-
Notifications
You must be signed in to change notification settings - Fork 1
2022 01 24 webex
#01/10/22 webex notes
Attending: Howard Pritchard, Thomas Hines, Dan Holmes, Isaias Urena, Trupeshkumar Patel, Brian Smith, Martin Schreiber, Aurelian Bouteiller
- Recently opened issues for MPI 4.1 ** https://github.com/mpi-forum/mpi-issues/issues/534 ** https://github.com/mpi-forum/mpi-issues/issues/535
- PMIx resource allocation thread – see https://github.com/pmix/pmix-standard/issues/386
- Followup to Dominik’s presentation on 1/10/22 – see https://www.martin-schreiber.info/pub/tmp/2022_01_11_dominik_huber_dynamic_sessions_interface_annotated.pdf
- Other
Dan thinks we may need a new subsection for 535. 534 (the second bullet) is more problematic. May not be able to return MPI_GROUP_NULL since its a pointer value. Maybe do return MPI_GROUP_EMPTY? Or should we instead return MPI_ERR_ARG? No one objected to striking the sentence and raising the error. This would be an errata. Its introducing a backwards incompatibility.
Discuss what MPI brings as a value add to what PMIx provides. Conceptual difference between MPI "thinking" and how PMIx handles resources. Compared to starting processes using mpiexec, spawn is much less standardized. Example of job script resource specification and then call an mpiexec to request number of MPI processes to spawn. Example of MPC MPI case too - using threads vs processes for MPI execution vehicles. the PMIx_Allocation_request maybe serves as an approach to standardize such an approach.
We do still want to dive into this more and see what we may want to have different from what he describes in his thesis.
Default error behavior - related to PR https://github.com/mpi-forum/mpi-standard/pull/629. Aurelian will get FT related PR open on mpi-standard.
- Howard will check Open MPI's implementation to see what it will do.
- Howard will consolidate feedback on mail list to Dominik's thesis for discussion next week or the one thereafter.