Skip to content

2018 01 24

Jump to bottom
Wesley Bland edited this page Jan 24, 2018 · 1 revision

Attendees

  • Intel - Wesley
  • UTK - Aurelien, George
  • Auburn - Nawrin
  • Argonne - Ken

Catastrophic

  • Discussed PR #9
  • What if we screw up the matching order?
    • Should we return MPI_ERR_OTHER forever?
    • Does this put us in an undefined state?
    • We decided to support returning an error forever even though it has a potential performance impact because in practice it will be a branch that is either predicted perfectly or we don't care about the missed branch penalty.
  • Should we deprecate MPI_ERR_INTERN? MPI_ERR_UNKNOWN?

FT Interoperability

Slides

  • How do we get all connected processes to participate in the call
    • e.g. Spawned processes free their intercommunicator but retain their MPI_COMM_WORLD while other spawned processes keep the intercommunitor. In this case, all of the spawned procs are connected but some might not know it.
    • A solution is to put all of the work on the user and remove the collective semantics.
      • Make the call local and the user should do their own synchronization beforehand (if necessary).
Clone this wiki locally