Meeting Notes

February 26 - March 1, 2018

MPI 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)

MPI Forum Plenaries

Intro

Martin began the plenary portion of the F2F with introductions, WG updates, and renaming/expanding the scope of the Collectives WG to include communicators and groups.

#67: MPI_T Consistent reference to MPI processes

Marc-Andre read an issue to clarify the term “processes” in the tools chapter. There were a few places in the text where people asked for additional changes, but agreed that they were “no-no” worthy.

#77: Add integer-based access to MPI_Wtime

Marc-Andre read new interfaces for integer-based timers. There were a few discussions:

  • Wesley: The definition of MPI_WTICK needs to be updated to mirror MPI_WTICKS_PER_SECOND.
  • Wesley: A changelog entry needs to be added.
  • Jim: Is there a performance benefit to this interface over the double-based MPI_WTIME?
    • Martin: Potentially doing the int to double conversion on some architectures (32 bit ARM) might be non-negligible.
  • A straw vote was take on whether to link the source of MPI_WTIME and the new functions.
    • The result was mixed. About 13 in favor and about 5 against.

#79: Callback-driven event interface for MPI_T

Marc-Andre presented slides for a callback-driven event interface. He also showed potential text to add these ideas to the tools chapter. There was some specific feedback on the text that will continue in the tools working group.

#1 & #3 No-No Reading

Wesley read the diff to be voted on as a no-no vote before the normal vote on #1 and #3.

#25: Adding Persistent Collective Communication

Tony read the persistent collectives PR.

  • Rich: It isn’t clear that allowing interleaved starts will be possible without ending up with deadlock or slowdowns.
    • We could add an info key to promise that we will always issue operations in the same order. Such an info key is already in the proposal but commented out. The persistence WG would bring it back.
  • Martin: Can you edit the MPI_INFO object after calling the INIT function for any of the collectives?
    • Yes, but the text doesn’t say that. The text should be updated to reflect that and that changing the info object will not change anything for the existing collective.
  • Should info keys be uniform for all INIT calls?
    • They don’t have to be uniform, but they could be if the info key that is being discussed to assert that is true.
  • The room agreed that this was a “successful reading” and would be in favor of voting on it a the next meeting (with some expected “no-no” votes).

#78: Add missing communicator operations (mostly non-blocking) to Groups, Contexts, Communicators, and Caching and Topology chapters

Tony read the proposal to add non-blocking versions of many communicator, window, and file based functions for construction, destruction, and manipulation.

#28: Catastrophic Errors

Wesley read the non-catastrophic errors proposal. There were minor wordsmithing comments, but the forum agreed with unanimous consent that the reading was “successful” and should be voted on at the next meeting.

#20: User-Level Failure Mitigation

Aurelien attempted a reading of the ULFM proposal and presented supporting slides. The general feedback from the forum is that ULFM still has too many problems to standardize as a whole. Individual pieces might be acceptable (e.g. MPI_COMM_FAILURE_ACK/GET_ACKED), but others are still objectionable (e.g. MPI_COMM_SHRINK and MPI_COMM_REVOKE).

The suggestion is to break this into three pieces:

  1. New error classes and MPI_FAILURE_ACK/GET_ACKED
  2. MPI_COMM_AGREE
  3. MPI_COMM_REVOKE and MPI_COMM_SHRINK

1 and 2 are probably standardizable as is, while 3 is the part that is a problem for many. If we have 1 and 2, there might be other solutions that can replace 3 (or tweaks to 3 that can fix the problems which we would discover in the meantime).

#80: Large Count

Jeff read the large counts proposal to add large counts versions of the collective functions. He pointed out that he is not attempting to replace int with MPI_Count everywhere. If we want that everywhere, we should do it with C++ and C11 generics and he could bring it as a future proposal.

Martin: If we’re going to add _X versions of all of the MPI functions, we should make sure we have all of the changes that we we’re going to want. What about:

  • Changing the rank from an int to MPI_Rank
  • Adding an MPI_INFO object

FP16

Atsushi presented an update on the FP16 work.

#76: Same Info in Collectives

Wesley presented a plenary on behalf of Jim Dinan to add an assertion that all info keys are identical across a communicator.

There was feedback about not wanting force all info keys to be uniform, particularly around object creation. A number of potential solutions were discussed:

  1. The values of this key are a list of other keys
  2. Duplicate all keys that should have the same value and the new version is required to be uniform
  3. Add an API for setting uniform keys (doesn’t work during object creation)
  4. Add new values to keys that we want to be uniform (true_everywhere, false_everywhere)

The State of Sessions (SoS)

Dan presented an update on the state of the sessions proposal. He asked for feedback on whether the Sessions WG should start working on text for the core proposal. There was support for this in the room.

#75: Interoperability with task-based runtimes

Dan presented an issue about how MPI can interoperate with task-based runtimes. The concern is that MPI’s thread-safety rules make valid task-based programs erroneous MPI programs. This is the example program that breaks:

#OMP parallel
{
	#OMP task
	{ MPI_SSEND(...) }
	#OMP task
	{ MPI_RECV(...) }
	#OMP taskwait
}

The solution is a new thread level MPI_TASK_MULTIPLE which implies that the MPI implementation will yield in the appropriate way to unblock other tasks.