Ok, thanks for explaining that. I think it should be possible for us to arrange things to avoid calling MPI_Cancel on send requests. We do this currently because we’re using MPI communication in an interactive environment where the user can tell us to stop at any time – the reproduction steps I posted were rather simplified compared to the actual case we were hitting. We already have mechanisms in place to “flush” messages at the receiving side when necessary, so that covers part of what we’ll need. Cheers, Edric. From: Zhou, Hui via discuss <[email protected]> Sent: 02 February 2022 00:16 To: [email protected] Cc: Zhou, Hui <[email protected]>; Edric Ellis <[email protected]> Subject: Re: [mpich-discuss] MPI_Cancel + MPI_Wait stalls when using ch4, not when using ch3 Hi Edric, After some pondering, we are closing https://github.com/pmodels/mpich/issues/5775<https://github.com/pmodels/mpich/issues/5775> as wontfix. For the following reasons: * Cancelling send is deprecated in MPI-4. * The potential fix will be very messy due to current design. * On the other hand, it is probably easier to work around in application. Example workaround could be to replace MPI_Issend with `MPI_Isend` plus separate MPI_Irecv for acknowledgement. You only need do this for the MPI_Issend that you may cancel, and I think it is a small inconvenience for more predictable behavior. Internally an MPI_Issend is not implemented as an MPI_Isend + MPI_Irecv to be more efficient and to exploit hardware acceleration where we can, but that also made cancelling it very complicated. If you have strong reasons to fix this issue, for example, no feasible work around, let's discuss. -- Hui Zhou ________________________________ From: Zhou, Hui <[email protected]<mailto:[email protected]>> Sent: Wednesday, January 19, 2022 12:28 PM To: [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>> Cc: Edric Ellis <[email protected]<mailto:[email protected]>> Subject: Re: MPI_Cancel + MPI_Wait stalls when using ch4, not when using ch3 Thanks for reporting. We'll fix this. Tracking issue: https://github.com/pmodels/mpich/issues/5775<https://github.com/pmodels/mpich/issues/5775> [https://opengraph.githubassets.com/abeb43664f034115c44e99f3e273e49dbfdeb9ebcc478dd9ede9ca4d439f78ab/pmodels/mpich/issues/5775]<https://github.com/pmodels/mpich/issues/5775> ch4: MPI_Cancel stalls MPI_Wait Issue #5775 pmodels/mpich<https://github.com/pmodels/mpich/issues/5775> Reported on discuss mailinglist: ---- quote ---- Hi, Running one of our test programs using MPICH 3.4.3 and ch4:ofi, I notice that MPI_Wait on an MPI_Request that has been MPI_Cancelled never compl... github.com ________________________________ From: Edric Ellis via discuss <[email protected]<mailto:[email protected]>> Sent: Wednesday, January 19, 2022 11:21 AM To: [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>> Cc: Edric Ellis <[email protected]<mailto:[email protected]>> Subject: [mpich-discuss] MPI_Cancel + MPI_Wait stalls when using ch4, not when using ch3 Hi, Running one of our test programs using MPICH 3.4.3 and ch4:ofi, I notice that MPI_Wait on an MPI_Request that has been MPI_Cancelled never completes (it does when using ch3). (The documentation for MPI_Cancel states "If a communication is marked for cancellation, then a MPI_WAIT call for that communication is guaranteed to return, irrespective of the activities of other processes (i.e., MPI_WAIT behaves as a local function)") Here's a simple example: #include <mpi.h> #include <stdio.h> #include <stdlib.h> void check(int const value) { if (value != MPI_SUCCESS) { fprintf(stderr, "Failed.\n"); exit(1); } } int main(int argc, char** argv) { MPI_Request r1; int payload = 42; int result; check(MPI_Init(0,0)); check(MPI_Issend(&payload, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &r1)); check(MPI_Test(&r1, &result, MPI_STATUS_IGNORE)); fprintf(stdout, "MPI_Test result: %d\n", result); check(MPI_Cancel(&r1)); check(MPI_Wait(&r1, MPI_STATUS_IGNORE)); MPI_Finalize(); return 0; } This stalls in MPI_Wait when executed using "mpiexec -n 1 ./a.out". Cheers, Edric. _______________________________________________ discuss mailing list [email protected]<mailto:[email protected]> To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss<https://lists.mpich.org/mailman/listinfo/discuss>