Re: [Starpu-devel] Problems with memory in CUDA

[Cyril Bordage <cyril.bordage@inria.fr>]

I tried with CHUNK_ALLOC_MAX=0 but no difference.

My StarPu version is the 1.1 bracnch.

When I say 2 passes, I mean I run my program by splitting it into two
programs. But when I have the two parts together (whih a
starpu_task_wait_for_all at the end of the first part), I get a "out of
memory". For the CPU, the memory consumption is about 2Go but for the
GPU it is more than 6Go, because the memory is not freed.


Cyril.


Le 21/10/2013 11:56, Nathalie Furmento a écrit :
> On 21/10/2013 02:09, Cyril Bordage wrote:
>> Hi,
>>
>> I call starpu_data_unregister_submit to free memory. It works well on
>> CPUs, but only on CPUs, not on GPUs. To be more precise, I see, with
>> nvidia-smi, only increasing in memory consumption. And it ends with a
>> "error: out of memory". If I split my program into two passes, it works
>> like a charm (no out of memory).
>>
>>
>> Cyril.
>>
>> _______________________________________________
>> Starpu-devel mailing list
>> Starpu-devel@lists.gforge.inria.fr
>> http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/starpu-devel
> 
> Cyril,
> 
> You should tell us more. Which version of StarPU ? Which machine ? What
> do you mean by  2 passes ?
> 
> There is currently a problem with the cuda allocation memory both in the
> trunk and the branch starpu-1.1 following the commit introducing the
> batch allocation of small CUDA allocations (<= 4MiB) to avoid the huge
> cudaMalloc overhead. You could try to disable it by setting the value of
> CHUNK_ALLOC_MAX to 0 in the file  src/datawizard/malloc.c
> 
> I have no ideas if both problems are linked though.
> 
> Cheers,
> 
> Nathalie
>