Re: [Starpu-devel] Combined workers and OpenMP code

[Nicolas Collin <nicolas.collin@inria.fr>]

On 04/25/2012 05:42 PM, Samuel Thibault wrote:
> Keisuke Fukuda, le Wed 25 Apr 2012 11:36:44 +0200, a écrit :
> > starpu_combined_worker_get_size() always returns 1
> I notice that your program uses a regression-based performance model,
> but it only submits tasks of the same size.  The scheduler does not
> realize that it's still a task of the same size, which will not help the
> performance model, and thus continues to try with 1, in the hope the
> performance model will eventually provide estimations.
>
> > and only 1 thread runs in #pragma omp parallel for loop.
> > OTOH, vector_scal example run 2-4 threads (although I expect 6 workers since
> > the CPU is 6-core sandy bridge).
> Doesn't starpu_perfmodel_display show that it at least tried 6 a few
> times, perhaps only to realize that it's less efficient than 2-4?
>
> > (2) Is there any way to let (or even force) StarPU to use a certain
> > number of workers, i.e.
> > just as the same number as of physical cores.
> You can use the execute_on_a_specific_worker field to force using a
> combined worker.  Nicolas, is there any way to print the combined
> workers tree?  Also, it seems that STARPU_MIN/MAX_WORKERSIZE are not
> documented at all, could you handle that?
There is no function to print the combined worker tree yet.
You can force StarPU to create the desired combined worker by declaring 
STARPU_SINGLE_COMBINED_WORKER to 0 and declaring 
STARPU_MIN/MAX_WORKER_SIZE to the desired number.
For example :
STARPU_CALIBRATE=2 STARPU_NCUDA=0 STARPU_NOPENCL=0 
STARPU_MIN_WORKERSIZE=6 STARPU_MAX_WORKERSIZE=6 
STARPU_SINGLE_COMBINED_WORKER=0 ./simple_omp

Indeed I'm going to write a paragraph about STARPU_MIN/MAX_WORKERSIZE.

Regards
 -- Nicolas