[Openmp-commits] [PATCH] D107121: [OpenMP] Fix performance regression reported in bug #51235

Sat Jul 31 11:06:00 PDT 2021

tianshilei1992 added a comment.

In D107121#2918487 <https://reviews.llvm.org/D107121#2918487>, @protze.joachim wrote:

> @tianshilei1992 when flaky tests fail/block in the build bots, this is often triggered by unusual thread interleaving due to over-subscription.
> To reproduce on my own system, it helped sometimes to oversubscribe the system by running multiple instances to the flaky test at the same time:
>
>   for i in $(seq 20); do ./gtid.cpp.out & done
>
> or to repeat until failing/hanging:
>
>   while for i in $(seq 20); do ./gtid.cpp.out & done; wait; do true; done
>
> Can you reproduce the hang on your system and verify that this change fixes the issue?

Yup. In my machine, it can be almost 100% reproduced via `numactl` to set the number of threads to 6, which is same as the test machine.

  $ numactl -C 0,1,2,3,4,5 ./a.out

After I applied the patch, I didn't see it again. Can you help try with `numactl` on your side to see if it can be easily reproduced and then this patch can fix it? Thanks.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D107121/new/

https://reviews.llvm.org/D107121