Re: [petsc-users] [PATCH 2/2] Call KSPSetFromOptions before setting fieldsplit-specific options

[Umut Tabak]

On 09/08/2013 03:38 PM, Dave May wrote:

Remove the option
   -fieldsplit_1_pc_type icc

It is clobbering the LU option Matt suggested you try first ( -fieldsplit_1_pc_type lu)

Dear Dave,

I also realized this but and changed that with

-fieldsplit_2_pc_type icc

but thanks for looking at that, need a break I guess ;-)

When debigging such solvers, you should always run with -ksp_view to check your solver is configured correctly.

Thanks for this

On 8 September 2013 15:32, Umut Tabak <u.ta...@tudelft.nl <mailto:u.ta...@tudelft.nl>> wrote:

    On 09/08/2013 03:00 PM, Matthew Knepley wrote:

    If you want to solve hard systems, you must learn to debug a
    solver yourself. What does "just hangs" mean? Start with a full
    Schur complement factorization with LU on block 1 and a tiny
    tolerance on block 2. This will converge in 1 iterate. Then
    start slowly making the preconditioner weaker.

    Hi Matt,

    Thanks for the suggestion but I did not completely understand the
    options to use for this, I am still learning the fieldsplit
    interface. It was my mistake to not to give more in formation on
    what 'Just hangs' means, sorry.

    For the previous message: that meant extremely slow iterations,
    the iteration process was not advancing at all.

    After your reply and looking at the fieldsplit options a bit more,
    I came up with these set of options, now iterations are going fast
    but the decrease in relative residual is so slow. But apparently
    this does not converge in 1 iteration, I am still mistaken at some
    point.

    ./ex3 -pc_type fieldsplit -pc_fieldsplit_type schur
    -fieldsplit_1_pc_type lu -fieldsplit_1_ksp_type preonly
    -fieldsplit_1_pc_type icc -fieldsplit_2_ksp_type minres
    -fieldsplit_2_ksp_rtol 1e-8
    -pc_fieldsplit_schur_factorization_type full
    -ksp_monitor_true_residual

    Maybe you can help me at this point again.

    Thanks.
    Umut

       Matt

        The system is a symmetric system in this case given as

        A =
        A11 C
        C^T A22

        C is of size 1490X2432

        Could you please comment on this point for me, do I still
        have an error on the problem setup or on the usage of the
        options?

        BR,
        Umut


             
ISCreateBlock(PETSC_COMM_SELF,bs,n1,inputindices,PETSC_COPY_VALUES,&is_row1);
             
ISCreateBlock(PETSC_COMM_SELF,bs,n2,inputindices,PETSC_COPY_VALUES,&is_row2);

            We could check for this case, but it would require a
            little code and
            isn't currently done.  This is probably what you wanted:

            diff --git i/ex2.c w/ex2.c
            index 94cea2c..0cce3e6 100644
            --- i/ex2.c
            +++ w/ex2.c
            @@ -45,7 +45,7 @@ int main(int argc,char **args)
                PetscInt bs = 2, n1=1, n2=1, inputindices[]={0,1}; /*
            block size is 2 */
            
ISCreateBlock(PETSC_COMM_SELF,bs,n1,inputindices,PETSC_COPY_VALUES,&is_row1);
            -
             
ISCreateBlock(PETSC_COMM_SELF,bs,n2,inputindices,PETSC_COPY_VALUES,&is_row2);
            +
             
ISCreateBlock(PETSC_COMM_SELF,bs,n2,inputindices+n1,PETSC_COPY_VALUES,&is_row2);
            ISView(is_row1,PETSC_VIEWER_STDOUT_SELF);
            ISView(is_row2,PETSC_VIEWER_STDOUT_SELF);

-- 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