[petsc-users] SNES convergence tol question

[Peter Brune]

What you're seeing is convergence in the step tolerance, not the relative
tolerance applied to the norm of the residual.  This means that the norm of
the steplength between successive iterations has gone below snes->xtol
times the norm of the solution.  This appears to be set through -snes_stol.

- Peter

On Tue, Mar 20, 2012 at 10:29 AM, Mark F. Adams <mark.adams at columbia.edu>wrote:

> We've been trying to drive the nonlinear residual down and SNES seems to
> return  prematurely.
>
> Appended is Dan's note.  We are asking for -snes_rtol 1.e-12 but only
> getting ~1.e-8.
>
> Anyone have any ideas?
>
> Mark
>
> > Nonlinear solve converged due to CONVERGED_PNORM_RELATIVE iterations 8
>
> However, here are the initial and final norms:
>
> >   0 SNES Function norm 1.999411334131e+08
> >   8 SNES Function norm 5.999267565004e-01
>
> 2e8/6e-1 = 3.3e8. So, I'm confused. Am I misunderstanding the convergence
> criterion?
>
> here's the dump of petsc options at the end:
>
> > #PETSc Option Table entries:
> > -ksp_converged_use_initial_residual_norm
> > -ksp_max_it 100
> > -ksp_monitor
> > -ksp_rtol 1.e-6
> > -ksp_type gmres
> > -mg_levels_ksp_max_it 4
> > -options_left
> > -pc_gamg_agg_nsmooths 1
> > -pc_gamg_sym_graph
> > -pc_gamg_threshold .01
> > -pc_gamg_type agg
> > -pc_gamg_verbose 2
> > -pc_hypre_boomeramg_agg_nl 1
> > -pc_hypre_boomeramg_coarsen_type HMIS
> > -pc_hypre_boomeramg_interp_type ext+i
> > -pc_hypre_boomeramg_no CF
> > -pc_hypre_type boomeramg
> > -pc_type gamg
> > -snes_converged_reason
> > -snes_mf_operator
> > -snes_monitor
> > -snes_rtol 1.e-12
> > #End of PETSc Option Table entries
>
>
>
> Mark
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120320/9312c64e/attachment.htm>