-
Notifications
You must be signed in to change notification settings - Fork 1
2021 03 15 webex
#03/15/21 webex notes
Attending: Howard Pritchard, Martin Schreiber, Martin Schulz, Jan Fecht, Rolf Rabenseifner, Aurelien Bouteiller
- https://github.com/mpi-forum/mpi-standard/pull/547
- recover where we were with bubbles and Martin Schreiber's app
Missed this discussion. It was merged in to the mpi-standard-rc-jun branch just prior to the zoom.
Coercive vs cooperative. We have first down with most all current MPI implementations, cooperative is a different story. Rolf asks whether the blue dots in the "Thoughts on global barrier slide" are in the WORLD process set. Green dots would be in a different process set. Dan and Howard don't think green dots can be in a "expanded" process set. Martin isn't so happy with this. Then are we back to versioning of process sets again? Martin Schreiber brings up issue of consensus if we do versioning. Something like MPI_Comm_agree from ULFM. Aurelien points out that at user level may not have ability to do collective ops, need to handle inside MPI implementation. MPI_Comm_create_from_group fails because different versions of process sets were used, handle internally or return an error to caller. Use a new procedure that goes directly from process set to communicator. Bypass group setup. Aurelien points out this is similar to MPI_Comm_shrink but more flexibility.
Aurelian had follow-on question about versions. How to pick which version to use? Does it have to use the most recent version? How to avoid deadlock if the groups being supplied to MPI_Comm_create_from_group were different versions, and hence different (possibly) groups. Dan thinks MPI/RT should be able to handle this. How's this versioning thing going to work with application initiated process sets like Jan Fecht described last week. Do we need to provide some kind of consensus function for these user-initiated process sets (if they could use version numbers). When should MPI give up with trying to get consensus on process set version numbers inside MPI_Comm_create_from_group. Aurelien says at some point fault events (that could impact process set formation) could cease and a consensus could be reached. Rapidly changing process sets? Dan thinks these wouldn't be usable - at least if you want to create a communicator based on these process sets.