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

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

Ok. I cannot really comment on the cudaEventDisableTiming flag but it 
would definitely explain why my experiment failed (unless I removed the 
flag but I do not have any memory of it).

However, I remember getting slightly worried over the fact that there 
was a potential delay between when a tasks starts/finishes and when the 
timing information is recorded. As you said, this probably does not 
matter when scheduling is concerned but I can see how it could impact 
the performance models.

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) but I have not evaluated its 
effect recently. I can see how this could cause problems if the CUDA 
threads is busy doing something else while a task starts/finishes.

 - Mirko

On 2020-06-05 13:52, Samuel Thibault wrote:
> Mirko Myllykoski, le ven. 05 juin 2020 13:49:00 +0200, a ecrit:
> > 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?
>
> Yes.  Last time we had a look, asking CUDA to record the time was
> bringing various additional synchronization.  Anyway, what StarPU needs
> to know for scheduling decisions is what it will be able to observe to
> release other tasks, it doesn't really need to care about what actually
> happens inside the GPU.
>
> Samuel