[Mpi-forum] why do we only support caching on win/comm/datatype?

[Holmes, Daniel John]

Hi Jeff,

When adding session as an object to MPI, a deliberate choice was made not to support attributes for session objects because “attributes in MPI suck”.

This decision was made despite the usage (by some tools) of “at exit” attribute callbacks fired by the destruction of MPI_COMM_SELF during MPI_FINALIZE in the world model and the consequent obvious omission of a similar hook during MPI_SESSION_FINALIZE in the session model (there is also no MPI_COMM_SELF in the session model, so this is not a simple subject).

Removal of attributes entirely – blocked by back-compat because usage is known to exist.
Expansion of attributes orthogonally – blocked by “attributes in MPI suck” accusations.

Result – inconsistency in the interface that no-one wants to tackle.

Best wishes,
Dan.

From: mpi-forum <mpi-forum-bounces at lists.mpi-forum.org> On Behalf Of Jeff Hammond via mpi-forum
Sent: 16 January 2023 14:40
To: MPI Forum <mpi-forum at lists.mpi-forum.org>
Cc: Jeff Hammond <jeff.science at gmail.com>
Subject: [Mpi-forum] why do we only support caching on win/comm/datatype?

I am curious if there is a good reason from the past as to why we only support caching on win, comm and datatype, and no other handles?

I have a good use case for request attributes and have found that the implementation overhead in MPICH appears to be zero.  The implementation in MPICH requires adding a single pointer to an internal struct.  This struct member will never be accessed except when the user needs it, and it can be placed at the end of the struct so that it doesn't even pollute the cache.

I wondered if callbacks were a hidden overhead, but they only called explicitly and synchronously, so they would not interfere with critical path uses of requests.

https://github.com/mpi-forum/mpi-issues/issues/664 has some details but since I do not understand how MPICH generates the MPI bindings, I only implemented the back-end MPIR code.

It would make MPI more consistent if all opaque handles supported attributes.  In particular, I'd love to have a built-in MPI_Op attribute for the function pointer the user provided (which is similar to how one can query input args associated with MPI_Win) because that appears to be the only way I can implement certain corner cases of MPI F08.

Thanks,

Jeff

--
Jeff Hammond
jeff.science at gmail.com<mailto:jeff.science at gmail.com>
http://jeffhammond.github.io/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpi-forum/attachments/20230116/f59d6a61/attachment.html>