Adaptive scheduler #85
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
petervdonovan
force-pushed
the
scaling-wrt-threads-runtime-experiments
branch
2 times, most recently
from
d32053a to
56bc3fb
Compare
A timer event seems to be occasionally dropped in TimeLimitThreaded.
TODO: compute_number_of_workers should be reduced correspondingly?
Marked improvement observed locally for RadixSort, Counting.
This optimization is lifted directly from the NP scheduler in response to visible differences in the number of cycles spent accessing mutexes. Clear improvement is observed locally for the Counting benchmark.
This is another step in the direction of the NP scheduler. The condition for level advancement should be that _everyone_ is sleeping; otherwise, we cannot optimize out the lock for level advancement without introducing a race condition.
Some of our benchmarks need either worker affinity OR use of just one worker, so that a particular reaction is always executed by the same worker. Perhaps worker affinity is easier to add. However, it must be combined with some sort of load-balancing strategy.
This is a very crude implementation that should be improved upon.
*Although this fixes a race condition, I have not tested if this fixes "the" race condition. The bug manifested itself so rarely that it was infeasible to reproduce.
Symptom: A thread goes to the event queue to wait for the next event and ends up waiting forever (while the other threads wait for work). This is possible only in SleepingBarber due to use of physical actions -- not in other benchmarks. This bug is not caught by our tests.
The level counter is part of the predicate associated with the condition variable. It is unsafe to change without acquiring a mutex.
If I am right, expected trials required to go from an intermediate number of workers to an extreme (1 or the maximum) number of workers was previously quadtratic in # of workers (cuz random walk) -- no good. This should make it (log(n))^2.
This kind of bug reflects lack of care on my part in finding a clean implementation. Some cleanup will be necessary if this is ever to be merged.
I did not observe a race condition, but in principle I think there was one. For this reason, this change could not be justified by empirical results.
Again, this is another optimization borrowed from the NP scheduler. The difference is that the NP scheduler is designed such that this optimization needs no explicit code to handle it; there, it just works.
petervdonovan
force-pushed
the
scaling-wrt-threads-runtime-experiments
branch
6 times, most recently
from
f7b48fd to
e2ee83f
Compare
petervdonovan
force-pushed
the
scaling-wrt-threads-runtime-experiments
branch
from
e2ee83f to
d682724
Compare
By "unrelated" I mean, "unrelated to the hard-to-reproduce race condition that is manifesting itself in LoopDistributedCentralized in CI."
petervdonovan
force-pushed
the
scaling-wrt-threads-runtime-experiments
branch
2 times, most recently
from
504885c to
5f3b2f0
Compare
petervdonovan
force-pushed
the
scaling-wrt-threads-runtime-experiments
branch
from
5f3b2f0 to
c19bb3d
Compare
petervdonovan
force-pushed
the
scaling-wrt-threads-runtime-experiments
branch
from
b6cb486 to
d23d17f
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
To clarify -- this is the scheduler that previously had the vague name "heuristic."
For details, please see PR #1207 in the main repository.