Re: [Starpu-devel] StarPU CUDA out of memory

[Usman Dastgeer <usman.dastgeer@liu.se>]

Thanks for the fix.

Just found a minor issue in workers.c:

                for (impl = 0; impl < STARPU_MAXIMPLEMENTATIONS; impl++)
                {
                        ...
                        case STARPU_CPU_WORKER:
                                if (task->cl->cpu_funcs[ i ] != NULL)
                                        ...

                        case STARPU_CUDA_WORKER:
                                if (task->cl->cuda_funcs[ i ] != NULL)
                                    ...

                        case STARPU_OPENCL_WORKER:
                                if (task->cl->opencl_funcs[ i ] != NULL)
                                    ...
                        ...
                }
Should be  "impl" instead of "i" as follow (?)

                for (impl = 0; impl < STARPU_MAXIMPLEMENTATIONS; impl++)
                {
                        ...
                        case STARPU_CPU_WORKER:
                                if (task->cl->cpu_funcs[ impl ] != NULL)
                                        ...

                        case STARPU_CUDA_WORKER:
                                if (task->cl->cuda_funcs[ impl ] != NULL)
                                    ...

                        case STARPU_OPENCL_WORKER:
                                if (task->cl->opencl_funcs[ impl ] != NULL)
                                    ...
                        ...
                }

regards,
Usman.

On 02/22/2013 12:15 PM, Samuel Thibault wrote:

Hello,

Samuel Thibault, le Mon 18 Feb 2013 16:36:02 +0100, a écrit :

Usman Dastgeer, le Mon 18 Feb 2013 10:22:07 +0100, a écrit :

I am using both vector and matrix interface. Here is a simple program to
reproduce the error.

It's a deep problem in CUDA. We didn't know that even just starting a
kernel could have to allocate memory and thus fail... So we're currently
having a look at limiting the amount of memory that StarPU uses.

This is now available. For now we don't limit automatically (since AFAIK
CUDA doesn't provide a way to know how much memory is left), you can
however set it by hand:

export STARPU_LIMIT_CUDA_MEM=1900

to limit to 1.9GiB over 2GiB for instance, thus leaving 0.1GiB for
CUDA-internal allocations.

Samuel