Notes 2016 09 21

Participants

MPI Forum members

Notes

• MPI_T_event_get_info
  • It should return more semantic information than just the list of datatypes
  • We could use MPI_T_enum to provide a description of each data element in the event data
  • Should we get a per-event version number for each event?
• MPI_T_event_get_handle_info
  • Forum suggested this is not needed, as the tool should track the info if it needs to
• MPI_T_event_is_registered
  • Extra call not needed
  • Other calls can return MPI_T_ERR_INVALID_HANDLE, if a handle is invalid
  • Errors should not be fatal
• MPI_T_event_copy
  • Call should allow non-contiguous (padded) data to copy
  • Size to be communicated through info call
  • Get a buffer element that can later be read via the read call
• MPI_T_get_timestamp
  • Explicit call to return the time associated with an event
  • Gets rid of explicit notion of 'synchronous' and 'asynchronous' events
    • For synchronous events the call will generate a time stamp (using the new high-resolution timer)
    • For asynchronous events the call returns the time stamp generated when it was recorded
• Callbacks vs. Polling
  • What are the restrictions to call backs
    • No further calls into MPI other than MPI_T_event_read/copy
  • Callbacks may be on the critical path
  • Polling-based interface may be able to take calls off the critical path
    • MPI runtime always writes event data to buffer (incl. timestamp) and tool queries it at less critical times (e.g., within selected PMPI wrappers)
    • Should tool/user be able to provide additional memory?
      • Do we need different buffers for different tools?
      • Do we need sessions so different tools do not interfere with each other?