2017 08 16

Attendees

• UTK - Aurelien Bouteiller
• LLNL - Ignacio Laguna
• ORNL - Geoffroy Valle

Catastrophic errors

• Pavan wants to keep supporting this idea.
• Ignacio asked why we need to have an intermediate state between fully catastrophic and fully defined.
  • In POSIX, only these 2 states exist (i.e. SEGFAULT: will abort, vs ENOMEM, you can continue/retry and all is well defined)
• Aurelien remarks that
  • Posix does not matter itself as much with distributed state (easier problem to remain globally defined for Posix than for MPI as a distributed state).
  • Posix favors software engineering vs performance, MPI tends to take the opposite tradeoff
  • Therefore it makes sense that some errors in MPI will blow-up the MPI state
• A third idea is to return an error (vs aborting silently) only if the state remains defined
  • Would work for resource errors
  • Would prevent some good post-failure use cases (i.e. MPI is disfunctional from now-on because of an error/failure/resource limit, but the application has a chance to save the dataset or continue w/o MPI).

TUM proposal in the Sessions group

• Aurelien gave a very short overview.
• In the Session WG, a proposition from TUM to add grow-shrink operations has been discussed
• we would like to evaluate how/if that can do something for error management
• Will discuss this with more deps at a future meeting where attendance is higher.