[CIG-SHORT] Linear convergence for nonplanar faults

Fri Nov 11 06:52:19 PST 2016

Matt,

Thank you for taking the time to look at this.

I went through your slides and the PETSc <https://www.mcs.anl.gov/petsc/>
manual and found that the parameters shown here, at the bottom of my
e-mail, solve my problem. I am happy with it. However, would you be able to
look at these parameters and let me know if you find anything strange about
it ?

Regarding your suggestions, I combined my parameters with the following:

pylith no_fault.cfg fault.cfg ./solvers/solver_fault_schur_custompc.cfg >
log.out

pylith no_fault.cfg fault.cfg ./solvers/solver_fault_fieldsplit.cfg >
log.out

pylith no_fault.cfg fault.cfg ./solvers/solver_fault_exact.cfg > log.out

and in all cases I get the following Error message:

KSPSolve has not converged

In this specific case, it seems that when I use field split option,
combined with the other parameter below, I can't get the linear solver to
converge. However, I am not sure why this happens. Can you clarify ? Can
you clarify also if I my result would be jeopardized by not using the field
split option that you suggest ?

Thank you,
Josimar

##################################################### The parameters below
works for me .
## Set the solver options.
[pylithapp.petsc]

## Preconditioner settings.
pc_type = asm
sub_pc_type=lu
sub_pc_factor_shift_type = nonzero

## Convergence parameters.
ksp_max_it = 4000
ksp_gmres_restart = 30

## Linear solver monitoring options.
ksp_type=gmres
ksp_monitor = true
ksp_converged_reason = true
ksp_error_if_not_converged = true
ksp_view=true
ksp_monitor_true_residual=true

# Nonlinear solver monitoring options.
snes_max_it = 10000
snes_monitor = true
snes_linesearch_monitor = true
snes_converged_reason = true
snes_error_if_not_converged = true
snes_view=true
#snes_max_funcs = 3000
snes_check_jacobian_view=true
snes_ls_monitor=true
snes_monitor_short=true

# PETSc summary -- useful for performance information.
log_view = true

# Uncomment to launch gdb when starting PyLith.
#start_in_debugger = true

#########Fault Paramerters

ksp_rtol = 1.0e-20
ksp_atol = 1.0e-11

snes_rtol = 1.0e-20
snes_atol = 1.0e-8

#########################

## ----------------------------------------------------------------------
## NOTE: There are additional settings specific to fault friction.
[pylithapp.petsc]

### Friction sensitivity solve used to compute the increment in slip
### associated with changes in the Lagrange multiplier imposed by the
### fault constitutive model.
friction_pc_type = asm
friction_sub_pc_type=lu

friction_ksp_type=gmres
friction_sub_pc_factor_shift_type = nonzero
friction_ksp_max_it = 500
friction_ksp_gmres_restart = 30

### Uncomment to view details of friction sensitivity solve.
friction_ksp_converged_reason = true

##############################################################################################

On Thu, Nov 10, 2016 at 6:33 PM, Matthew Knepley <knepley at rice.edu> wrote:

> On Wed, Nov 9, 2016 at 9:50 AM, Matthew Knepley <knepley at rice.edu> wrote:
>
>> On Wed, Nov 9, 2016 at 9:42 AM, Josimar Alves da Silva <
>> jsilva.mit at gmail.com> wrote:
>>
>>> Brad,
>>>
>>> I changed the boundary conditions for Dirichlet BC with constrained
>>> displacement perpendicular to the surfaces, but the problem still persists.
>>> I am starting to wonder if the nonplanar nature of the fault surface may be
>>> causing the problem.
>>>
>>> The reason for this hypothesis is that I have the same set up but now
>>> with flat surfaces, and then in this case I don't find any convergence
>>> issue. Everything else was kept the same, except the fault geometry was
>>> changed from slightly smoothed (the version I am having problems) to not
>>> smoothed.
>>>
>>
>> You NEED to use the exact version of the solver and check for
>> convergence, just as in the slides. This will tell us whether your
>> formulation is screwed up.
>>
>
> I want to follow up here. What i mean is that you need to run with exact
> solvers for the two blocks of the
> Schur form, and I would test without the custom PC. This is described in
> the slides, and also given in the
> schur_exact cfg file. Have you been able to run this?
>
>   Thanks,
>
>      Matt
>
>
>>    Matt
>>
>>
>>> Let me know what you think about this.
>>> Thank you
>>> Josimar
>>>
>>>
>>>
>>>
>>> On Wed, Nov 9, 2016 at 10:20 AM, Brad Aagaard <baagaard at usgs.gov> wrote:
>>>
>>>> On 11/09/2016 07:15 AM, Josimar Alves da Silva wrote:
>>>>
>>>>> Dear Brad, Matt and Charles,
>>>>>
>>>>> Thank you again for the suggestions. After reviewing your comments I
>>>>> performed the following task in order to solve my problem:
>>>>>
>>>>> Brad mentioned:
>>>>>
>>>>>  "1. Without a fault, check convergence with using ML for
>>>>> the preconditioned."
>>>>>
>>>>>          Result: the solver converges fine without the fault and using
>>>>> the specified preconditioned.
>>>>>
>>>>> "2. Add the fault and with zero prescribed slip (FaultCohesiveKin),
>>>>> examine the convergence with the solver_fault_fieldsplit.cfg
>>>>> parameters."
>>>>>
>>>>>          Results: no convergence for any solver. Error message:  Linear
>>>>> solve did not converge due to DIVERGED_ITS iterations 10000
>>>>>
>>>>
>>>> This suggests there is something wrong with your problem setup. Make
>>>> sure you have sufficient Dirichlet BC on each side of the fault to prevent
>>>> rigid body motion.
>>>>
>>>> Brad
>>>>
>>>>
>>>> _______________________________________________
>>>> CIG-SHORT mailing list
>>>> CIG-SHORT at geodynamics.org
>>>> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/cig-short
>>>>
>>>
>>>
>>> _______________________________________________
>>> CIG-SHORT mailing list
>>> CIG-SHORT at geodynamics.org
>>> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/cig-short
>>>
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.geodynamics.org/pipermail/cig-short/attachments/20161111/d06805bd/attachment-0001.html>