Regression: MPI_Win_start()/complete() with MPI_GROUP_EMPTY

[dalcinl]

I don't consider myself an expert on OSC, but the following code used to work across MPI implementations. Now it is failing with current master.

Reproducer

from mpi4py import MPI  # Implies MPI_Init_thread() with MPI_THREAD_MULTIPLE

buf = MPI.Alloc_mem(1024)
win = MPI.Win.Create(buf, 1, MPI.INFO_NULL, MPI.COMM_SELF)
win.Start(MPI.GROUP_EMPTY)
win.Complete()
win.Free()
MPI.Free_mem(buf)

Output:

Traceback (most recent call last):
  File "/tmp/rma.py", line 5, in <module>
    win.Complete()
  File "mpi4py/MPI/Win.pyx", line 592, in mpi4py.MPI.Win.Complete
mpi4py.MPI.Exception: MPI_ERR_RMA_SYNC: error executing rma sync

[jsquyres]

@hjelmn Does this code look correct?

[dalcinl]

Note that the following code runs to completion. If Post() with the empty group works, then I would argue that Start() with the empty group should also work.

from mpi4py import MPI
win = MPI.Win.Create(MPI.BOTTOM, 1, MPI.INFO_NULL, MPI.COMM_SELF)
win.Post(MPI.GROUP_EMPTY)
win.Wait()
win.Free()

[dalcinl]

@jsquyres I know the test looks a bit silly. Those are the silly things and corner cases I test for. I get what I deserve.

[dalcinl]

@hjelmn Does this code look correct?

An MPI expert always needs to double check with a colleague about OSC semantics 🤣.

[bosilca]

All these version look legit.

[hjelmn]

While odd that should work as I understand the standard.

[hjelmn]

When fixed we should add both a C (if it is missing) and fortran version of this to ompi-tests to cover.

[jsquyres]

An MPI expert always needs to double check with a colleague about OSC semantics 🤣.

I stay far, far away from all MPI one-sided things. 😃

[hjelmn]

Why? One-sided is fast communication without the slowness that is MPI two-sided semantics 😁 . I mean, seriously. The no-overtaking rule is explicit deleted. We kind of fixed that flaw using communicator info keys but it is still the default.

[jsquyres]

This is how I think of MPI one-sided:

[awlauria]

I addressed this in #8756 for osc/rdma.

The sync type wasn't being set, causing the SYNC error.
Also there was a couple issues freeing an empty group, the attached basic test runs clean for me now.

#include<mpi.h>
#include<stdio.h>
#include<stdlib.h>

int main() {
    MPI_Win win;
    char *buf;
    int rank;

    MPI_Init(0,0);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Group comm_group, group;

    if (rank == 0) {
        MPI_Win_create(buf,sizeof(int)*3,sizeof(int),
            MPI_INFO_NULL,MPI_COMM_WORLD,&win);
    }
    else {
        MPI_Win_create(NULL,0,sizeof(int), MPI_INFO_NULL,MPI_COMM_WORLD,&win);
        MPI_Win_start(MPI_GROUP_EMPTY,0,win);
        MPI_Win_complete(win);

    }
    /* Free window and groups */
    MPI_Win_free(&win);
    return 0;
}