> I tried this a long time ago and it turns out that it doesn't make much of a > difference to start CG with the interpolated previous solution.
I am curious. I have and still do often invert the preconditioner to get a good enough guess to start my iterations. Typically, it yields a faster convergence especially inside a nonlinear scheme (newton/picard). I've never really needed this for just linear systems so far but would've thought that with a good guess, the iteration count should drop. Perhaps with strongly varying solutions, the projection on a refined mesh might not have been sufficient to capture all the scales but it is much better than a zero guess I would think ! I'm implementing the SolutionTransfer for my problem and can verify if I see any significant gains. Vijay On Thu, May 10, 2012 at 2:14 PM, Wolfgang Bangerth <[email protected]> wrote: > On 05/10/2012 02:54 PM, Vijay S. Mahadevan wrote: >> >> With PETSc, you will have to specify that you are using a non-zero >> initial guess using >> >> >> http://www.mcs.anl.gov/petsc/petsc-dev/docs/manualpages/KSP/KSPSetInitialGuessNonzero.html >> KSPSetInitialGuessNonzero(ksp, PETSC_TRUE) >> >> If this is not set, it will reset the vector to zero and sure, then >> there is nothing to be gained. I haven't looked if a similar behavior >> exists for the Deal.II internal solvers. > > > The deal.II solvers simply start with whatever values are in the solution > vector whenever you call > cg.solve (system_matrix, solution, system_rhs, preconditioner); > > Best > > W. > > ------------------------------------------------------------------------ > Wolfgang Bangerth email: [email protected] > www: http://www.math.tamu.edu/~bangerth/ > _______________________________________________ dealii mailing list http://poisson.dealii.org/mailman/listinfo/dealii
