Re: [Starpu-devel] CPU core binding mask and worker creation.

[Mirko Myllykoski <mirkom@cs.umu.se>]

Hi,

On 2018-10-24 10:21, Samuel Thibault wrote:

> I'm wondering whether that should be set the default, or the current
> behavior should be set the default (an environment variable would allow
> to choose which behavior is wanted anyway). I'd tend to prefer 
> migrating
> to that new behavior.

The current approach can lead to odd behavior if the binding mask is set 
to something else that all available PUs. It seems that multiple workers 
end up running on a same PU and StarPU does not even warn about that. I 
think that the library should at least warn that this is happening.

My modifications do not handle the case where the binding mask is empty 
(hwloc-bind does not even allow that). However, I am not sure what 
happens if hwloc/kernel is somehow miss-configured/buggy. The library 
should probably report an error and abort.

In addition, if STARPU_NTHREADS_PER_CORE is set to a larger value than 
1, then my modification will probably cause the code to hag if the cores 
do not contain enough "activated" PUs and/or the CPU does not support 
hyperthreading.

 - Mirko