dolfin team mailing list archive

[Jed Brown <jed@xxxxxxxx>]

On Thu 2008-09-25 09:59, Nuno David Lopes wrote:
> I'm using uzawa methods with hood-taylor (P2-P1) so we don't really deal with 
> the complete indefinite system 
>            (A B')(u)=(f)
>            (B  0)(p)  (0). 
> 
> The idea is to work with something like 
> Au^n=f+B'p^n            (this is the problematic case in big problems) 
> p^(n+1)=p^n-\rho B u 

This is actually equivalent to a Richardson iteration on the Schur
complement.

> In the simpliest case, A a comes from -(grad(u),grad(v)), so it should be 
> a "good" system to deal with. 
> And it works "well" with other preconditioners. 

And using -ksp_type gmres -pc_type hypre -pc_hypre_type boomeramg is
failing for solving with A?  This doesn't make sense to me, I regularly
use it for this problem.  Is the memory blowup occurring during PC
assembly?  Are you sure the matrix is correct?  Is it parallel or
serial?  How much storage does the matrix alone take?  Does it work
correctly for smaller problem sizes?  Have you tried with ML (-pc_type
ml -mat_no_inode)?  Multigrid should work very well for this problem.

Jed

Attachment: pgpRB0UWFWt6L.pgp
Description: PGP signature