Parallelizes MDAnalysis.analysis.msd

[tanishy7777]

Fixes #4676

Changes made in this Pull Request:

• Added the split-apply-combine technique to parallelize the MDAnalysis.analysis.msd.EinsteinMSD
• Added boilerplate fixture(s) to testsuite/analysis/conftest.py, analogous with existing ones
• Added a client_EinsteinMSD, fixtures to all tests using in testsuite/MDAnalysisTests/analysis/test_msd.py, and modified the way run() method is called to run(**client_EinsteinMSD)

PR Checklist

• Tests?
• Docs?
• CHANGELOG updated?
• Issue raised/referenced?

Developers certificate of origin

• I certify that this contribution is covered by the LGPLv2.1+ license as defined in our LICENSE and adheres to the Developer Certificate of Origin.

---

📚 Documentation preview 📚: https://mdanalysis--4896.org.readthedocs.build/en/4896/

[pep8speaks]

Hello @tanishy7777! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2025-01-20 21:03:12 UTC

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 93.41%. Comparing base (7fb3534) to head (ba68a93).

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #4896      +/-   ##
===========================================
- Coverage    93.42%   93.41%   -0.01%     
===========================================
  Files          177      189      +12     
  Lines        21865    22945    +1080     
  Branches      3079     3079              
===========================================
+ Hits         20427    21435    +1008     
- Misses         986     1059      +73     
+ Partials       452      451       -1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[tanishy7777]

Just wanted to remind you that this is ready to be merged I think. Please do so at your convenience.
@RMeli @orbeckst

[orbeckst]

Thanks for your work. I'm currently quite busy, so might not be able to review in the next few days. Please be patient.

[orbeckst]

@talagayev / @marinegor can you have a look at this PR, please?

[talagayev]

Checked the code and ran locally, looks all good.

https://github.com/tanishy7777/mdanalysis/blob/18a2e516d914f6dc438b409b403be2a1a3429e77/testsuite/MDAnalysisTests/analysis/test_msd.py#L155

Here @tanishy7777 you could also add the **client_EinsteinMSD to cover the parallelization in test_simple_start_stop_step_all_dims and test_fft_start_stop_step_all_dims, but here I would rely on what @orbeckst suggests if it needs to have the **client_EinsteinMSD or not.

From my side it looks good, good job @tanishy7777 :)

[tanishy7777]

Checked the code and ran locally, looks all good.

https://github.com/tanishy7777/mdanalysis/blob/18a2e516d914f6dc438b409b403be2a1a3429e77/testsuite/MDAnalysisTests/analysis/test_msd.py#L155

Here @tanishy7777 you could also add the **client_EinsteinMSD to cover the parallelization in test_simple_start_stop_step_all_dims and test_fft_start_stop_step_all_dims, but here I would rely on what @orbeckst suggests if it needs to have the **client_EinsteinMSD or not.

From my side it looks good, good job @tanishy7777 :)

Thanks a lot for reviewing my PR, will wait for the suggestions as you mentioned.

[tanishy7777]

From my side it looks good, good job @tanishy7777 :)

Also could you please review this PR #4884 its pretty similar or tell me if it needs any more work to be done. Thanks again

[talagayev]

From my side it looks good, good job @tanishy7777 :)

Also could you please review this PR #4884 its pretty similar or tell me if it needs any more work to be done. Thanks again

Hey @tanishy7777, yes the PR is similar, I can take a look at it as well.

[hmacdope]

Blocking here, I just need to check the implementation IIRC there is a reason MSD algo itself is non-parallelisable, but may not apply if only the collection of particle positions is parallelised.

[tanishy7777]

Blocking here, I just need to check the implementation IIRC there is a reason MSD algo itself is non-parallelisable, but may not apply if only the collection of particle positions is parallelised.

Like the tests were passing so I thought it has been parallized

[marinegor]

hi @tanishy7777, sorry for long review -- many life things got in the way.

I think you're on a good path, I mentioned few minor things in the comments.

Main action items:

• move out hacky @staticmethod def f(arrays): pass out of the class
• using datafiles in MDAnalysisTests (imported on top of test_msd.py, check that parallelized run produces exactly the same results as non-parallelized one (add code snippet to comments that anyone can run to check, and its results)
• add this check as a test (if it's too slow, we can always mark it this way and not run by default)

Please ask if you have questions, and ping me here if I don't reply for more than 48 hours.

[marinegor]

move this outside of the class and explicitly name it something like __noop to make sure no one uses that:

def __noop(arrs) -> None: pass

[tanishy7777]

move this outside of the class and explicitly name it something like __noop to make sure no one uses that:

def __noop(arrs) -> None: pass

I tried renaming it to __noop and moving it outside the class, but this leads to NameError due to how name mangling works in python. I renamed it _noop instead, hope that isnt a problem?

we could also do something like this if we want to keep the name as __noop

from MDAnalysis.analysis import msd

def _get_aggregator(self):
    return ResultsGroup(
        lookup={
            "_position_array": ResultsGroup.ndarray_vstack,
            "msds_by_particle": msd.__noop,
            "timeseries": msd.__noop,
        }
    )

Not sure if this is as clean as above method though.

[marinegor]

Since there are indeed concerns that the algorithm is indeed parallelizable, I would suggest you explicitly test for that: namely, ensure that all results.<something> attributes are exactly the same, regardless of how you run your analysis. You can do this yourself first, and than later add it as one of the tests here.

I should also say that if you make a convenient way to test that, it'd be nice to have it added to other parallelization tests, to ensure additional correctness.

[tanishy7777]

hi @tanishy7777, sorry for long review -- many life things got in the way.

I think you're on a good path, I mentioned few minor things in the comments.

Main action items:

• move out hacky @staticmethod def f(arrays): pass out of the class
• using datafiles in MDAnalysisTests (imported on top of test_msd.py, check that parallelized run produces exactly the same results as non-parallelized one (add code snippet to comments that anyone can run to check, and its results)
• add this check as a test (if it's too slow, we can always mark it this way and not run by default)

Please ask if you have questions, and ping me here if I don't reply for more than 48 hours.

Hey, sorry for the late response. I had semester exams so I was quite busy the last 2 weeks. Will start working on this soon!

[tanishy7777]

@marinegor

As suggested in #4896 (comment)
I have added the tests for checking the equivalence of the msd algorithm with both the serial and multiprocessing backends.

After running the tests, the results seem to be equal across both backends. Could you please review the implementation and confirm if this is correct?

Also, is this a good way of adding this test? Should I add it for all other methods that were parallelized?

I have added the following test for comparing the result.attributes
This test is passing.

Thank you!

[tanishy7777]

@marinegor

As suggested in #4896 (comment) I have added the tests for checking the equivalence of the msd algorithm with both the serial and multiprocessing backends.

After running the tests, the results seem to be equal across both backends. Could you please review the implementation and confirm if this is correct?

Also, is this a good way of adding this test? Should I add it for all other methods that were parallelized?

Thank you!

@marinegor just a gentle reminder

[orbeckst]

@marinegor tests pass, codecov is good, linter is happy – could you please have another look?