-
Notifications
You must be signed in to change notification settings - Fork 0
Draft Agenda for MPI RMA Working Group: September 2016
William Gropp edited this page Sep 21, 2016
·
14 revisions
Working group issues
- #1 Attributes on MPI windows
Type of the attributes (integer, but what kind of integer)
Next steps: Bill to propose text - #2 Deprecate usage of MPI_MINLOC and MPI_MAXLOC with accumulate
Long discussion that veered into the rationale behind the restrictions on MPI_Accumulate an friends to only allow predefined operators (including these two).
Next steps: Still pending. Could benefit with a summary of the discussion, separating out the different issues.
MPI Forum issues
- #23 relax constraints on MPI_WIN_SHARED_QUERY
Allow this routine to be called for any (non-Dynamic) Window in case the implementation has used shared memory.
Next steps: make reading for a reading; need precise text and all issues addressed. - #24 new assertions for passive-target epochs
A collection of additional assertions to be used in the passive target case. Consider each independently.
Next steps: Take first 4 asserts and make them into a proposal with complete text. Remaining two need more discussion. - #31 MPI_COMM_TYPE_ADDRESS_SPACE
New comm_split_type type for MPI processes that share an address space (not strictly RMA, but impacts RMA chapter)
Next steps: Ignore while the hybrid group considers. - #36 MPI Accumulate operation restrictions
Provide a way to allow any combination of operations to the same location in RMA accumulate operations (possibly as the default)
Next steps: Pavan to provide text. - #37 Errata: MPI Active Target synchronization requirements
The infamous 434; the todo is to draft some clarifying text. Need to assign this task.
Next steps: Bill to draft text. - #38 Errata: MPI Unified window clarification
Clarify synchronization text.
Next steps: Text shold describe the problem in more detail (the specific shared memory issues). Pavan to lead text change with help from Bill and Rajeev and rest of RMA WG. - #39 Interpretation of MPI_WIN_BASE for shared memory windows
What does MPI_WIN_BASE mean for MPI_Win_allocate_shared? A preference was selected at a previous meeting, but next steps are needed.
Need complete and specific text for option 1. Ask Jeff to create. - #40 deprecate MPI_WIN_SYNC
Rely on the language rather than Win_sync
Next step: Close as will-not-do. - #41 extend MPI_WIN_SYNC for Fortran
An alternative to the above, intended for Fortran.
Next step: Close as will-not-do - #42 advice to users about MPI shared memory and C(++)11
As above, but targeted at C/C++ 11 users
Next step: Need specific text for the advice to users. Bill to do? - #46 fine-grain specification of RMA op+type in window info(x
Place holder for a proposals from the old wiki, but not yet in a form to discuss
Pavan and Jeff to handle; disposition is to close this one or greatly restrict it. - #47 RMA sync ops with vector of windows
Provide a way to use a single call to synchronize a vector of MPI_Win objects
Next steps: Link with discussion of non-blocking epochs and decide in that context. - #48 RMA Notification
Add notification to passive-target
Next steps: Add note that closed because (a) can use send/receive and (b) too tightly connected to specific hardware capabilities. - #49 Allow MPI_Win_sync outside of RMA epoch
As it says, though the proposed text is broader and probably needs to be tightened.
Next steps: Ask for use case; note that one approach is to call lockall immediately after creating the window and never unlocking until right before the win_free. - #59 RMA Notification
Mostly a duplicate to the RMA Notification above, but with more comments
Next steps: as #48 above. - #61 Errata for Passive Target RMA on shared memory
Clear errata on text with simple fix
Next steps: straw vote to consider as level 0 errata, though still needs to be added to errata page.