[petsc-users] KSP "randomly" not converging

[Matthew Knepley]

On Tue, Jun 2, 2015 at 12:26 PM, Italo Tasso <italo at tasso.com.br> wrote:

> I made a code to solve the Navier-Stokes equations, incompressible,
> non-linear, all coupled, finite differences, staggered grid.
>
> I am running the code with:
>
> -ts_monitor -snes_monitor -ksp_monitor_true_residual
> -snes_converged_reason -ksp_converged_reason -pc_type fieldsplit
> -pc_fieldsplit_type schur -pc_fieldsplit_detect_saddle_point
>
> It works very well most of the time. But in some cases, the solver halts
> for a long time then KSP does not converge.
>
> See output1.txt. It seems that the residual is already very small, close
> to machine zero, but KSP doesn't stop.
>
> So I added -ksp_atol 1e-10. See output2.txt. Now it fails on a different
> time step.
>
> I also tried -ksp_norm_type unpreconditioned. It works for this case
> (grid size), but fail for other cases.
>
> I also tried building the Jacobian and including null space. It fixes some
> cases but causes others that worked before to fail. Seems really random.
>
> It feels like this is related to the PC, because the code halts for a long
> time at the first KSP step, then diverges.
>
> Any suggestions?
>

Yes, this is related to your preconditioner. If you have a null space, you
have to project it out. However,

      0 KSP preconditioned resid norm 1.044238592402e-03 true resid
norm 9.729564145362e-11 ||r(i)||/||b|| 1.000000000000e+00
      1 KSP preconditioned resid norm 1.044238592401e-03 true resid
norm 6.820544672134e-10 ||r(i)||/||b|| 7.010123547399e+00
      2 KSP preconditioned resid norm 1.044238592401e-03 true resid
norm 8.445969119028e-10 ||r(i)||/||b|| 8.680727104362e+00

this is something strange. Your preconditioner has changed the solution to
your problem. It appears ILU (which I assume you are using.
you should ways send -ksp_view) has broken down completely. It is
unreliable in the extreme.

   Matt

-- 
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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20150602/1c18b275/attachment.html>