> Direct solvers are not immune to numerical stability issues. -ksp_type
> preonly does not check whether the system has been solved, it just uses
> whatever the preconditioner (direct solve in this case) returned. Putting it
> inside a Krylov method generally makes it more robust.

This is very useful to know. I now can manage to converge with
-ksp_type fgmres and -ksp_max_it 1. It does not seem to perform any
slower than -ksp_type preonly.

Thanks,
Dominik

Reply via email to