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

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

Ok, thanks!

I remember reading through the StarPU code some time ago (at least one 
year ago, perhaps two) while attempting to solve performance model 
calibration related problems with K80 GPUs. My understating is that 
StarPU places several tasks into a CUDA stream and checks the statuses 
of associated CUDA events to figure out if tasks have finished. My 
memory is that StarPU does not actually extract the execution time from 
the events. Instead, StarPU records the wall-time when the CUDA event is 
checked (in _starpu_driver_end_job?). Is this correct?

I think I attempted to modify the StarPU code such that it would extract 
the execution time from the CUDA events but it did not fix the problem. 
Again, this was some time ago and I had just a limited amount of time 
available.

 - Mirko

On 2020-06-05 13:29, Samuel Thibault wrote:
> Mirko Myllykoski, le ven. 05 juin 2020 13:20:10 +0200, a ecrit:
> > Am I understanding correctly that these mean == 0.0 entries are the result
> > of a cleanup performed before the models gets calibrate?
>
> Yes, so my fix should be fixing the issue you were having.
>
> Samuel