[petsc-dev] configuring PETSC for a cluster with GPUs

[mary sweat]

I'm trying to use Petsc to solve a system with ksp preconditioned with jacobi.  Is this algorithm effivient in a GPU cluster?


Il Martedì 13 Maggio 2014 17:32, Karl Rupp <rupp at iue.tuwien.ac.at> ha scritto:
 
Hi,

> I'm trying to configure PETSc in an cluster with multiple GPUs, one per
> node.
> I read it is possible to use PETSc with N processes, one per GPU, each
> process accesses to its own GPU. Isn't it?

yes, this is possible and the recommended way of running it.


> Can anyone help me in that configuration?
> should I change the code to realize communication and exchange between
> processes on different GPUs?

No, this is all handled internally by PETSc. If your code handles 
'standard PETSc with MPI on the CPU', then the use of GPUs is just a 
matter of providing the right flags (-vec_type XXX -mat_type YYY).

Which algorithms do you intend to run? Many solver-related algorithms 
don't map well to massively parallel architectures such as GPUs, so you 
may better want to run a more efficient CPU-based algorithm.

Best regards,
Karli
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20140514/61e671f3/attachment.html>