[petsc-users] Generalized eigenvalue problem with spectral transformations and mpiaij compatible preconditioners

[Jose E. Roman]

El 22/05/2013, a las 22:15, Heikki Virtanen escribió:

> Hi, I have been trying to solve a generalized eigenvalue problem using
> SLEPc's EPS object. I have tried to parallelize my solver, so I should 
> use PETSc's mpiaij matrices instead of seqaij matrices. 
> Unfortunately, PETSc's LU preconditioner does not support MPI matrices. 
> (this is said in the manual and if I use it with mpiaij matrices I 
> get an error message) I get the eigenvalue problem solved with 
> LU preconditioner and seqaij matrices but, I also have to use a 
> spectral transformation ( shift-invert ) to improve convergence. 
> 
> But, as I mentioned before I cannot use LU preconditioner with mpiaij 
> matrices. Thus, I have changed preconditioner to Hypre's BoomerAMG, 
> for example. ( which supports mpiaij matrices ) When I use this 
> combination ( BoomerAMG/shift-invert transformation/
> Krylov-Schur solver) I get an early convergence failure after a couple 
> of iterations. I have also tried other solvers and preconditioners, 
> (Hypre's pilut, euclid, bjacobi and Jacobi-Davidson solver) but the result is 
> the same. Without any preconditioner I also get the early convergence
> failure and without the spectral transformation convergence is
> too slow. Any comments or suggestions?
> 
> -Heikki

Try a parallel LU such as MUMPS. This is mentioned in the manual.
Jose