Re: [Starpu-devel] Overlapping communication with computation on GPU

[Nabeel Alsaber <nabeel.alsaber1@gmail.com>]

Hi Samuel,

If I remove the stream synchronizations cudaStreamSynchronize(starpu_cuda_get_local_stream(), I get overlapping with any scheduler. But when I use the synchronizations (even with the dmda scheduler) I don't get overlapping. Is there a different way to synchronize the CUDA kernels which allows overlapping?

Thanks,

Nabeel

On Thu, Oct 30, 2014 at 10:22 AM, Samuel Thibault <samuel.thibault@ens-lyon.org> wrote:

Hello,

Nabeel Alsaber, le Wed 29 Oct 2014 15:58:10 -0400, a écrit :
> I am trying to overlap communication with computation in the provided matrix
> multiplication example.

Did you make sure to use a scheduler which takes decisions in advance,
such as dmda?  Since the default eager scheduler takes decisions late,
StarPU can't prefetch the data, and thus no overlapping can be done
since the task has to wait for the data to be transferred.

Samuel