1.1.0 release is not working for me

Hi there, I had major issues on our compute cluster with the newest release of threadpoolctl. I don’t quite know what is happening because I don’t get any error messages and I have not been able to reproduce the error with a minimalist example.

What is happening is basically - nothing. I am spawning a bunch of background workers which are supposed to generate data and put it into some queue. The code that runs in the background workers is embedded in with threadpool_limits(limits=1, user_api="blas"): to limit the number of threads within each worker. This works perfectly on 1.0.0, but in 1.1.0 no data will arrive in the queue, no error message is printed, cpu is idle (so the really is nothing happening) and I get large core.304403 files in the folder the script was running in (as if there was a segmentation fault).

This issue only occurs on our compute cluster (LSF schedule with modular software stack (maybe this is the problem?), centOS). On my local workstation (Ubuntu 16.04) everything is fine.

I would greatly appreciate if you have any idea why this happens or how this could be resolved.

Best, Fabian

[jeremiedbb]

It may be because before your were only invoking OpenMP in child processes whereas now you also invoke OpenMP in the parent process by using threadpoolctl. This is known to cause deadlocks, see https://stackoverflow.com/questions/49049388/understanding-openmp-shortcomings-regarding-fork

[FabianIsensee]

Apologies for the delay. I have allocated time on Friday to look into this in more depth. I will let you know how that goes. Best, Fabian