[petsc-users] PCFIELDSPLIT and RHS

[Matthew Knepley]

On Thu, Aug 2, 2012 at 6:58 AM, tribur <burckhardt at itis.ethz.ch> wrote:

> Dear PETsc team,
>
> We are confrontated with the problem of a complex-valued equation system M
> * z = (A + i*B) * (x + i*y) = b + i*c which shows slow convergence using
> GMRES and, e.g., Block Jacobi or SOR as preconditioner.
>
> To be able to use algebraic multigrid preconditioning, we are trying to
> solve the equivalent real-valued system (A -B; B A) * (x; y) = (b; c) using
> PCFIELDSPLIT and, e.g.,
> -ksp_type fgmres -pc_type fieldsplit -pc_fieldsplit_type schur
> -pc_fieldsplit_schur_fact_type full -fieldsplit_real_ksp_type gmres
> -fieldsplit_real_ksp_max_it 10 -fieldsplit_real_pc_type bjacobi
> -fieldsplit_real_sub_pc_type lu -fieldsplit_imag_ksp_type fgmres
> -fieldsplit_imag_ksp_max_it 10 -fieldsplit_imag_pc_type none
>
> However, PCFIELDSPLIT does'nt work due to the lacking convergence of
> ksp(A) with the right hand side bTmp given by the current FGMRES iteration
> (outer loop). kspSolve(kspA, b, x) converges acceptably fast using, e.g.,
> GMRES and Block Jacobi, but kspSolve(kspA, bTmp, xTmp) shows stagnation
> after about 3 iterations.
>

Always check with LU for all solves first, and then back off. This behavior
sounds like you have a singular A.

   Matt


> What can we do? Do think a good initial estimate for (x,y) helps?
>
> Thanks a lot for any hints,
>
> Kathrin
>



-- 
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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120803/5092e3c6/attachment.html>