Re: [Starpu-devel] Problem with performance model calibration

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

I will check if that --disable-cuda-memcpy-peer flag is still relevant 
with the last versions of StarPU. It will probably take awhile since I 
have only limited access to GPUs right now.

 - Mirko

On 2020-06-05 14:19, Samuel Thibault wrote:
> Mirko Myllykoski, le ven. 05 juin 2020 14:13:35 +0200, a ecrit:
> > As you said, this probably does not matter when scheduling is
> > concerned but I can see how it could impact the performance models.
>
> Yes, sure, but we use the performance models *for* the scheduling :)
>
> > I should also mention that I am compiling StarPU with the
> > --disable-cuda-memcpy-peer flag enabled. This seems to improve the
> > performance (sometimes significantly)
>
> That's surprising. It'd be useful to investigate in traces how you can
> get a performance improvement since this option basically adds *more*
> synchronization within CUDA.
>
> > I can see how this could cause problems if the CUDA threads is busy
> > doing something else while a task starts/finishes.
>
> Yes, that would perturbate performance models a lot since it'd make CUDA
> report kernel termination only after some other terminations as well.
> Again, the performance model getting affected is actually what we want
> since that's what the scheduler will be faced to.
>
> Samuel