[petsc-users] Doubts - Petsc involving mumps in a GPU environment.

Barry Smith bsmith at mcs.anl.gov
Fri Mar 17 16:39:13 CDT 2017


  I did a quick google search and could find no indication that MUMPS can run on GPUs.

   However SuperLU_Dist does run on GPUs; You can install it with the additional options 

   --download-superlu_dist -download-superlu_dist-gpu --with-cuda --with-openmp --download-parmetis=yes --download-metis=yes

  Note that we have no experience with using SuperLU_Dist with GPUs and OpenMP so you'll need to direct questions about performance .... directly with Sherry Li.

   Barry

BTW: most systems have blas and lapack installed so you should not use --download-fblaslapack=yes

> On Mar 17, 2017, at 4:16 PM, Das, Rajeev <rajeev.das at cnl.ca> wrote:
> 
> UNRESTRICTED / ILLIMITÉE
> 
> Hi,
>  
> I configured Petsc using the following flags
>  
> --download-mpich=yes --with-cc=gcc --with-cxx=g++ --with-fc=gfortran --with-shared-libraries=0 --with-debugging=0 --download-mumps=yes --download-scalapack=yes --download-blacs --download-parmetis=yes --download-metis=yes --download-fblaslapack=yes --with-c2html=0 --with-cuda=1 --with-cusp=1 --with-thrust=1 COPTFLAGS=-O3 CXXOPTFLAGS=-O3 FOPTFLAGS=-O3 -march=native -mtune=native
>  
> The matrix that I am using to run a program is of type MATAIJ. The command prompts provided are  -ksp_type preonly  -pc_type lu  -pc_factor_mat_solver_package mumps. When I do a nvidia-smi , it is indicating that the program is using a GPU .
>  
> My question is with the above configuration and options, is Petsc really using the direct solver in GPU to solve or it is just making copies and that is why I am finding the GPU usage.
>  
> In addition, I would also like to know that if the direct solver is not using a GPU in the above case, then how do I use a direct solver in a GPU (using the command prompts, if possible).
>  
> Regards,
>  
> Rajeev Das.



More information about the petsc-users mailing list