Accéder au contenu.
Menu Sympa

starpu-devel - [Starpu-devel] CUDA Allocate problem

Objet : Developers list for StarPU

Archives de la liste

[Starpu-devel] CUDA Allocate problem


Chronologique Discussions 
  • From: Berenger Bramas <berenger.bramas@inria.fr>
  • To: starpu-devel@lists.gforge.inria.fr
  • Cc: Emmanuel Agullo <emmanuel.agullo@inria.fr>
  • Subject: [Starpu-devel] CUDA Allocate problem
  • Date: Tue, 26 Mar 2013 12:50:12 +0100 (CET)
  • List-archive: <http://lists.gforge.inria.fr/pipermail/starpu-devel>
  • List-id: "Developers list. For discussion of new features, code changes, etc." <starpu-devel.lists.gforge.inria.fr>

Hi,

We face a problem with StarPU and CUDA in some simulations.
The error is:
"oops in allocate_vector_buffer_on_node (datawizard/interfaces/vector_interface.c:264)... 4: unspecified launch failure"
In this case the error runs in a infinite loop (creating output file of Giga bytes! my best shot was 200 Go for an output file thanks to that).

But I also sometime get a message saying that CUDA memory is full:
"CUDA Error: out of memory"
(and even after this message the simulation continue sometimes)

We do not allocate memory by our-self during the simulation.
(But we have to mem copy to symbol at the beginning).

The size of block we are declaring with the handle are around 50MB.
I do not know if it is meaningful, but the ratio TIME TO COMPUTE  / SIZE OF DATA is small (huge data for small computational time => lots of transfer)

Any ideas?

(I can give .cu files or any other things if needed).

Thanks a lot.

Bérenger Bramas

HiePACS Project

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





Archives gérées par MHonArc 2.6.19+.

Haut de le page