Draft Agenda for MPI RMA Working Group: December 2015

Draft Agenda for MPI RMA Working Group: December 2015

Copied over from the September meeting. Needs update.

1. #396 - RMA needs new assertions for passive-target epochs. Current status (9/15): Proposed text available. Please review. Also needs input from implementations about how they might make use of the assertions. Only assertions that some implementation can exploit will be considered.
   

2. #397 - extend the use of MPI_WIN_SHARED_QUERY to all windows. Text proposed, straw vote 4/2/0. Current status (9/15): Additional rationale added about why Win_create_dynamic is excluded.
   

3. #401 - MPI_IN_PLACE in MPI_Get_accumulate. Can we add this as an Errata included in MPI 3.1? Current status (9/15): Votes in June had too many abstentions. Needs some specific use cases.
   

4. #416 - MPI Accumulate operation restrictions. Status (3/15): No Incorrect statement about MPI-2 removed. If the RMA group is favorable to this proposal, we need the exact text, including the advice to users and implementers to ensure that the performance opportunities are not lost. Current status (9/15): Needs specific proposal text and discussion of potential performance issues.
   

5. #434 - MPI 3.0 Errata: MPI Active Target synchronization requirements. See #456. Current status (9/15): needs to be rewritten to focus on clarity issue and consider whether the chapter should be reorganized to make this clearer.
   

6. #436 - MPI 3.0 Errata: MPI Unified window clarification. See #456. Current status (9/15): Add a Rationale or Advice to Users on how/why this "eventually synchronized" occurs.
   

7. #438 - Interpretation of MPI_WIN_BASE for shared memory windows. Current status (9/15): From the June meeting, new text is needed before this proposal can be further considered.
   

8. #439 - RMA Notification. There is an alternative for this without a ticket; the alternative is linked from this ticket. A new paper added to the ticket; More discussion needed. Current status (9/15): Need a clear use case and a discussion of why send-recv is not enough for this.
   

9. #459 - RMA sync ops with vector of windows. Current status (9/15): Some information added on performance impact. Also need to consider whether a global sync/flush_all across all created windows could be proposed instead.
   

10. #475 - Allow MPI_Win_sync at any time. New since 3/15. Current status (9/15): Needs additional discussion.
    

11. Nonblocking RMA epochs (see http://sc14.supercomputing.org/schedule/event_detail?evid=pap705 )
    

12. We still need the handle the issue raised by #456.
    

13. There are also remaining items from the 3.1 discussion; see https://docs.google.com/document/d/1MX6NRDaq6GxslM06q_COu5uFtJjPyCqtlh09g1pc-L8/edit
    

14. Require only one synchronization mode for a window (e.g., only passive, or only active). This would not be a backward compatible change, but it is not clear that any users need it, and it is difficult to implement correctly and efficiently.