Meeting Notes
June 11 - June 14, 2018
| Logistics | Agenda | Presentations | Attendance | Votes | Notes |
MPI Forum Working Groups
Various working groups met. Notes can be found on their respective wikis:
- Sessions
- Tools
- Persistence
- Hardware Topology
- Fault Tolerance
- Collectives (and groups and communicators)
- Hybrid
MPI Forum Plenaries
Upcoming Meeting Dates and Locations
- March 4 - 9: Chattanooga, TN
- May 28 - 31: Chicago, IL
- Shifted to Tuesday - Friday to avoid Memorial Day holiday in the US
- September 5 - 9: Zürich, Switzerland
- Thursday - Saturday before EuroMPI to avoid Labor Day holiday before and conflicts after the conference
- December 9 - 12: Albuquerque, NM
Introduction
Martin began the plenary portion of the F2F with introductions and working group updates.
Reading: Errata: broken link for rfc1832
- No objection
Reading: Errata: Possible ambiguity on size of array_of_indices and array_of_statuses?
- No objection
Reading: Errata: MPI_STATUS_SET_CANCELED F08 binding: flag param intent
- No objection
Reading: Errata: Behavior of MPI_Dims_create when ndims is zero
- Open MPI and MPICH are already doing this.
- No objection
Reading: Errata: Clarification to allow MPI_PROC_NULL as neighbor in (distributed) graph topologies
- Rich raised performance concerns that implementations which do not use the existing point-to-point code to implement this (which MPICH and Open MPI already do) will be impacted because they will need to shuffle around buffers and do additional checks for
MPI_PROC_NULL. - Agreed that this is not an errata change and should go through the full voting procedure.
- This will give everyone a chance to discuss this before the first vote in September.
Reading: No-No: Noncatastrophic Errors
- No objections
Reading: No-No: MPI_T: Consistent reference to MPI processes
- No objections
Reading: Deprecate MPI_SIZEOF in MPI-4.0
- Asked for minor whitespace change and additional text to point out that some compilers may not support
storage_size. - Pushback from Jeffs Squyres and Hammond that no available, current compiler falls into this category and we should not make those changes.
Discussion on Draft Standard and MPI-next
- Tentative targets
- Draft standard at SC19
- Release candidate meeting Dec 2019
- Ratification meeting March 2020
Plenary on status, open issues (13, 17, 18, 19, 30, 64)
- #13 Deprecate MPI_COMM_JOIN
- This is a good idea and we should do it.
- The collectives working group will pick this up.
- #17 Clarify what functions can be called on uncommitted datatypes
- It seems to have become clear recently that not many implementations are using the commit function.
- The fix here is not completely clear, but it seems that it should involve listing the functions where uncommitted datatypes are allowed.
- Wesley will move this forward.
- #18 Longer types for use with
MPI_MINLOCandMPI_MAXLOC- Pavan will raise this in the RMA WG.
- #19 Make MPI_THREAD_MULTIPLE a requirement for all implementations
- Forum agreed to close the ticket.
- #30 When are attribute callbacks invoked?
- The collective WG will take this.
- #64 Threading guarantees of MPI_User_function
- The Forum agreed that we want to clarify the existing state of the art for this issue.
- Martin will work on a proposal and bring it back to a future meeting.
No-no: Errata: Possible ambiguity on size of array_of_indices and array_of_statuses?
- Dan made a few small modifications to the reading from the previous day to make the text more consistent with the rest of the standard.
- More changes were needed and it was agreed to fix this for a future meeting.
Reading: New functions to create communicators from groups
- Dan gave an overview of tickets #93, 94, and #95
- The question was raised about deprecating
MPI_COMM_CREATE_GROUPand Dan responed that it is more likely thatMPI_COMM_CREATE_GROUPwould be changed to allowMPI_COMM_NULLand change the name of these new functions to dropFROM. - Rich raised concerns that relying on group manipulation would have serious performance implications on hardware.
- Many folks weren’t sure that this proposal made sense outside the context of Sessions because it
didn’t provide a big difference over
MPI_COMM_CREATE_GROUP. - Further notes here
Reading: New functions to create files from groups
- Decided not to read this proposal in light of the discussion for #93
Reading: Info key(s) for Persistent Collective Communication
- Tony read this proposal. There were a number of changes suggested.
- The proposal was well received but determined that it needed to be reread in September.
Reading: Non-blocking communicator/file/win constructors/destructors/etc
- Tony read this proposal.
- The decision was made to re-read the proposal based on the feedback from the room.
Reading: Nonblocking functionality for the Process Creation and Management Chapter
- Tony read this proposal.
- The decision was made to re-read the proposal based on the feedback from the room.
Reading: Nonblocking constructor/destructor(s) for the One-sided Communication RMA Chapter
- Tony read this proposal.
- The decision was made to re-read the proposal based on the feedback from the room.
Discussion: Front matter text clarifications for the “Collective Communication” and “Topology” Chapters
- Puri read the document. Originally this was scheduled to be an official reading, but changes were made during the working group session and the was converted to an unofficial reading.
- A straw poll was taken on whether to re-organize Chapter 2 to put similar operations together. It passed almost unanimously.