[WIP] This fix enables running tests on the CI with pytest-parallel

[akarbown]

Description

This proposal of using the pytest-parallel plugin could be the temporary WA for the slowdown in the tests (#20092) after removal of LLVM OpenMP.
Using pytest-parallel plugin (instead of pytest-xdist):

• needs the WA connected with python described here. Thus, the '--timeout-method=thread' option appeared in the command line and the fixes in the dataloader.py file.
• might result in the deadlocks, as some of the tests may not be thread safe. The one I've encountered is i.e test_ndarray_saveload and this is why I've mark that test as serial. There might be some others but it needs definitely longer testing.
• make the tests run faster than when using pytest-xdist.

For now, it's only the draft that will check/test that fix in the production. In the next steps I'll try to find the root-cause of the slowdown when MxNET is compiled with GNU OpenMP.

Comments

Where can I find any installation guide or other documentation that needs mentioning that change? I'm asking as I couldn't find any.

[mxnet-bot]

Hey @akarbown , Thanks for submitting the PR
All tests are already queued to run once. If tests fail, you can trigger one or more tests again with the following commands:

• To trigger all jobs: @mxnet-bot run ci [all]
• To trigger specific jobs: @mxnet-bot run ci [job1, job2]

---

CI supported jobs: [windows-cpu, miscellaneous, sanity, centos-gpu, unix-cpu, edge, clang, unix-gpu, centos-cpu, windows-gpu, website]

---

Note:
Only following 3 categories can trigger CI :PR Author, MXNet Committer, Jenkins Admin.
All CI tests must pass before the PR can be merged.

[akarbown]

@mxnet-bot run ci [unix-gpu]

[mxnet-bot]

Jenkins CI successfully triggered : [unix-gpu]

[Zha0q1]

hi @akarbown , what's the status of this pr? Thanks!

[akarbown]

I guess that @bgawrych fixed the issue in the following PR: #20367. However, there might be a potential to speed up the CI, but it will require more time and effort to figure out the details. Thus, I'm thinking if this might be useful and dig into it more (here the GPU cases), or withdraw it. @Zha0q1, what do you think?

[Zha0q1]

I guess that @bgawrych fixed the issue in the following PR: #20367. However, there might be a potential to speed up the CI, but it will require more time and effort to figure out the details. Thus, I'm thinking if this might be useful and dig into it more (here the GPU cases), or withdraw it. @Zha0q1, what do you think?

Thanks for the reply! Do we know if #20367 is the only cause of slowdown? I think faster ci will be favorable but I can monitor the CI runtime dash board for a few days and update back.