[petsc-users] Better PCGAMG Settings in 2D?

[Mark Adams]

On Thu, Jul 19, 2018 at 9:44 PM Weston, Brian Thomas <weston8 at llnl.gov>
wrote:

> Hi Mark,
>
> I’m using the default settings with GAMG for 2D problems and it’s
> performing much worse than BoomerAMG for my problem. I was told, however,
> that GAMG’s defaults were optimized for 3D problems, so I was wondering if
> you knew of better settings for 2D or which knobs might be most significant
> to play with? Is pc mg levels the way to control the number of aggressive
> coarsening levels? I tried using unsmoothed aggregation, but it was only
> marginally better than the default.
>

Are you looking at iteration count or solve time?

This is a 5-point stencil. You want to square the graph:
'-pc_gamg_square_graph 20' (20 is the number of levels to square the graph,
which is infinity really).

You can use:

-mg_levels_ksp_type richardson
-mg_levels_pc_type sor

and one processor this is a prefect smoother, just for debugging.

And you can run with -info and grep on GAMG and send me the result (about
20 lines). This should give me an idea of what is wrong if these ideas
don't work.

Mark


> The discretization is a essentially a 2nd-order cell-centered
> finite-volume scheme and I’m primarily solving a pressure-velocity Schur
> complement system.
>
> best,
> Brian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20180720/7c06df93/attachment.html>