[petsc-users] Help with ML/BoomerAMG

[Jed Brown]

Gaetan Kenway <gaetank at gmail.com> writes:

> Hi Again
>
> I did try running LU instead of ILU on the sub-domains. (This is the result
> of a RANS simulation at M=0.85, SA model, frozen turbulence and 96768 cells
> or matrix dimension or 483840). However, the ILU does not seem to be the
> cause of the issue. I've attached a plot of the two convergence histories.
> Both take 426 iterations of full GMRES (restart was set at 600). Column F
> is ASM(1), ILU(1) Column L is ASM(1), LU. This is right preconditioning and
> a check of the residual after the solver has completed confirms the actual
> residual has reduced by 6 orders of magnitude.

I assume that methods other than GMRES are not making much progress for
you.  Have you tried using FGMRES (or GCR) and nesting an iterative
method as the preconditioner, to reduce the size of the GMRES subspace
that needs to be stored?

It would be interesting to plot the spectrum, and more interesting to
look at the "problematic" eigenmodes.  What sort of system do you use
for visualization?  I'm working on a tool that would help with this sort
of analysis so I'm interested in enough of your workflow to make it work
for you.