Wenduo Wang
Wenduo Wang
Reading through everyone's opinions above, my impression is that none of us could predict how real workloads will use GPU, leading to the fear that a short-term fix will prove...
> PMIx already calculates the distances from porc->gpu and proc->nic. So we can retrieve this information and If GPU is set && GPU selection enabled find the closest proc->nic distance...
> I cannot address what one might find with a simple Google search - the terms of the search, the specific search engine, etc all tend to make the results...
@rhc54 Could you provide some insights?
Removing v5.0.x label - this will be a main-only change.
Also passed AWS CI
@sdonoso We have ingested multiple runtime fixes since 5.0.1. Can you reproduce on 5.0.3? For example, we fixed this a while ago https://github.com/open-mpi/ompi/issues/12064
Thanks. I updated the issue title.
Should be fixed in 5.0.4 scheduled in 7/2024
Just realized that ~~this~~ the original issue happened on AWS. I will take a look.