[aspect-devel] Using MUMPS through Trillions instead of PETSc
jbnaliboff at ucdavis.edu
Wed Feb 22 10:21:23 PST 2017
Great, at least on the ASPECT side this is quite straightforward. I'll
try both MUMPS and SuperLU and report back on how things go with both
the installation process and the results.
My interest in using the direct solvers is to see how much they improve
convergence behavior for 2-d problems with plasticity. Many lithospheric
deformation codes use direct solvers and for the purposes of code
comparisons it would be helpful to have a parallel direct solver
option. With pressure-dependent plasticity, I imagine many setups are
in fact quite ill-conditioned. Models trying to replicate the analogue
sandbox experiments (GeoMod) are often a horror show of spurious results
(particularly with low-order elements).
Certainly not thinking about using direct solvers for large 3-D models
with ASPECT, although a number of groups have used MUMPS and PARDISO for
Thanks again for the input!
Assistant Project Scientist, CIG
Earth & Planetary Sciences Dept., UC Davis
On 02/22/2017 06:08 AM, Timo Heister wrote:
>> I saw deal.II has the capability to use MUMPS through Trillions through the
>> SolverDirect class:
> We currently don't expose this, but if you go into solver.cc you will
> find the line
> TrilinosWrappers::SolverDirect solver(cn);
> If you change that to
> TrilinosWrappers::SolverDirect::AdditionalData data(false, "Amesos_Mumps");
> TrilinosWrappers::SolverDirect solver(cn, data);
> and you configure Trilinos with mumps (not enabled by default), you
> should be good to go.
>> However, the few discussions I’ve seen regarding using MUMPS through
>> Trillions are not entirely encouraging :)
> The first link you posted is about using MUMPS directly in deal.II
> (which we removed) and has nothing to do with using it through
> Trilinos. But yes, MUMPS is a nightmare to install/use.
>> In reality, my interest is not explicitly with MUMPS but rather a parallel
>> direct solver option (MUMPS, PARDISO, etc).
> You might try superlu (also through Trilinos). But let me ask, what is
> the reason you are looking into direct solvers? While they might help
> for very ill-conditioned problems, I doubt that you will have good
> results on large 3d problems.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Aspect-devel