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

[Xavier Lacoste <xl64100@gmail.com>]

Hello, 

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)

As I'm statically deciding which GEMM go on GPUs before computation, should I 
use STARPU_COMMON + cost_function instead of STARPU_HISTORY_BASE + size_base 
? 

XL.

Le 10 mars 2014 à 12:44, Samuel Thibault <samuel.thibault@ens-lyon.org> a 
écrit :

> Xavier Lacoste, le Mon 10 Mar 2014 09:40:23 +0100, a écrit :
>>        So, in my algorithm, I have several GEMM tasks unlocking a 
>> GETRF+TRSM
>>        unlocking an other list of GEMMs etc.
>> 
>>    Ok, and are all the GEMMs happening on the same MPI node?
>> 
>> Yes all GEMMs applying on one column block are performed on the MPI node 
>> that
>> owns the column block (it's the only "write access" data of the GEMM task).
> 
> Ok, so there's nothing to optimize at the MPI layer here.  For now I
> have made StarPU-MPI simply ignore the flag, it should work just fine
> (it works for Cholesky at least, I believe your use case is just the
> same).
> 
> Samuel