Skip to content

2018 01 31

Jump to bottom
Wesley Bland edited this page Feb 3, 2018 · 2 revisions

Attendees

  • Intel - Wesley, Rob
  • Argonne - Ken, Yanfei
  • Sandia - Keita
  • Auburn - Nawrin
  • ORNL - Geoffroy
  • UTK - Aurelien, George
  • UTC - Tony

FT Interoperability

  • How do disjoint groups agree on the FT model before connect/accept?
    • Tony suggested that restricting things to require both sides of an intercommunicator to have the same FT model is too restrictive.
      • If the two sides never create an intracommunicator, we could just break the intercommunicator when a fault occurs.

MPI_COMM_ISHRINK

  • The semantics of MPI_COMM_ISHRINK are confusing because there is no MPI_COMM_ISPLIT to piggy-back in the same way we do between MPI_COMM_SHRINK and MPI_COMM_SPLIT.
    • Tony will work on a separate proposal to create non-blocking communicator creation variants of all functions.
  • Copying attributes is difficult because there is no way to get a list of them.
    • Could add a function to copy attributes, but this doesn't help the problem of knowing which attributes to copy.
    • The user still has to know which attributes apply so they will need to copy attributes themselves.
  • MPI_COMM_ISHRINK won't copy attributes, info keys, name, topology, etc.

February F2F Plans

  • Error Handlers - Vote
  • ULFM - Reading
  • Catastrophic Errors - Plenary and possible reading
  • FT Interoperability - Plenary
  • 1/2 working group time
Clone this wiki locally