Re: [Starpu-devel] StarPU code hangs when using a lot of GPU memory

[Mirko Myllykoski <mirkom@cs.umu.se>]

Hi,

On 2019-03-02 16:39, Samuel Thibault wrote:
> [...] I guess you are not using a lot of streams
> on the same CUDA device with STARPU_NWORKER_PER_CUDA, and not using a
> very large CUDA pipeline with STARPU_CUDA_PIPELINE? I also guess you 
> are
> not using starpu_data_acquire on the CUDA GPU?

I have not touched the STARPU_NWORKER_PER_CUDA nor STARPU_CUDA_PIPELINE 
enviromental variables during these test runs. The code does not contain 
any starpu_data_acquire() function calls either. I tried setting 
STARPU_CUDA_PIPELINE to 0 at one point but it did not have any effect.

> Could you re-run with the latest gdbinit version where I added more
> details?

Here:

(gdb) starpu-memusage


Node 0:
  Total used: 28, 190MiB
  WT: 0, 0MiB
  home: 0, 0MiB
  OOC: 28, 190MiB
  diduse: 28, 190MiB
  redux: 0, 0MiB
  relax: 0, 0MiB
    noref: 0, 0MiB
  normal: 28, 190MiB
    owner: 26, 175MiB
    shared: 2, 14MiB
    invalid: 0, 0MiB
    nosubdataref: 14, 94MiB
    nodataref: 14, 94MiB
    reading: 14, 95MiB
    writing: 0, 0MiB
    overwriting: 0, 0MiB

  cached: 0, 0MiB


Node 1:
  Total used: 1209, 958MiB
  WT: 0, 0MiB
  home: 0, 0MiB
  OOC: 1209, 958MiB
  diduse: 1203, 941MiB
  redux: 0, 0MiB
  relax: 2, 29MiB
    noref: 0, 0MiB
  normal: 1207, 929MiB
    owner: 1205, 914MiB
    shared: 2, 14MiB
    invalid: 0, 0MiB
    nosubdataref: 0, 0MiB
    nodataref: 0, 0MiB
    reading: 0, 0MiB
    writing: 0, 0MiB
    overwriting: 0, 0MiB

  cached: 0, 0MiB

 - Mirko