homepage

multiprocessing's Condition uses a couple of semaphores to keep track of
processes which are sleeping and have woken up whilst waiting for the
condition. These counters however are only ever decremented in the
notify(_all) functions, via a loop which results in somewhat unexpected
behaviour where said functions take longer to run than expected.
The proposed patch simplifies notify(_all) functions such that time complexity
is still O(N) but crucially N is the number of currently sleeping processes
only (rather than number of wait() calls since last notify(_all) call).
Note: This also affects Event.wait() & Event.set() in the same fashion since a
Condition is used under the hood.
I've attached mp_sync_condition.patch based on "in-development" branch as of
98840:2028aed246c0. I have run unit tests on said commit (via debug build)
using:
./python -bb -E -Wd -m test -r -v -uall\
 test_multiprocessing_fork\
 test_multiprocessing_fork\
 server test_multiprocessing_spawn
Additionally I've run condition_test.py before & after to illustrate performance
of the proposed change as well as demonstrate the issue.

A few additional notes:
1) On an AMD E-450 under Ubuntu 14.04 x64 with a current in-development build, when running condition_test.py, Condition.notify_all() takes (measure using time.clock) 0.0003 with the patch and 0.3219 without it. Additionally the in-script timeit calls seem to suggest that the patch does not adversely affect Condition.wait().
2) Unit tests have also been run against test_multiprocessing_forkserver (typo in original message).
3) A scenario where this might become apparent would be to have many worker processes which have a loop along the lines of:
while not event.is_set():
 do_more_work()
 # Pause because the task is only supposed to happen every X seconds
 # or a wait is in effect due to an error. Do not want to use
 # time.sleep so loop is exited promptly rather than blindly sleeping.
 event.wait(some_amount)
other_work_or_shutdown()
If many processes call event.wait() with a timeout and the event is not set often (e.g. since it's a shutdown flag), a long-running program with many worker processes can end up having event.set() take a considerable amount of time.

I realise that having type bug type of "performance" means it's less likely to be looked at than if I had marked is as "behaviour" or "resource usage" (which would not be completely unfair, I think).
Also, the patch should be applicable to 2.7 and 3.2 through to 3.5 (despite versions having been removed from the bug report).

I've added a regression test for the proposed patch along the lines of the example script (i.e. fails before and passes with patch). Apologies if the test is a bit clumsy - maybe there is a more elegant way?