[petsc-users] Unreproducible number of iterations for consecutives solves (GCR + ML)

[Dave May]

Hi Thomas,
Does this behaviour only occur when using ml, or do you see this with other
preconditioners as well?



On Thursday, 27 June 2013, Thomas DE-SOZA wrote:

>
> Dear PETSc users,
>
> We've been using the Krylov solvers in PETSc for a while in our software
> package for implicit structural mechanics and recently added algebraic
> multigrid preconditioning through the ML and Hypre libraries provided as
> part of PETSc.
> Though we're quite happy with this new feature, we've encountered a
> strange behaviour when using KSPGCR in conjunction with the PCML
> preconditioner.
> Indeed, identical and  consecutives solves inside the same run do not
> display the same number of Krylov iterations even in sequential and we were
> wondering why :
>
> 1st solve
> 723 KSP unpreconditioned resid norm 2.911385065051e-02 true resid norm
> 2.911385065051e-02 ||r(i)||/||b|| 9.979426047969e-09
>
> 2nd solve
> 787 KSP unpreconditioned resid norm 2.896035212670e-02 true resid norm
> 2.896035212670e-02 ||r(i)||/||b|| 9.926810982197e-09
>
> 3rd solve
> 721 KSP unpreconditioned resid norm 2.913123687343e-02 true resid norm
> 2.913123687343e-02 ||r(i)||/||b|| 9.985385566274e-09
>
>
> Would you say this case requires a great number of iterations and
> therefore reproductibility is not ensured (indeed it is ill-conditioned and
> is not very suited for algebraic multigrid) ?
> Or is there a seed somewhere in ML that would explain this ? I searched
> the ML manual for that and couln't find any. We're using the uncoupled
> coarsening scheme in ML (call PetscOptionsSetValue('-pc_ml_CoarsenScheme',
> 'Uncoupled', ierr)).
>
> Final notes : several consecutive runs of the program do display identical
> behaviour (that is the 1st solve always require 723 iterations, the 2nd
> 787, etc). Moreover other cases that are well-conditioned require the same
> number number of iterations for each consecutive solve (though the
> converged residual differs a bit).
>
> Thanks for any hints,
> Thomas
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20130627/e6c3f66f/attachment-0001.html>