[petsc-users] parallel complex preconditioner needed

[Jed Brown]

On Tue, 20 Jul 2010 11:06:53 -0400, Craig Tanis <craig at tanistech.com> wrote:
> This is a frequency domain EM simulation.

This is an especially difficult problem if you have large jumps in
coefficients.

> I think my main problem is interpreting the options available..  PETSc
> apparently has a built-in direct solver, which is how ILU works on a
> single processor, right?

The I in ILU stands for Incomplete.  The direct solver is LU which you
can access with -pc_type lu.

> Are the parallel direct solvers in some 3rd party package,like hypre,
> that I need to install separately?  Do I somehow specify a direct
> solver to use for the built-in ILU ?

Yes, the native LU is not parallel, you can get one with
--download-{blacs,scalapack,mumps} or --download-superlu_dist, and

  -pc_type lu -pc_factor_mat_solver_package mumps (or superlu_dist).

To try a 1-level domain decomposition method, run with -pc_type asm,
check the -help options under -pc_asm_*, and control the subdomain
solver with -sub_{ksp,pc}_*

Jed