[Trilinos-Users] [EXTERNAL] Re: Tpetra::VbrMatrix interface

Hoemmen, Mark mhoemme at sandia.gov
Tue Aug 5 11:34:21 MDT 2014



On 8/4/14, 10:31 PM, "C.S. Natarajan" <csnataraj at gmail.com<mailto:csnataraj at gmail.com>> wrote:

Alan, Mark,
            Thank you very much for your responses. I do have constant block sizes, however, for now I think I will just use the Crs interface. Just curious, if you are doing in-house development (FEM), what is your preferred methodology for problems with multiple DOF per node (Problems span the range from electrostatics to fluid flow)?

The VBR approach in both Epetra and Tpetra (once we get the bugs worked out) emphasizes the "V" for "variable": that is, the blocks can vary arbitrarily.  It tends to perform better for larger blocks.  "Larger" means "large enough that calling the BLAS is worthwhile."  VBR will NOT be fast for 2x2 or 3x3 blocks.  Neither Epetra nor Tpetra currently has a complete solution that is optimized for small blocks.  Our users would like this, but demand other things more loudly, so we haven't focused energy on optimizing sparse matrices for multiple DOFs per node.  Trilinos' Teko package has physics-based / block preconditioners that deal with different kinds of DOFs, and I think MueLu may have some support for multiple DOFs per node as well.

I am going with the Tpetra, Intrepid (+shards), Muelu, Belos stack; is this a reasonable choice?

As mentioned, we don't have good support for block data structures at the moment.  We are working on improving Tpetra's support for small blocks, and will soon add a block relaxation (Gauss-Seidel / SOR) preconditioner that can use this data structure.  However, this is not done yet.  It will take a while to propagate support for block matrix data structures up through the software stack.  The data structures will work in solvers, but things like preconditioner construction might not be optimized for a while.

I can answer more specific questions about the Tpetra, MueLu, Belos solver stack, but will defer to others for Intrepid + Shards questions.

Thanks for your interest!  Please feel free to ask more questions.
mfh
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://software.sandia.gov/pipermail/trilinos-users/attachments/20140805/c0c1c451/attachment.html>


More information about the Trilinos-Users mailing list