Thanks Matthew! The problem I'm working on is the Dirac equation, in various number of dimensions. I'm going to take a stab at multigrid preconditioning.
Best regards Torquil Sørensen On 27 October 2013 14:48, Matthew Knepley <[email protected]> wrote: > On Sat, Oct 26, 2013 at 3:05 PM, Torquil Macdonald Sørensen < > [email protected]> wrote: > >> Hi! >> >> I have a linear problem Ax=b, with complex values, that is solved very >> well using the ILUTP and GMRES implementations in GMM++. None of the other >> preconditioners available in GMM++ would work (e.g. ILU and ILUT) for this >> problem. >> >> Then I tried the same problem using ILU and GMRES in PETSc, with no >> success, despite a lot of attempts at adjusting ILU settings. I always end >> up with gigantic residual norm values. The same PETSc program works well >> when I apply it to a different matrix A. >> >> I'm now suspecting that the ILU options cannot be set so as to obtain >> ILUTP behaviour. >> >> What would be the recommended method to access an ILUTP preconditioner >> from PETSc? >> >> According to the PETSc website, a preconditioner named ILUDT is available >> by using the external package Hypre, but I had to deselect Hypre during the >> PETSc build due to my use of complex numbers... So do you guys think that I >> should transform everything to a real representation and try Hypre/ILUDT? >> > > I suggest you look for a different preconditioner. The fact that you have > so much trouble reproducing the behavior shows > you just how fragile the performance of ILU is. It may work for a certain > size, but fail for larger or smaller problems, or slightly > different parameters. What problem are you solving? Usually the best > option is to consult the literature for preconditioners > tailored to your problem, and then reproduce them. > > Matt > > >> Best regards >> Torquil Sørensen >> >> > > > -- > What most experimenters take for granted before they begin their > experiments is infinitely more interesting than any results to which their > experiments lead. > -- Norbert Wiener >
