ThreadMonitor : Add class for monitoring threads used for processes #5425
Conversation
|
|
||
| s = monitor.plugStatistics( random["outFloat"] ) | ||
| self.assertEqual( len( s ), IECore.tbb_global_control.active_value( IECore.tbb_global_control.parameter.max_allowed_parallelism ) ) | ||
| self.assertEqual( sum( s.values() ), 100001 ) |
There was a problem hiding this comment.
This test would be a bit more convincing if there was some check of the distribution - fiddling with a Poisson distribution a little, if the selection of threads to run were independent, it should be basically guaranteed that every bucket get at least 0.5 * 100000 / max_allowed_parallelism ... maybe lack of independence breaks this though? You've probably tested it in other ways so you know it works, I was just thinking that this would pass on [ 100001, 0, 0, 0, ... ]
There was a problem hiding this comment.
This does guarantee that each thread took at least 1 task (not 0), because otherwise there wouldn't be an entry for that thread at all (it's the processing of a task that creates the per-thread container in the first place).
I agree that we do want to test the quality of task distribution as well, but that's not really our concern when testing ThreadMonitor itself. I intend to use ThreadMonitor to provide such tests in ValuePlugTest and ComputeNodeTest in an upcoming TaskCollaboration pull request.
There was a problem hiding this comment.
I did completely mess up the process mask test though - that's fixed in f7341df.
|
Just to be clear, I don't see any problems with merging this ( though maybe I should be being more careful with stuff like the process mask test issues :P ) |
We don't want to clutter up the public API with Process subclasses.
johnhaddon
force-pushed
the
threadMonitor
branch
from
f7341df
to
a5f94ac
Compare
I intend to use this to beef up our test coverage for TaskCollaboration, as part of some other work I'm doing. Hopefully it is flexibile enough to be generally useful as a debugging tool, but I may have designed it too much with my initial use case in mind. If you have any thoughts on how it might be made more useful, feel free to let me know.