RMA - standard compliance and empty groups

[thomasgillis]

• version: 4.1.0 and 4.5.0
• installation: easybuild
• OS: ubuntu and docker on macOS
• bug reproducers: https://github.com/thomasgillis/cori-mpi-rma

My issue (which is more a question) is twofold and originates from the small reproducer code osc.cpp that fails with OpenMPI.

1. I think it relates to the issue Regression: MPI_Win_start()/complete() with MPI_GROUP_EMPTY #8677 which has been solved and merge since then (cfr in the reproducers empty.cpp). Do you have any idea of the release timeline for 4.1.2 (which I guess will contain the merge)?

2. On some machines, the reproducer osc.cpp still fails with various MPI implementations while it goes through with others (like pt2pt). To me it raises the question of the compliance of the code with the MPI Standard. Would it be possible for a osc expert to take a look and confirm that the code complies to the standard? In particular to the type deletion right after the MPI_Get and a single rank that is at the same time target and origin of RMA calls.

Thank you very much for your time and your help!

[devreal]

I can confirm that the osc.cpp hangs with the osc/ucx component. @janjust Any chance you can take a look at it?

[hppritcha]

I have a fix in the works.

[gpaulsen]

@devreal, Howard said he's working on a fix now

[hppritcha]

closed via #9556 and related PRs to release branches