Dear all, attached you find Howard's notes on a call we just had among the people involved with PRs currently in need of MPI_Info before init (MPI_T Events, Session, FT). Cheers, Marc-Andre -------- Forwarded Message -------- Subject: [mpiwg-sessions] notes from 7/29/19 webex Date: Mon, 29 Jul 2019 18:29:45 +0000 From: Pritchard Jr., Howard via mpiwg-sessions <[email protected]> Reply-To: MPI Sessions working group <[email protected]> To: MPI Sessions working group <[email protected]> CC: Pritchard Jr., Howard <[email protected]> Hello Folks Some notes from the 7/29/19 webex. Attending: Dan Holmes, Wesley Bland, Aurelien Bouteiller, Marc-Andre Hermanns, Howard Pritchard Dan turned on recording of the meeting. Dan pointed out a conflict about info keys – namely how they will be accessed before MPI initialization and after MPI finalization. Dan and Mark-Andre had discussed at ISC and were of the opinion that changes for MPI info keys should go into the sessions (PR 55) proposal. Sessions is being considered as a gating factor for 4.0. We need to figure out what is the difference between the info object changes in the. Aurelien says all the FT WG is wanting is to have MPI_INFO_ENV available before MPI initialization so the app can figure out what error handler has been set. Tools and Sessions don’t really care so much about this being available or not before MPI initialization. Dan discusses compromise that Sessions has taken. Wesley says FT needs MPI_Error_class and MPI_Error_string to be invocable prior to MPI initialization as well. Aurelien thinks this is in PR 50. Dan thinks the issue of argc/argv and how to fill in MPI_INFO_ENV is the key question. Aurelien thinks maybe add a new function to set MPI_INFO_ENV? Maybe a new ticket to handle this. FT group will work on this. Dan says we need to go through discrepancies between tickets. Sessions will need to pull in some of the FT adds to callable-prior-to-MPI Init to the table in Chapter 10 of PR 55. Mark-Andre raised some concerns about MPI_Info_get and error cases, but we think this will be covered by the setting-default-error handler before MPI initialization approach. Howard -- Howard Pritchard HPC-ENV Los Alamos National Laboratory -- Dr. rer. nat. Marc-André Hermanns IT Center Group: High Performance Computing Division: Computational Science and Engineering RWTH Aachen University Seffenter Weg 23 52074 Aachen Phone: +49 241 80-24381 [email protected] www.itc.rwth-aachen.de