Re: [Starpu-devel] Using STARPU_COMMUTE with starpu_mpi_task_build()

[Xavier Lacoste <xl64100@gmail.com>]

Le 11 mars 2014 à 11:14, Samuel Thibault <samuel.thibault@ens-lyon.org> a 
écrit :

> Xavier Lacoste, le Tue 11 Mar 2014 11:04:23 +0100, a écrit :
>> Le 11 mars 2014 à 10:57, Samuel Thibault <samuel.thibault@ens-lyon.org> a 
>> écrit :
>>> Xavier Lacoste, le Tue 11 Mar 2014 10:51:02 +0100, a écrit :
>>>> I updated my trunk to have the STARPU_COMMUTE ignoring in MPI but now I 
>>>> have a lot of this errors :
>>>> [1,0]<stderr>:[starpu][_starpu_update_perfmodel_history] Too big 
>>>> deviation for model DGEMM on cpu0_parallel1_impl0: 1212.959000, vs 
>>>> average 989.918214 over 42 samples (+22.531234%), flushing the 
>>>> performance model. Use the STARPU_HISTORY_MAX_ERROR environement 
>>>> variable to control the threshold (currently 10%)
>>>> (Which slows down a lot computation)
>>> 
>>> Yes, this is a new thing in the trunk, the +-10% error rate may be too
>>> small, perhaps you'd rather set STARPU_HISTORY_MAX_ERROR to 25 or even
>>> more.
>> Ok, I even have 100% error on some cases.... So should I go with 
>> STARPU_HISTORY_MAX_ERROR = 100 ?
> 
> Yes.
> 
>>> Are you also deciding which GPU the GEMM should go to?  StarPU still
>>> needs to decide between the cost of moving the data and the computation
>>> cost.  Do you have a good cost function ?
>> 
>> Yes I'm deciding all about GPU mapping,
> 
> How are you deciding it?  By prefetching the data or by using
> use_a_specific_worker?  If the latter, you could perhaps as well just
> use the eager scheduler, which doesn't need a performance model.

Ok I'll do that (I use STARPU_EXECUTE_ON_WORKER, workerid when building a GPU 
task (and let StarPU decide when it's a CPU task) and can activate prefetch 
or not).
> 
> Samuel
> 
> _______________________________________________
> Starpu-devel mailing list
> Starpu-devel@lists.gforge.inria.fr
> http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/starpu-devel