[Dune] SuperLU with CMake

Dominic Kempf dominic.r.kempf at gmail.com
Fri Feb 10 17:00:20 CET 2017


Hey,

I think there is some misconception here:

SuperLU and UMFPack are *direct solvers*. As such, they are quite limited
in the problem sizes they can solve.
As linear solvers they are usually not feasible beyond small toy problems.

However, as a *coarse grid solver*, they can signifcantly improve the
convergence of your AMG solver.
So, if cmake has not detected either of the two you should really fix this
to have your AMG give best performance.
Once, the direct solver is detected, AMG automatically uses it for the
coarse grid level.

Best,
Dominic

On Thu, Feb 9, 2017 at 4:12 PM, Christoph Grüninger <
christoph.grueninger at iws.uni-stuttgart.de> wrote:

> Hi Gergő,
> on my computer, I can solve systems in 2d up to 150'000 unknowns, in 3d
> (more nonzero entries) its less then 100'000. Unless your matrix is
> special, i.e., symmetric, AMG is probably the best you can get.
>
> That's what you can find in my opts file (without installing, those line
> gets added to CMAKE_PREFIX_PATH):
>
> >   -DSUPERLU_INCLUDE_DIR=/temp/gruenich/dune/zusatz/SuperLU_5.2.1/ \
> >   -DSUPERLU_LIBRARY=/temp/gruenich/dune/zusatz/SuperLU_5.2.1/build/SRC/libsuperlu.a
> \
>
> Have you tried the packages from your distribution or homebrew? Usually
> that's the easiest way.
>
> Bye
> Christoph
>
> --
> When I worked at Harvard's Writing Center, we joked that the single
> most useful piece of equipment for a writer was a bucket of glue.
> First you spread some on your chair, and then you sit down.
> [Joan Bolker, Writing Your Dissertation in 15 Minutes a Day]
>
>
> _______________________________________________
> Dune mailing list
> Dune at dune-project.org
> http://lists.dune-project.org/mailman/listinfo/dune
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.dune-project.org/pipermail/dune/attachments/20170210/442e8f8d/attachment.htm>


More information about the Dune mailing list