Re: [Starpu-devel] Unexpected behavior - Not making task ready when they could be

[Berenger Bramas <berenger.bramas@inria.fr>]

Hello,
I understand, and I do not want to have concurrent access to my data!

But the behavior is not the one expected.
Please have a look to the next code (coming from the code I sent):
=====================================================
    for(int idxHandleA1 = 0 ; idxHandleA1 < nbA ; ++idxHandleA1){
        starpu_insert_task(&slowCodelete,
                (STARPU_RW|STARPU_COMMUTE), handleA[idxHandleA1],
                0);
        for(int idxHandleA2 = 0 ; idxHandleA2 < nbA ; ++idxHandleA2){
            if(idxHandleA1 != idxHandleA2){
                starpu_insert_task(&normalCodelete,
                        (STARPU_RW|STARPU_COMMUTE), handleA[idxHandleA1],
                        (STARPU_RW|STARPU_COMMUTE), handleA[idxHandleA2],
                        0);
            }
        }
    }
=====================================================
with nbA = 10, we can imagine to have a huge amount of task that can be 
executed at the same time (without using the same handle of course).
Because we have 10*10 tasks! and each is working on two handles only.
For example, one worker computes [0][1], another [2][3] and another [4][9], 
etc....
But, if you look to the trace I sent, it is sequential!

In the trace, if we look to the number of ready tasks, it looks like, when 
one task is over, only one ready task become available.
But, the dag looks fine.

Thanks.

(I repost the link to the code and the trace and the dag 
https://zimbra.inria.fr/home/berenger.bramas@inria.fr/Briefcase/Public/StarPU 
)

Bérenger Bramas

HiePACS Project

Tel (05 24 57) 40 76
INRIA BORDEAUX Sud Ouest


----- Mail original -----
| De: "Samuel Thibault" <samuel.thibault@ens-lyon.org>
| À: "Berenger Bramas" <berenger.bramas@inria.fr>
| Cc: starpu-devel@lists.gforge.inria.fr
| Envoyé: Vendredi 13 Février 2015 14:48:10
| Objet: Re: [Starpu-devel] Unexpected behavior - Not making task ready when 
they could be
| 
| Hello,
| 
| Berenger Bramas, le Fri 13 Feb 2015 11:51:43 +0100, a écrit :
| > In the test I am submitting lots of tasks that use two handles in
| > STARPU_RW|
| > STARPU_COMMUTE.
| 
| COMMUTE does not mean that the tasks can access the data concurrently.
| It only means that they can be run in either order, but still in
| exclusion: there can be only one writer.  This is because in a
| distributed environment it does not make sense to let several tasks
| write to the same data. Implementing some STARPU_CONCURRENT_WRITE flag
| should be possible, it has been on the TODO list for some time already,
| the "only" difficult part is preventing it from letting tasks go over
| different memory nodes. In the particular case of a CPU-only task, it
| would probably be relatively easy to implement.
| 
| Samuel
|