[Trilinos-Users] Exploring Trilinos GPU Capabilities with GPUs

[Scott Kruger]

I am a developer on the NIMROD MHD code from the Office
of Fusion Energy Sciences.  We are heavily implicit with
CFL numbers ~ 100 - 10,000.  With these CFL numbers, we
spend about 80% of our time in solvers.  We currently
use 3D GMRES with our preconditioning done with a
2D matrix.  For the preconditioner, we are currently using
SuperLU.  We get reasonable scaling up to 65k processors
when we use OpenMP threading, but our preconditioner is
a big bottleneck.

We are currently writing an INCITE proposal for time
on Titan, and would like to have a plan for solvers
that includes either threaded direct solves, or
GPU-accelerated direct solves.  We have other ideas
on improving our scaling using ML, but the low hanging
fruit would be any direct solve improvements (we
tried Pastix without success -- currently
reinvestigating MUMPS).  I am interested in learning
about the state of the art of what Trilinos is doing
as Google searches are turning up material that is out
of date.

Thanks,
Scott



-- 
Scott Kruger
Tech-X Corporation               kruger at txcorp.com
5621 Arapahoe Ave                Phone: (720) 974-1841
Boulder, CO 80303                Fax:   (303) 448-7756