Skip to content
Howard Pritchard edited this page Aug 3, 2020 · 1 revision

8/2/18 webex notex

Reviewing Ralph Castain's PMIx group extensions

Attending: Vishnu, Dan, Howard

Discuss rearranged chapter 10 (now has MPI Init, etc.)

https://github.com/mpiwg-sessions/mpi-standard/tree/RearrangeInitialisationTextForSessions

Group agreed to have a new 10.4 (moving current 10.4 to 10.5) where we'd have a whack at Sessions.

  • MPI_Session_create/init (we're not entirely happy about this)
  • MPI_Session_get_names
  • MPI_Session_get_info
  • MPI_Group_create_from_session
  • MPI_Comm_create_from_group(? no probably leave where it was in comm chapter)

Dan and Tony discuss options for creating comm from group - do we really need all the direct file from group and window from group. Dan gives rationale for why we would like to do this - it only helps if we are somehow saving some resource, memory, etc. by going straight from the group to the window, file. More obvious for the group to communicator with topology. The user can still create a communicator from group, then get file or window from this communicator. Tony is for the group to resource direct without an intermediate communicator.

Talk about examples of workflows with RMA, non-volatile memory, etc.

Dan suggests having intro text additions for thing from group about where the group came from (reference back to groups from other groups, etc. or from sessions). Discussion of need for a runtime for this. Example of CM and SM in Infiniband and its similarities.

Dan on vacation next two weeks, but we'll try to meet on the 9th, may skip 16th.

Tony pushed having something about sessions in the SC'18 MPI BOF. Didn't have anything about it at ISC BOF though.

Discussion on how to get Sessions into a 4.0 standard. We'd need to push out release of 4.0 to get sessions pulled in. - synergy discussion. need to get a proto-type and papers published.

Clone this wiki locally