[mpiwg-sessions] Minutes for MPI Session meeting with special guest Rolf
martin.schreiber at univ-grenoble-alpes.fr
Mon Jun 26 10:57:42 CDT 2023
we intensively discussed (thanks Rolf!) the different options for the
changes in the MPI standard and I thought it's good to send this around
Attendees (until the end): Joseph, Rolf, Dominik, Martin (Schreiber)
I’m trying to summarize it here:
First of all, there’s an alternative A and B:
* Alternative A: Changing MPI Sessions (starting/finishing) might lead
to non-optimal utilization of MPI derived data types. This would be
communicated to the implementors and they are supposed to fix.
** Obviously, this puts more effort for the implementors to get it
** We discussed the realization of this. There have been some thoughts
that this would lead to overheads each time a communication is done to
check for this, but then came to the conclusion that these overheads
are always required: Data Types are defined without being specific to
communicators. Hence, they need to be optimized in an lazy way or a-
priory (or something in between). As a consequence, there would be
always overheads and we don’t expect any additional overheads for this
** This also follows the main principle of the MPI implementation to be
* Alternative B: Derived data types might not be optimal anymore after
changing MPI Sessions.
** This would put a lot of burden on the users.
** In particular, it’s not clear how to “recommit” datatypes in order
to optimize this.
** Users don't want to live with the situation that data types might
suddenly be less efficiently usable.
=> We (Rolf, Joseph, Dominik & Martin) all agreed to "Alternative A"
(which doesn't mean that this is final)
The second issue is about "with or without option C":
* With option C:
** Requires users to care about everything on their own to free
handlers at finalize.
** This breaks with libraries / tools to easily switch to sessions.
* Without option C:
** There would be just a recommendation to users to free handlers
** This recommendation could be incentivize by tools such as MUST
(pointing out the user that with Sessions, the handles, etc. should be
=> We all agreed that without option C would be the best solution.
Again, this is just our opinion and shouldn't be seen as a final
decision (unless you all agree).
All the best,
Martin & the others
Prof. Dr. Martin Schreiber
Applied Mathematics | High Perf. Scientific Comp. for PDEs
Université Grenoble Alpes | Lab. Jean Kuntzmann | INRIA, France
For Time-X Euro-HPC project:
TUM School of CIT, Technical University of Munich, Germany
More information about the mpiwg-sessions