Skip to content

2015 09 24

Jump to bottom
Wesley Bland edited this page Sep 29, 2015 · 1 revision

Plenary Discussions

  • Presented Tickets 384 & 477
    • Lots of feedback for small clarifications
    • Only major objection was from Rolf about backward compatibility if MPI_COMM_WORLD has MPI_ERRORS_RETURN and MPI_COMM_SELF has MPI_ERRORS_ARE_FATAL
      • In this case, we would be changing existing behavior
    • Plan to do a full reading at the December meeting
  • ULFM Plenary
    • Wesley presented the updates to ULFM in the past year (RMA, dynamic processes, etc.)
    • George discussed the new additions to the ULFM implementation & presented the events going on in the community around ULFM (SC, tutorials, etc.)
    • Geoffroy presented work from one of his students to figure out how usable ULFM is for applications
      • Really would like to see MPI_COMM_IS_REVOKED

Working Group Discussions

  • Discussed feedback from 384 & 477 plenary
    • Decided that we wouldn't try to fix backward compatibility because the workarounds were a lot of extra work and not worth the effort
      • One workaround was to say that the default error handler for operations without a communicator is MPI_COMM_NULL.
      • Another workaround was to introduce a new error handler MPI_ERRORS_ARE_MPI_COMM_WORLD (or something) that would be the default on MPI_COMM_SELF which would just call the error handler for MPI_COMM_WORLD.
  • Discussed MPI_*_IS_REVOKED
    • There was some objection related to thread safety, but we decided that we could avoid this issues as long as the function will not return true until some thread has seen MPI_ERR_IS_REVOKED.
    • There was more discussion about whether the API is necessary, but we decided that it's worth adding since there have been so many requests and it removes the need for the user to replicate information that the implementation must track anyway.
Clone this wiki locally