[Trilinos-Users] Symmetric matrices

[Kevin Long]

Bart,

The Ifpack_ICT class does incomplete Cholesky, so you might look at 
that. It's pretty easy to use and seems to work well for problems like 
typical FEM stiffness and mass matrices.

However, I should say that one of my students, Kim Kennedy, recently did 
a master's project in which she used Ifpack_ICT with CG rather heavily, 
and in the course of testing her code she ran into a number of problems 
where Ifpack's ICT failed due to blowup during the incomplete 
factorization. She tested it for all SPD problems in the matrix market 
test suite; the Ifpack_ICT failure rate was nontrivial, probably at 
least 25% of the MM SPD test problems. Matlab's cholinc worked fine for 
all. (Mike H: now that her thesis is done, Kim will be putting together 
a bug report describing the failures.)

I'll reiterate, though, that she had no problem with Ifpack_ICT on some 
large FEM calculations so if that's the sort of problem you're doing, 
Ifpack_ICT may work well for you.

Kevin



On 11/07/2012 02:46 PM, Bart Janssens wrote:
> On Wed, Nov 7, 2012 at 8:53 PM, Heroux, Michael A <maherou at sandia.gov> wrote:
>> Do you want Trilinos preconditioners for symmetric operators, or just the
>> matrix-vector multiplication?
> Mike,
>
> Yes, I think some preconditioning will be needed. Currently, I use ILU
> which converges in about 15 CG iterations for the tests done so far.
> Anyway, I'll just use the complete matrices for now, and see if the
> memory actually becomes a problem.
>
> Cheers,
>


-- 
Kevin Long
Associate Professor
Department of Mathematics and Statistics
Texas Tech University

"The end of fear is the beginning of wisdom"
- Bertrand Russell