2016 12 12 webex

Meeting

• Attendees

  • Jeff Squyres
  • Wesley Bland
  • Martin Schulz
  • Ignacio Laguna
  • Brian Smith

• Slides for today

• Discussion points for today:

  • Tools!

Notes from meeting

• (Re-)Discussed the tools ideas from the last meeting (no overlap in people with the last meeting)

• Martin raises a good point: several of the ideas we mentioned last time could be considered syntactic sugar for a tool intercepting all MPI calls and tracking MPI handles itself. E.g., do we really need session-conversion functions and additional hooks if a tool intercepts all MPI calls -- including session creation, communicator creation, ...etc. In that way, a tool can make its own shadow communicators without needing session conversion functions.

• Martin agrees that the tools WG should take over some of these topics and discuss them further.

  • E.g., the hook conversation (for knowing when MPI is gone gone gone / valgrind clean) should move to the Tools WG, and be part of their events proposal.

• Intruiged by the whole "should MPI_T go to a local-state-based model?" question.

• Going through the 5 stages of grief: we're likely going to have to rename the sessions WG to something else (because of MPI_T sessions). :frown:

• Point was also made that in the last MPI Forum meeting, FT may save itself via sessions. I.e., each session in an MPI proc could use a different FT mechanism. This would allow you to have an app with multiple libraries, and they don't all have to agree on the FT mechanism.

  • It strikes me (Jeff Sq) that if the FT group uses this argument to save their proposal, then sessions will likely get used this way in the future, too -- i.e., that MPI libraries become much more run-time polymorphic / multimorphic, rooted in sessions (i.e., session A has behavior X, but sessions B has behavior Y).