Notes 2016 05 12

Jump to bottom Edit New page

Marc-Andre Hermanns edited this page May 12, 2016 · 2 revisions

Participants

Jean-Baptiste Besnard
James Custer
Marc-Andre Hermanns
Kathryn Mohror
Soren Rasmussen

Topics

Preparation for next face-to-face meeting
- Marc-Andre to get MPIR line-number ticket ready for a reading
- Marc-Andre to create PR for whitespace ticket to be decided upon by the chapter committee
- We won't push tickets #9 and #10 for next meeting, but maybe for Edinburgh
  - Soren wants to update ticket #9 with comments on where clarifications in the MPI_T chapter could be helpful
- As Kathryn and Marc-Andre won't be on site for the face-to-face meeting, we should organize proxies to lead the discussions
  - Marc-Andre to contact Anh, Martin, and Jeff for the debugger and performance tools sessions, respectively.
- It would be best if performance topics get scheduled in the morning for Marc-Andre & Jean-Baptiste to join
  - Have MPI_T Events first and then "The interface formerly known as PMPI".
MPI_T Events
- Marc-Andre updated the [wiki page](MPI_T Events).
- Callback requirements adopted from PERUSE should be OK for now. Some will need further discussion when the interface has developed further
- Open questions:
  - It should be safe to have the event interface to be initialized by the normal MPI_T_init_thread() call
    - Some events may not be available right away
    - Maybe provide an 'initialization callback' to be triggered once a callback becomes available?
  - As no specific callbacks are planned to be specified, how do tools discover the structure of the event information?
    - Have the query routine return a data layout description (like for MPI_Type_create_struct) to enable runtime-detection of data fields?
    - Tool will still need explicit (outside) help for interpreting the semantics of a data field
    - Type envelope could be used for sanity check
    - Information specific to all events could the specified in a header structure common to call events across all implementations (has to be high-level not to prescribe implementation detail)
      - Implementation may choose to indicate via unused_field that a specific header field is not used/supported
    - Tracing tools could copy blob to trace and interpret data post-mortem
    - Profiling tools can only act on information they know how to process
  - We should also port the multi-stage concept of PERUSE
    - a certain event could indicate which 'event set' is belongs to, and which other events may precede and succeed it
- Kathryn indicated that John Mellor-Crummey wanted to get in touch with us on asynchronous signal-safe callbacks