() decision() run_pass() stats_gen()* next_pass() mod_stats()* decision() end() * may be gen_features() version / target version / target program features new features pass list init running loop finish
  28. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() init() init() gen_features() decision() run_pass() stats_gen()* next_pass() mod_stats()* decision() end() end() * may be gen_features() version / target version / target program features new features pass list init running loop finish
  29. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Overall Design Compiler Coordinator Machine Learner plugin_init() init() init() gen_features() decision() run_pass() stats_gen()* next_pass() mod_stats()* decision() end() end() end() * may be gen_features() version / target version / target program features new features pass list init running loop finish
  30. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Pass Constraints • The challenge for an arbitary order pass manager: only using valid combinations. • The machine learner can additionally learn about what options to avoid when selecting passes.
  31. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Flags • Today's optimisaction passes optimize for speed... again, why not have passes dedicated to energy? gcc -fenergy (or a more descriptive name)
  32. Copyright © 2013 Embecosm and University of Bristol Freely available

    under a Creative Commons license Community Involvement

More Decks by Simon Cook

Other Decks in Technology