> FWIW: I have rarely seen this to be the issue.
Been bitten by similar situations before. But it may not have been OpenMPI. In any case it was a while back.
> In short, programs are erroneous that do not guarantee that all their
> outstanding requests have completed before calling finalize.
Agreed 100%. The barrier won't prevent the case of unmatched sends/receives or outstanding request handles, but if the logic is correct it does make sure that everyone completes before anyone leaves.
In any case, I also tried code #2 and it completed w/o issue on our cluster. I guess the next thing to ask Kong is regarding what version he is running and what is the platform.