Correct, but that’s not what I was suggesting. I was saying that the root alone
could track completion of any number of collectives individually and then
broadcast notice of those completions to all ranks.
This is but one example use case where freeing requests associated with NBCs
makes sense.
Jeff
Sent from my iPhone
> On Aug 8, 2020, at 2:14 PM, Richard Graham <richgra...@nvidia.com> wrote:
>
>
> The non-blocking reductions are required to start in the same order,
> completion order is not specified. So I don’t think you can infer from the
> completion of the last non-blocking operation (ibarrier in this case), that
> all others have completed.
>
> Also, completion is local, and even in a synchronized operation like a
> reduction, not all ranks will complete at exactly the same time. All have to
> enter the operation before any can complete, but the fact that one completed
> does not guarantee the other have.
>
> Rich
>
> From: mpi-forum <mpi-forum-boun...@lists.mpi-forum.org> On Behalf Of Jeff
> Hammond via mpi-forum
> Sent: Saturday, August 8, 2020 3:56 PM
> To: Main MPI Forum mailing list <mpi-forum@lists.mpi-forum.org>
> Cc: Jeff Hammond <jeff.scie...@gmail.com>
> Subject: Re: [Mpi-forum] MPI_Request_free restrictions
>
> External email: Use caution opening links or attachments
>
> The argument that there is no portable way to detect completion is false. It
> is completely portable to detect completion using only a subset of processes,
> and there may even be use cases for it.
>
> For example, an application can use Ibarrier to determine whether all
> processes have reached a phase of the program but only detect this at one
> rank.
>
> MPI_Request req;
> {
> foo();
> MPI_Ibarrier(comm, &req);
> if (rank) MPI_Request_free(&req);
> ..
> }
> if (!rank) {
> MPI_Wait(&req);
> printf("foo finished everywhere\n");
> }
>
> Similarly, one might have a bunch of Ireduce operations where the root is the
> only rank that tests completion, after which it notifies all processes that
> the bunch has completed.
>
> Are there any implementations that can't trivially support this? I recall
> that MPICH implements requests as nothing more than pointers to internal
> state, and I assume that Open-MPI does something similar.
>
> Jeff
>
> On Sat, Aug 8, 2020 at 9:45 AM Bangalore, Purushotham via mpi-forum
> <mpi-forum@lists.mpi-forum.org> wrote:
> I see discussion of this issue here:
>
> https://github.com/mpi-forum/mpi-forum-historic/issues/83
>
> Puri
> From: mpi-forum <mpi-forum-boun...@lists.mpi-forum.org> on behalf of Balaji,
> Pavan via mpi-forum <mpi-forum@lists.mpi-forum.org>
> Sent: Saturday, August 8, 2020 8:51 AM
> To: mpi-forum@lists.mpi-forum.org <mpi-forum@lists.mpi-forum.org>
> Cc: Balaji, Pavan <bal...@anl.gov>
> Subject: [Mpi-forum] MPI_Request_free restrictions
>
> Folks,
>
> Does someone remember why we disallowed users from calling MPI_Request_free
> on nonblocking collective requests? I remember the reasoning for not
> allowing cancel (i.e., the operation might have completed on some processes,
> but not all), but not for Request_free. AFAICT, allowing the users to free
> the request doesn’t make any difference to the MPI library. The MPI library
> would simply maintain its own refcount to the request and continue forward
> till the operation completes. One of our users would like to free NBC
> requests so they don’t have to wait for the operation to complete in some
> situations.
>
> Unfortunately, when I added the Rput/Rget operations in the RMA chapter, I
> copy-pasted that text into RMA as well without thinking too hard about it.
> My bad! Either the RMA committee missed it too, or they thought of a reason
> that I can’t think of now.
>
> Can someone clarify or remind me what the reason was?
>
> Regards,
>
> — Pavan
>
> MPI-3.1 standard, page 197, lines 26-27:
>
> “It is erroneous to call MPI_REQUEST_FREE or MPI_CANCEL for a request
> associated with a nonblocking collective operation.”
>
> _______________________________________________
> mpi-forum mailing list
> mpi-forum@lists.mpi-forum.org
> https://lists.mpi-forum.org/mailman/listinfo/mpi-forum
>
>
> --
> Jeff Hammond
> jeff.scie...@gmail.com
> http://jeffhammond.github.io/
_______________________________________________
mpi-forum mailing list
mpi-forum@lists.mpi-forum.org
https://lists.mpi-forum.org/mailman/listinfo/mpi-forum
