Author: Antonio Cuni <anto.c...@gmail.com> Branch: extradoc Changeset: r5263:5acdd3db487c Date: 2014-05-24 12:10 +0200 http://bitbucket.org/pypy/extradoc/changeset/5acdd3db487c/
Log: STM slides diff --git a/talk/pycon-italy-2014/talk.rst b/talk/pycon-italy-2014/talk.rst --- a/talk/pycon-italy-2014/talk.rst +++ b/talk/pycon-italy-2014/talk.rst @@ -183,6 +183,8 @@ - ~80% of numpy implemented + * 2336 passing tests out of 3265 + * http://buildbot.pypy.org/numpy-status/latest.html - just try it @@ -237,13 +239,114 @@ - 3x faster than CPython -STM ---- +The future: STM +---------------- -TODO +- Software Transactional Memory -Q&A ---- +- Strategy to solve race conditions -Any question? +- "Finger crossed", rollback in case of conflicts +- On-going research project + + * by Armin Rigo and Remi Meier + +STM semantics +------------- + +- N threads + +- Each thread split into atomic blocks + +- Sequential execution in some arbitrary order + +- In practice: + +- Parallel execution, conflicts solved by STM + + +Unit of execution (1) +--------------------- + +- Atomic blocks == 1 Python bytecode + +- Threads are executed in arbitrary order, but bytecodes are atomic + +- ==> Same semantics as GIL + +- "and this will solve the GIL problem" (A. Rigo, EuroPython 2011 lighting talk) + +Unit of execution (2) +---------------------- + +- Larger atomic blocks + +- ``with atomic:`` + +- Much easier to use than explicit locks + +- Can be hidden by libraries to provide even higher level paradigms + + * e.g.: Twisted apps made parallel out of the box + +Race conditions +--------------- + +- They don't magically disappear + +- With explicit locks + + * ==> BOOM + + * you fix bugs by preventing race conditions + +- With atomic blocks + + * ==> Rollaback + + * Performance penalty + + * You optimize by preventing race conditions + +- Fast&broken vs. Slower&correct + + +Implementation +--------------- + +- Conflicts detection, commit and rollaback is costly + +- Original goal (2011): 2x-5x slower than PyPy without STM + + * But parallelizable! + +|pause| + +- Current goal (2014): 25% slower than PyPy without STM + +- Yes, that's 800x faster than original goal + +- mmap black magic + +Current status +--------------- + +- Preliminary versions of pypy-jit-stm available + +- The JIT overhead is still a bit too high + +- Lots of polishing needed + + +Thank you +--------- + +- http://pypy.org + +- http://morepypy.blogspot.com/ + +- http://antocuni.eu + +- Any question? + _______________________________________________ pypy-commit mailing list pypy-commit@python.org https://mail.python.org/mailman/listinfo/pypy-commit
