-
Notifications
You must be signed in to change notification settings - Fork 1
2024 07 01 webex
- Forum related items
- Moving start time for fortnighly meetings
- local MPI session free discussion
- other
Dan said only discussion of Sessions came up in ULFM presentation on Thursday
The group discussed moving the meeting time to 1 hour earlier to accomodate some European participants. It was decided to try meeting at 8 AM US C(D/S)T.
Also discussed recording meetings. Howard said he'd see what LANL policy is on this. We may need to switch to Univ. Tenn. Knoxville zoom - which Aurelian offered to do.
This discussion is related to Need to define a local variant of MPI_Session_finalize
Note these were notes Howard tried to take and may definitely be missing points.
Group discussed non-blocking session "finalize" as an alternative to a local-only MPI_Session_free. Would be useful for FT cases. Also discussed was a session revoke operation. Aurelian noted one variant proposed would destroy everything associated with the session, there was a second but he couldn't recall details.
Dan and others pointed out some of the issues Rolf has raised about the MPI_Session_free are valid and there are lots of problems.
Tony pointed out the need to separate out progress (which is one of the things that makes session free problematic) from modern language concerns. Requests are currently the only way to "communicate" between the app and the MPI library. This led to discussion of callbacks.
Sessions attributes - or lack thereof - were discussed in the context of callback functions. Howard noted that we'd dropped these from the original sessions proposal upon advice from others that attributes were broken anyway and may be deprecated as part of MPI 5.
Returning to a non-blocking session free/finalize type function, we would need to add some of the test/wait functions to the list of routines that can be called after finalizing a session.