Notes 2018 09 21

Plenary Discussion on MPI_T UIDs

• discussion on equivalency

  • adjust APIs to deal with sets of UIDs to handle cases where a variable adopts a foreign UIDs without changing semantics (e.g., implementation initialy uses own UID and changes to one from different vendor [e.g., MPI Forum] later on)

• Straw poll

  • 6 as it is here
  • 10 adding equivalency
  • 0 like the idea and go back to drawing board
  • 0 don't like the idea at all

WG Discussion on MPI_T Events

• Info Argument on get_info & handle_alloc

  • Use case for info object
    • communicate minimum and maximum saftey levels of tools
    • communicate maximum level safety of tool callback
  • Available Options:
    1. Get rid of info
       • bad
       • loss of information and communication btw. implementation and tool
    2. Add text to clarify that 'info' can only be used after Init
       • bad
       • how to handle return of info parameters?
       • a lot of events will be registered before MPI_init
    3. have additional calls '_with_info'
       • bad
       • No real difference to (2)
       • drawback of additional calls
    4. have explicit arguments (instead of info)
       • handle_alloc gets passed a cb safety level with maximum support of tool
         • if recommended or stronger level is passed alloc is guaranteed to succeed
         • if passed weaker that recommended call may succeed or return NOT_SUPPORTED
       • event_get_info call returns the "recommended" cb_safety level for a type
         • level is guaranteed to be supported
       • recommended level should reflect the level that an implementation is comfortable to support without dropping "too many" event instances
         • but that depends on what the user code does (message bursts, etc.)
    5. have separate calls (like for UID)?
       • return upper/lower bound in separate call?

• How to handle dropped count?

  • Include count into normal callback?
  • Count would be a minimum count (as others might have been missed)
  • Dropped callback safety should be the same
  • Count should be in normal CB as the number of dropped invocations is tied to the source, i.e., count is per source!