>    Then I modified this code (following tutorial 17) to run it in
>    parallel using PETSC (just as a sand-box be sure that I understood
>    how the parallel tools work). Principally nothing have changed
>    except CG solver (for "inversion" of the mass matrix) now is used
> from PETSC and it uses another preconditioner.

So does the solver work if you run the program on a single machine, or does 
it work there as well?

PETSc's CG solver by default uses a BlockJacobi preconditioner which 
computes an ILU of the locally stored block on each processor. In other 
words, on a single machine, the BlockJacobi preconditioner is equivalent 
to an ILU of the entire matrix, but if you have more than one machine 
communicating via MPI, the BlockJacobi preconditioner moves away from ILU 
and towards a simple Jacobi preconditioner, i.e. the preconditioner 
becomes worse and worse as there are more processors. As a consequence, 
you have to expect more linear iterations, and you have to make sure you 
allow more iterations to achieve the same accuracy.

I would not be surprised if a lack of accuracy in the CG solver can lead to 
instabilities even if the time step was otherwise small enough.

Best
 W.

-- 
-------------------------------------------------------------------------
Wolfgang Bangerth                email:            [email protected]
                                 www: http://www.math.tamu.edu/~bangerth/

_______________________________________________
dealii mailing list http://poisson.dealii.org/mailman/listinfo/dealii

Reply via email to