-
Notifications
You must be signed in to change notification settings - Fork 1
2020 07 06 webex
#7/06/20 webex notes
The PR is against the MPI standard (https://github.com/mpi-forum/mpi-standard/pull/161). This was discussed/presented at the MPI Forum last week.
The main question is how to improve the wording starting with "It is erroneous to...". Jeff Squyres was suggested more positive wording.
The wording in this PR was a result of a user case with the MVAPICH crew. They were trying to bind a T var
to MPI_COMM_WORLD prior to invocation of MPI_Init. Dan wants to drill down into what the
Tools group actually wants to achieve.
Example of
MPI_Comm foobar_comm = MPI_COMM_WORLD
Dan cites 2.5.4 about named constants. What is allowed by MPI_T prior to MPI_Init? Pretty much anything.
Dan suggests asking what Jeff thinks about section 2.5.4. Simple examples shows about foo_comm thing works for Open MPI. Also go back to MVAPICH folks and see if they can "fix" the problem that started this ticket. If they say it can't be easily fixed, then rework proposed text to more accurately trap the issue.
Are there potential race conditions for freeing MPI_T cvar/pvars that were delayed bound to a handle like MPI_COMM_WORLD? Maybe do need an error code for case of late binding.
Martin was unable to attend so delayed further discussion till next week.