[Trilinos-Users] Help choosing a matrix format (CRS/VBR)

[Heroux, Michael A]

Hi Pat,

I think there are a couple of reasons people choose CrsMatrix over
VbrMatrix.  

First is that it is simpler to work with CrsMatrix objects.  VbrMatrix adds
another layer of data structures that many people don't want to bother with.

Second, the Vbr matrix classes in Epetra are do not perform as well as I
would like them to.  We chose adding a serial dense matrix abstraction to
manage the block entries, which improves usability but makes it hard to get
performance out of small block sizes.  This aspect should improve with
Tpetra.

As for advice, I think going with the VbrMatrix path is best for new
development, especially on modern and emerging architectures.  VbrMatrix
gives us the opportunity to exploit a source of data regularity that is
going to be increasingly important.

As for examples, I know we have them, but other might be able to point you
to real code that uses VbrMatrix, from which you can get a good starting
point.

Mike


On 3/20/11 11:19 PM, "Pat Notz" <patnotz at gmail.com> wrote:

> Hi,
> 
> I'm very new to developing against Trilinos. I started out by watching
> Mike's day-1 intro talks from TUG2010 and reading the Epetra tutorial
> from the didasko project. In Mike's presentation, he mostly discusses
> Epetra_CrsMatrix but mentions the Epetra_VbrMatrix and the FE versions
> of those matrices. In particular, the FEVbrMatrix sounds appropriate
> because I have multiple unknowns per mesh point and I'm doing a
> distributed FEM assembly.
> 
> I'm looking for example code that uses either VbrMatrix or FEVbrMatrix
> but it seems most people prefer CrsMatrix. Even the FEI package
> appears to use CrsMatrix and so does the Albany project, which is a
> multivariate FEM code.
> 
> I'm curious what motivates the choice of CrsMatrix for general FEM
> codes (if anyone here knows).
> 
> Thanks ~ Pat
> 
> _______________________________________________
> Trilinos-Users mailing list
> Trilinos-Users at software.sandia.gov
> http://software.sandia.gov/mailman/listinfo/trilinos-users
>