[Openmp-commits] [openmp] [OpenMP] Use distributed fork/join barrier for large teams by default (PR #195473)

[Terry Wilmarth via Openmp-commits]

TerryLWilmarth wrote:

In that case, you would probably want to get some evidence together that the change doesn't have a detrimental effect on other applications, the SPEC benchmarks, on other architectures.  Hopefully other members of the community here can help with that. 

I'm happy to help for Intel architecture; hopefully you can find others to help cover others. I'm a bit biased because I developed the distributed barrier, and would love to see it adopted.

I've tested distributed barrier extensively (and exclusively) on Intel architectures and for Intel at least, it is mostly beneficial, but it does strongly depend on what application you are running, and how the threads map to cores across the machine.  I will dig up my numbers to remind myself what situations it doesn't handle well and report back.

https://github.com/llvm/llvm-project/pull/195473