Dear Bärbel,

> we would like to use ML preconditioner from Trilinos for problems where we 
> use cubic finite elements. Is it a good idea to use this preconditioner 
> for such problems? If it is: are there any suggestions how to choose the 
> parameters appropriate?

The ML preconditioner is certainly not a bad choice - from what I've
tested, it performs better than anything else I've tried for elliptic
and elliptic-dominated problems (either scalar or vector-valued), except
geometric multigrid.

The iteration count increases if problem sizes increase, but still
modestly (14, 16, 19, 24, 28 on a Laplacian problem in 3D when halving
the mesh size). What explains this behavior is that ML coarsens too much
because too many elements couple to each other. ML has a lot of
parameters to tune, but I haven't found any real impact besides the
setting
TrilinosWrappers::PreconditionAMG::AdditionalData::aggregation_threshold,
where you can tune the parameter somewhere in the range 1e-1 to 1e-3 in
order to define which dofs should form one "coarse-grid-point". See the
step-31 tutorial program in the section "build_stokes_preconditioner"
for more detail.

Generally, ML still works for Q3 elements in 3D, but for Q4 or Q5 it
fails anyway - when there are more than a few thousand entries per row.
In 2D things are a bit better.

Best,
Martin

_______________________________________________
dealii mailing list http://poisson.dealii.org/mailman/listinfo/dealii

Reply via email to