Fix `run` for `impi_rt` and `run_exports` for `impi-devel`

[leofang]

See Step 2 & 4 of https://github.com/conda-forge/mpi-feedstock/issues/11#issuecomment-1769101283. This fix is needed to avoid hitting issues like this.

cc: @conda-forge/intel_repack, @conda-forge/mpi, @conda-forge/mpi4py, @dalcinl @ax3l @beckermr

Checklist

• [x] Used a personal fork of the feedstock to propose changes
• [x] Bumped the build number (if the version is unchanged)
• [ ] Reset the build number to 0 (if the version changed)
• [x] Re-rendered with the latest conda-smithy (Use the phrase code>@<space/conda-forge-admin, please rerender in a comment in this PR for automated rerendering)
• [x] Ensured the license file is being packaged.

[conda-forge-webservices[bot]]

Hi! This is the friendly automated conda-forge-linting service.

I wanted to let you know that I linted all conda-recipes in your PR (recipe) and found some lint.

Here's what I've got...

For recipe:

• When defining a source/url please add a sha256, sha1 or md5 checksum (sha256 preferably).
• When defining a source/url please add a sha256, sha1 or md5 checksum (sha256 preferably).
• When defining a source/url please add a sha256, sha1 or md5 checksum (sha256 preferably).
• When defining a source/url please add a sha256, sha1 or md5 checksum (sha256 preferably).
• When defining a source/url please add a sha256, sha1 or md5 checksum (sha256 preferably).
• When defining a source/url please add a sha256, sha1 or md5 checksum (sha256 preferably).
• When defining a source/url please add a sha256, sha1 or md5 checksum (sha256 preferably).
• When defining a source/url please add a sha256, sha1 or md5 checksum (sha256 preferably).
• When defining a source/url please add a sha256, sha1 or md5 checksum (sha256 preferably).

[leofang]

@conda-forge-admin, please rerender

[oleksandr-pavlyk]

@beckermr Is there a way to retrieve conda package artifacts built in this PR for testing?

[leofang]

Yes. Give me a minute.

[leofang]

@conda-forge-admin, please rerender

[leofang]

Is there a way to retrieve conda package artifacts built in this PR for testing?

@oleksandr-pavlyk See the Azure page: https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=811656&view=artifacts&pathAsName=false&type=publishedArtifacts You can move your cursor to the right, and a 3-dots icon would show up for you to download the artifacts.

[leofang]

Btw, isn't Intel MPI binary compatible with MPICH by design?

I think so, based on the ABI Initiative, but I've never validated it myself.

[beckermr]

I've had enough trouble with that in various spots that it'd be safer to ship separate builds for now even though they may formally be Abi compatible.

[jeffhammond]

Please report all MPICH ABI issues to MPICH via GitHub. The ABI initiative is important and I have relied on it for a decade. I don't know what you're seeing that's not working but it's a surprise.

[dalcinl]

Btw, isn't Intel MPI binary compatible with MPICH by design?

I did not manage not figure out how we could take advantage of that in conda-forge. Maybe using features: [mpich_abi], but that's probably not enough.

I've had enough trouble with that in various spots

I can confirm that, at least for mpi4py (which uses almost all of the MPI 3.1/4.0 API), the MPICH ABI is working perfectly fine when replacing MPICH with Intel MPI 2021.10. This claim is under CI testing here.

PS: Debian/Ubuntu have had semi-broken the Linux MPICH ABI, simply because they use a different name for the MPI shared library than libmpi.so.12, and they do not provide a symlink. That that would only affect conda-forge users if they install MPICH external dummy package because they want to use the distro-provided MPICH.

[beckermr]

Sorry @jeffhammond! I should have been more specific. The issues i've seen are not ABI but are instead system/vendor specific quirks in how MPI is packaged and interacts with the scheduler on various sytems. This makes me hesitant to rely on run-time swapping. However, from the point of view of conda-forge's packages, this indeed may not be an issue.

[leofang]

@oleksandr-pavlyk let me know when you're done with local testing, and I'll revert the artifact change 🙂

[oleksandr-pavlyk]

@leofang I have downloaded artifacts, feel free to revert the change

[leofang]

@conda-forge-admin, please rerender

[leofang]

CI is green except for the linter (which I assume has been ignored for a long time here?) @conda-forge/intel_repack merge?

[leofang]

Thanks to all! 🙂

[ax3l]

Thank you all! :tada:

I think the same issue that we see with Intel MPI is also a problem for Intel OpenMP. Opened a new issue here: #59