[petsc-users] Convergence problem in nonlinear problem, with PCBDDC.

Barry Smith bsmith at mcs.anl.gov
Mon Oct 13 11:31:44 CDT 2014


On Oct 13, 2014, at 11:24 AM, De Groof, Vincent Frans Maria <Vincent.De-Groof at uibk.ac.at> wrote:

> 
> But if it is indefinite, shouldn't the factorization algorithm complain about a zero pivot?

   It may not. Cholesky generally just chugs along “pushing” the zero pivot to the final entry at the lower right corner of the matrix, when it gets there if through roundoff the final value becomes slightly positive then you will never know that the problem is (technically) singular. So just cause Cholesky succeeds doesn’t mean that much.

   Barry

> 
> 
> 
> Vincent
> Von: Matthew Knepley [knepley at gmail.com]
> Gesendet: Montag, 13. Oktober 2014 17:23
> An: De Groof, Vincent Frans Maria
> Cc: petsc-users at mcs.anl.gov
> Betreff: Re: [petsc-users] Convergence problem in nonlinear problem, with PCBDDC.
> 
> On Mon, Oct 13, 2014 at 10:18 AM, De Groof, Vincent Frans Maria <Vincent.De-Groof at uibk.ac.at> wrote:
> Hi, 
> 
> 
> I am having a convergence problem in a nonlinear FE problem (we only use ksp from petsc, not snes). We don't have any problems using a direct solver with a cholesky factorization (Petsc factor, cholmod, mumps). Now we want to solve it using the bddc preconditioner. No problems on some problems. So far so good.
> 
> 
> Unfortunately, in some problems, the first increment in the newton iteration does not converge. It exits due to divergence (-4) after the first solve (0 iterations). This happens with CG, using the natural or preconditioned norm. Using the unpreconditioned norm, or minres, the problem converges with no problems. Checking the Dirichlet and Neumann errors of the subdomains (using -pc_bddc_check_level 1)  delivers small errors for all subdomains (~1e-16 - 1e-11).
> 
> 
> Because minres works, and cg doesn't, it seems like I am missing a null space. But since the subdomain solvers work without problems and do not complain
> 
> MINRES works for symmetric indefinite problems, whereas CG needs SPD problems. Could your problem be indefinite? If not, and it is close,
> its possible that the preconditioned problem is indefinite.
> 
>   Matt
>  
> about a null pivot this seems unlikely. Also when I check the number of constraints on the subdomains, they should be fully-determined...
> 
> 
> 
> any suggestions on what might be going wrong here? Am I really missing a nullspace? Or might there be something else going wrong?
> 
> 
> 
> kind regards,
> Vincent
> 
> 
> 
> 
> -- 
> 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



More information about the petsc-users mailing list