[Trilinos-Users] Sparse direct solver and language options for GPU

Tom Anderson tomacorp at gmail.com
Wed Sep 17 10:49:04 MDT 2014


Is C++ and Amesos2 the only GPU-capable sparse direct solver option using
Trilinos?
Have I missed any good alternatives?

Here is a longer version of the same question with a description of the
application:

I am working on a thermal simulation problem, using a fine square mesh
instead of a more complex course mesh.

I think about the thermal problem as the solution to an equivalent
electronic circuit.

I have constructed a matrix using Modified Nodal Analysis as described in
http://www.swarthmore.edu/NatSci/echeeve1/Ref/mna/MNA2.html .
My PyTrilinos matrix code agrees with the Xyce circuit simulator.
I could just use Xyce instead of writing my own matrix code,
but Xyce gets slow on large problems.
My equivalent circuit only needs voltage sources, current sources, and
resistors.
The solution I want is the DC operating point of the circuit.

The matrix is real, Hermitian, positive semi-definite, if I understand
matrix lingo correctly.
The condition number is about 430 and I expect it to always be less than
about 7000.

I am aiming for 200e6 unknowns (circuit nodes) and 1.4e9 matrix entries
(1.2e9 resistors).
As of now, I can solve 1e6 nodes using Amesos/PyTrilinos on an old Mac
mini, single threaded.
Solve time is reasonable and linear in the number of unknowns, and memory
usage is about 1kB per unknown.
My existing code uses PyTrilinos and is 2D.  I plan to move to 2.5D (3D
planar) soon.
I expect to deploy the application on Linux.

Next I would like to use GPU processing. I didn't find a way to use the GPU
with PyTrilinos.
I have been avoiding Fortran because of the install process on OSX.
It seems that C++ and Amesos2 is the only GPU-capable sparse direct solver
using Trilinos.
Have I missed any good alternatives?

Any advice about my project is welcome, including suggestions for different
approaches, platforms, or software.

Thanks,

Tom Anderson
tom_anderson at keysight.com
Keysight Technologies
Santa Rosa, California
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://software.sandia.gov/pipermail/trilinos-users/attachments/20140917/3181f008/attachment.html>


More information about the Trilinos-Users mailing list