On Wed, Dec 1, 2010 at 15:44, Laurent Michel <laurent.michel at epfl.ch> wrote:
> You can find all the details about the problem I am solving in attachment. > Do you know Guillaume Jouvet (recent graduate at EPFL, worked on a similar problem)? It is unlikely that BoomerAMG will work for this problem, ILU will be very sensitive to ordering. Since you use equal-order velocity and pressure spaces, your best bet for purely algebraic solvers is to interlace the fields (u0,v0,w0,p0,u1,...) and use a solver that respects these blocks (use a BAIJ format or call MatSetBlockSize(A,4)). I know people who claim success using ML for these equal-order methods, as long as the block structure is respected, but this method is fragile and fails completely in some non-extreme cases. Note that most glaciological applications require a pretty strong conservation statement, and the P1-P1 GLS methods that you describe using are notoriously bad at conservation, especially with the irregular geometry and steep gradients in coefficients that occur in practice. Basically, what I want to solve is a time-independent Stokes problem, non > linear in the viscosity (for the moment, I solve a fixed point for this). I > am using the finite element library of my research group. The mesh is > generated with gmsh (btw would you have any advice for a free, parallel, > easy-to-use mesher?) > No, I use Cubit, but it takes a long time to get a license. Jed -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20101201/656bd24e/attachment.html>
