On Thu, 15 Sep 2011, David Andrs wrote:
> 1) Improving AztecOO solver:
> - using more of its built-in preconditioners
> - using more solver types - like CG, BISTABCG
> 2) Improving NOX
> - can do preconditioned JFNK (IFPACK must be enabled AFAIK)
> - can do jacobian
There's a bug here: you forgot to change PetscNonlinearSolver to
NoxNonlinearSolver in a couple START_LOG lines.
> 3) Adding supprt for ML package
The configure part of the patch didn't apply for me here, presumably
because I ran a bootstrap or someone else checked in a bootstrap using
a different autoconf. Applying the rest of the patch and
> Attached is also modified ex19, where I played and tested those
> patches. There are commented lines with a few preconditioners to try
> out (ILU, AMG).
No time to play much with them myself, but I don't see any other
obvious problems, if you're ready to commit the main patches.
Trilinos support still isn't in great shape (anyone know why I'm
seeing AMR convergence stall on ex14?) but you're at least uniformly
improving it. :-)
> Notes on preconditioners: I'm not an expert here, but there are only
> two preconds implemented: ILU (thru IFPACK) and AMG (thru ML).
> IFPACK supports a lot more than just ILU, however there are no
> PRECOND_ enum entries for those and I'm not quite sure if there
> should be one. What would be a good thing to do? Add new ENUM
> entries and support those?
I think so. Derek ought to have the last word there, though; he's
more on top of the Preconditioner classes than I am. The only problem
> There is a possibility there is an overlap, but the names are just
> slightly different: like (ICC in PETSc) and ICT in Trilinos. I do
> not know. If someone knows, let me know or just build on top of my
It would be nice to get that straightened out before adding new enums;
otherwise any fix to consolidate the enums later would require a
backwards-incompatible API change or a collection of "deprecated"
enums in every test.