[CP2K:4215] Not enough GPU device memory

Urban Borštnik urban.b... at gmail.com
Mon Dec 3 08:54:22 UTC 2012


Hi Miro,

Have a look at cuda_tools/README if you haven't found it yet.  Regarding
the memory keyword: yes, you should limit it.  That's only for some
parts of the CUDA code, and if it takes all the memory, then others can
not.

Just a few parts of the code are CUDA-enabled for now and these may not
be the ones you need for your calculation.

One of the implementation choices so far has been that one MPI rank can
use no more than 1 GPU.  MPI ranks can share a GPU.  Besides trying
that, you may want to try compiling a hybrid (MPI+OpenMP) code and using
1 MPI rank per GPU.  In any case it's recommended to compile with the
machine architecture support (see machine/README--you'll need libnuma or
preferable hwloc) to better bind processes (and maybe threads) closer to
the GPUs).

Best regards,
Urban.

On Fri, 2012-11-30 at 14:30 -0800, mirix wrote:
> Hello,
> 
> I am new to CP2K and I am not a developer, just a chemist. I have just
> compiled CP2K in a Debian Testing 64-bit workstation with 64GB of RAM.
> I have created a parallel binary, which, in theory is also
> CUDA-enabled. In the workstation right now there is only an NVIDIA
> Quadro FX4800 with 1.5GB of memory, but I have 3 Teslas in a drawer
> waiting for CP2K to become useful.
> 
> However, I have just run the test C.inp, which apparently ran fine,
> but there are many output lines containing the following:
> 
> NOT ENOUGH GPU DEVICE MEMORY!
> 
> Searching through the manual, I have only found a CUDA/GPU-related
> keyword, which is memory. The default is -1. What does it mean? That
> the GPU should not be used or that the memory is unlimited?
> 
> In fact, modifying the input like this seems to solve the problem:
> 
> &GLOBAL
>   PROJECT C
>   PRINT_LEVEL LOW
>   &CUDA
>       MEMORY 533504
>   &END CUDA
> &END GLOBAL
> 
> I would like to know if the CUDA options are documented somewhere and,
> in particular, if I can choose which GPU to use and if I can use more
> than one GPU per calculation.
> 
> Kind regards,
> 
> Miro
> 
> 
> -- 
> You received this message because you are subscribed to the Google
> Groups "cp2k" group.
> To view this discussion on the web visit
> https://groups.google.com/d/msg/cp2k/-/-d0tO7pkzVoJ.
> To post to this group, send email to cp... at googlegroups.com.
> To unsubscribe from this group, send email to cp2k
> +unsub... at googlegroups.com.
> For more options, visit this group at
> http://groups.google.com/group/cp2k?hl=en.





More information about the CP2K-user mailing list