On Fri, Mar 15, 2013 at 1:19 PM, Jens Lohne Eftang <[email protected]> wrote:
> On 03/15/2013 03:16 PM, John Peterson wrote:
>>
>> On Fri, Mar 15, 2013 at 1:15 PM, John Peterson <[email protected]>
>> wrote:
>>>
>>> On Fri, Mar 15, 2013 at 1:03 PM, Roy Stogner <[email protected]>
>>> wrote:
>>>>
>>>> On Fri, 15 Mar 2013, Jens Lohne Eftang wrote:
>>>>
>>>>> On 03/15/2013 02:01 PM, Roy Stogner wrote:
>>>>>>
>>>>>> I'd start by turning on PETSc's monitor flags ('-ksp_monitor' etc?).
>>>>>> Endless residual reevaluations suggests to me that the initial
>>>>>> linear solve converged poorly and didn't give you a descent direction,
>>>>>> and a subsequent line search is failing to find a reduced nonlinear
>>>>>> residual.
>>>>>
>>>>> It seems to me that the residual gets assembled after the solver has
>>>>> converged, for some reason.
>>>>
>>>> Yes: After you take a Newton step, the Newton solver typically wants
>>>> to make sure that that step gave you a (nonlinear) residual reduction,
>>>> and to see whether that new residual meets your tolerances so that the
>>>> Newton iteration can claim convergence and exit.
>>>>
>>>> If you got a good linear solve (so the new proposed solution is a
>>>> descent direction) but the next nonlinear residual is lousy, then
>>>> typically the next operation is a line search, which does a whole
>>>> bunch more residual evaluations at points in between the solution at
>>>> the previous nonlinear solve and the failed proposed next step.
>>>>
>>>> But it looks like you're getting a good linear solve yet not getting a
>>>> good line search result.  Is it possible your residual and Jacobian
>>>> aren't consistent?
>>>
>>> Petsc 3.3?  You can also try "-snes_linesearch_type basic" to turn
>>> linesearch off completely
>>>
>>> Prior to 3.3 it was "-snes_ls basic"
>>
>> Wait, nevermind, you said FEMSystem, which is base on ImplicitSystem,
>> sorry.  Those options won't do anything.
>
> Well, it did do something ... "-snes_linesearch_type basic" makes the solver
> do multiple linear solves instead. That does not make any sense to me as I'm
> just trying to solve Ax=b.

Well I think FEMSystem solves everything as though it were nonlinear.
So if you make your ksp_rtol (linear tolerance) about the same as your
nonlinear tolerance, the "nonlinear" solver should converge in 1
iteration.

-- 
John

------------------------------------------------------------------------------
Everyone hates slow websites. So do we.
Make your web apps faster with AppDynamics
Download AppDynamics Lite for free today:
http://p.sf.net/sfu/appdyn_d2d_mar
_______________________________________________
Libmesh-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/libmesh-users

Reply via email to