Hm, I attempted to attach my full code, but maybe it didn’t get through correctly. Weirdly though when I look at the web view of this discussion group, *only* the code shows up (initially):

https://lists.mpich.org/pipermail/discuss/2024-April/006578.html

Hopefully that is visible? (I think the problem shows up only when using 2 hosts)

(On looking again at that code, there’s a stray logging message about MPI_Comm_free – initially I was using separate communicators for the ISend and the Barrier – but that seems to make no difference).

Cheers,

Edric.

Hi Edric,

I don’t see anything wrong in your pseudo code. I believe it is a correct pattern. I ran some experiments myself on some local machines and could not cause a hang, so if you come up with a reproducer please send it along.

Ken

I'm trying to understand if a change in behaviour I'm seeing is expected or not. My code initiates an MPI_Isend on rank==1, and before waiting for completion of that send, all ranks perform an MPI_Barrier. This works fine on ch3:nemesis. It works fine on ch4:ofi when the message is small (presumably using the "eager" protocol). When using ch4:ofi (the embedded tcp provider) and the message is large (presumably switching to "rendezvous" protocol), rank 0 never leaves the MPI_Barrier call. (I think the SHM piece of ch4 does not show the problem)

Should this work? I cannot find anything in the MPI standard that says it should not, but perhaps I'm not looking in the right place. 

I'm using mpich-4.1.2 in both cases, either in "--with-device=ch3:nemesis" mode or "--with-libfabric=embedded --with-device=ch4:ofi:tcp". 

Here's a sketch of the problematic section of code, I'll attempt to attach a full reproduction (but I'm not sure if that works?)

// setup code...

if (rank == 1) {

    MPI_Isend(data, count, MPI_INT, 0, TAG, comm, &req);

} else {

    std::this_thread::sleep_for(std::chrono::seconds(1));

}

MPI_Barrier(comm);

MPI_Barrier(comm);

// cleanup code, receive the message etc...

Cheers,

Edric.