-
Notifications
You must be signed in to change notification settings - Fork 652
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Limit threads usage in numpy during test to avoid time-out #4584
base: develop
Are you sure you want to change the base?
Conversation
Linter Bot Results:Hi @yuxuanzhuang! Thanks for making this PR. We linted your code and found the following: Some issues were found with the formatting of your code.
Please have a look at the Please note: The |
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## develop #4584 +/- ##
===========================================
- Coverage 93.87% 93.84% -0.04%
===========================================
Files 173 185 +12
Lines 21428 22494 +1066
Branches 3980 3980
===========================================
+ Hits 20116 21109 +993
- Misses 858 931 +73
Partials 454 454 ☔ View full report in Codecov by Sentry. |
Hello @yuxuanzhuang! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:
Comment last updated at 2024-09-09 23:10:45 UTC |
@orbeckst @IAlibay I now believe that these three adjustments can reduce the chance of timeouts:
I have re-run the GitHub CI three times consecutively, and they all pass. I have also patched it to #4162 that are more likely to time-out in yuxuanzhuang#6 and it also seems to work well. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thank you for looking into this pesky issues. If your changes improve our CI then I am all in favor.
I have one comment on how to set the number of used processes for pytest-xdist (see comments) but I am not going to block over it.
I am also shocked at how insensitive the ENCORE tests are: you changed them dramatically and they still pass without changing the reference values. Maybe a thing to raise on the mdaencore repo....
.github/workflows/gh-ci.yaml
Outdated
export OMP_NUM_THREADS=1 | ||
export MKL_NUM_THREADS=1 | ||
# limit to 2 workers to avoid overloading the CI | ||
export PYTEST_XDIST_AUTO_NUM_WORKERS=2 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Could this be set to auto, because that's what we are already using. Does auto not work correctly?
If auto does not work then I'd prefer we define a variable GITHUB_CI_MAXCORES or similar and then use it invoking pytest -n $GITHUB_CI_MAXCORES
. I prefer commandline arguments over env vars when determining code behavior because you immediately see what affects the command itself.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I was planning to find a way to find the default number of workers pytest will use and reduce it by one because the Ubuntu runner has 4 cores and Mac has 3 cores. but I ended up setting it to 2 and found the performance acceptable.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
ok
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Ubuntu runner has 4 cores and Mac has 3 cores.
It's unclear from this conversation that altering the value of -n
is actually resulting in any difference in the issue we're seeing with the the one multiprocessing test failing (i.e. the only cause of timeouts), could you confirm this please?
If it's not affecting things, then my suggestion is to stick with auto unless there's a substantial difference in performance that we've missed.
azure-pipelines.yml
Outdated
# limit to 2 workers to avoid overloading the CI | ||
$env:PYTEST_XDIST_AUTO_NUM_WORKERS=1 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The comment is confusing: if you set workers to 1 but say you limit to 2 workers. Change comment, otherwise same as above: perhaps just add to commandline args?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I forgot to correct the comment line when I realized Azure has only 2 cores so I changed # workers to 1.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
As above, is this actually affecting timeouts or is this some separate optimization?
@IAlibay would you please shepherd the PR to completion? |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for doing this @yuxuanzhuang
I'm unfortunately pretty swamped this weekend so I can only very briefly look at this today but I'll try to get back to you asap with a longer review and better explanation of where I think those env exports should go so they apply globally.
.github/workflows/gh-ci.yaml
Outdated
@@ -108,6 +108,13 @@ jobs: | |||
- name: run_tests | |||
if: contains(matrix.name, 'asv_check') != true | |||
run: | | |||
export OPENBLAS_NUM_THREADS=1 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
todo: there's another place to put this so it applies everywhere
@IAlibay with my PR shepherding hat on, just pinging here, no rush though 😄 |
@IAlibay just pinging here with my review coordination hat on. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Based on the PR main comment and the follow-up comment, it's unclear to me how this PR specifically fixes #4209
The timeout issues are specifically related to the multiprocessing test going over the time limit, whilst the optimizations to the encore tests are nice, could you please detail what is being done here that specifically changes the behaviour of the multiprocessing test?
Since this PR doesn’t really resolve the time-out issue, I will repurpose it to focus on reducing the test duration. |
fix #4209
Changes made in this Pull Request:
Similar to #2950, allowing numpy to make full usages of the resources clogs the machine and likely leads to the
occassionalfrequent time-out in multiprocessing-related testings.See the comparison between duration below; it takes a lot more time in test_encore for numpy calculations before.
new test duration
old test duration (https://github.com/MDAnalysis/mdanalysis/actions/runs/8869719078/job/24350703650)
gonna run it three times.
pass once
pass twice
PR Checklist
Developers certificate of origin
📚 Documentation preview 📚: https://mdanalysis--4584.org.readthedocs.build/en/4584/