2016 12 20

FTWG Con Call

Agenda

1. Recap Forum Meeting
2. Go over existing tickets/work
3. ULFM

Notes

• #1/#3 - Error Handlers
  • Handful of edits requested and made
  • Still discussing return codes for error handlers
• #28 - Catastrophic Errors
  • More small tweaks requested and made
• Error Handler Greenfielding Work
  • Decided to abandon the effort as is due to complexity of design
  • Will pursue error handler inheritance separately
• ULFM
  • Recapped discussion about ULFM problems raised during reading.
    • https://github.com/mpi-forum/mpi-issues/issues/20#issuecomment-265773234
  • When including ULFM, there are four competing models for FT:
    1. ULFM
    2. ULFM + auto recovery
    3. Communicators with holes
    4. Reinit
  • Problem with holes model
    • Increases non-failure cost because of algorithm changes
  • Problems with ULFM
    • Cleaning up MPI state is a large problem (or uses PMPI) if reinit-like behavior is desired.
    • Revoking globally is not currently possible
    • Error handlers are still insufficient for fault tolerance
      • Communicators may be too small for specifying recovery model
    • C-style exception handling is insufficient, but cannot easily be extended and remain compatible with Fortran.

Next Steps

• Investigate fleshing out the three main options (excluding holes) and discussing how they would work together.
  • ULFM: Aurelien
  • ULFM + auto recovery: Wesley
  • Reinit: Ignacio
• Once we have a good idea of some more details of each proposal, consider how they would interact and overlap.
  • Would there be anything in one proposal that would prevent the other.
• For now, don't allow more than one model in a single MPI instance.
  • This might be relaxed with sessions to allow each session to choose an FT model.
  • Work through the exercise of a sessions-style app and library which use different FT models.