Updated the documentation about shared memory.

[xbw22109]

This PR fixes errors regarding the shared memory file path in the document, and also updates two other parts of the shared memory documentation:

1. Corrected the error about the file path.

previous .. error:: notes:

shared-memory.rst:78 .. error:: TODO Is this correct?
shared-memory.rst:86 .. error:: TODO The filename above will certainly be wrong.
shared-memory.rst:91 .. error:: TODO The MCA param name above is definitely wrong.

2. Moved the explanation about Shared Memory Mechanisms from docs/launching-apps/localhost.rst to docs/tuning-apps/networking/shared-memory.rst.

previous .. error:: notes:

localhost.rst:41 .. error:: TODO This should really be moved to the networking section.

3. Fixed titles of the Shared Memory section.

previous .. error:: notes:

shared-memory.rst:4 .. error:: TODO This section needs to be converted from FAQ Q&A style
shared-memory.rst:5            to regular documentation style.

Additional Notes on Modification 1:

1. The default output path is /dev/shm with filename sm_segment.nodename.user_id.job_id.my_node_rank, not the session directory:

mca output:

code opal/mca/btl/sm/btl_sm_component.c:192:

code opal/mca/btl/sm/btl_sm_component.c:381:

2.About the session directory:

The "session directory" is a concept from OpenPMIx. It can be set manually via --pmixmca, or automatically by mpirun/prun or Open MPI (in singleton mode). While not central to this doc section, it's relevant because fallback paths may use the session dir. So, the MCA param is mentioned for completeness.

3.About single-copy methods and the backing file

After enabling single-copy methods, it might seem that a shared memory mapping file is no longer needed.
However, from reading the source code, I confirmed that it is still required.
This information is relevant for performance tuning, so I added it to the documentation.

[github-actions]

Hello! The Git Commit Checker CI bot found a few problems with this PR:

aaf5f27: Updated the documentation about shared memory.

• check_signed_off: does not contain a valid Signed-off-by line

Please fix these problems and, if necessary, force-push new commits back up to the PR branch. Thanks!

[xbw22109]

@bosilca Hello, this is my first pr and I need to be approved for workflow execution.

[edgargabriel]

Question: do we need to add in here something about the smsc/accelerator component that will be available in 6.0?

[rhc54]

Might be worth explaining a tad more about the session directory and this shmem backing file. The reason we created a session directory (way back in the beginning of OMPI - predates PMIx by more than 10 years) was to provide a location where we could collect all files OMPI creates, thus providing a simple way to clean them all up upon termination. The daemon just whack the entire session directory tree after the job completes (normally or abnormally) before it exits. We changed that for the shmem backing file (to use /dev, if it exists) due to size limitations (IIRC).

The problem with moving that file is that we now lose the ability to ensure cleanup. If the process itself abnormally terminates, the daemon that shepherds it has no idea that the shmem backing file was created somewhere outside the session directory - and therefore has no way to clean it up. This has generated some user complaints, but I'm not sure of the best solution. Perhaps we should use the PMIx_Job_ctrl API to register the backing file for cleanup upon termination when it isn't in the session directory? Or maybe that has already been done?

Regardless, might be worth noting that in the docs somewhere. If we don't register the cleanup location and your app abnormally terminates, then you may need to ensure (somehow) that you go back and remove those shmem backing files. Otherwise, they will just build up over time (as other jobs abnormally die), consuming disk space.

[bosilca]

We can certainly do work to improve this documentation., But as a first contribution this is already awesome as it is.

[jsquyres]

Suggested change

	filesystem such as /tmp or /dev/shm (default: (linux) /dev/shm,
	filesystem such as ``/tmp`` or ``/dev/shm`` (default: (linux) ``/dev/shm``,

[jsquyres]

Is this the right MCA param name? I didn't think we had any orte_ primary names any more -- is that an alias?