[petsc-users] snes failures

[Barry Smith]

> On May 20, 2016, at 5:17 PM, Juha Jäykkä <juhaj at iki.fi> wrote:
> 
>>>>  1) Grid sequencing: This is easy if you use a DMDA. You just use
>>>> -snes_grid_sequence
>>>>      and its automatic. Since you report that smaller grids converge,
>>> this
>>>> is usually enough.
>>> Unfortunately, no.
>> Isn't this a 1D problem with no geometry? You should use DMDA. It would make
>> it easier.
> 
> I think you misunderstood my too-short-a-comment. What I meant was 
> unfortunately -snes_grid_sequence was not enough to make it converge. I went 
> up to -snes_grid_sequence 10.

   Interesting. How much is the solution changing in each refinement? Is there a singularity in the solution at one end point, or elsewhere?  If you know the form of the singularity perhaps you can subtract it out from the solution (and this hence changes the function) to get a simpler problem without a singularity that Newton works better on.

   Barry

> 
> I always use DMDA and yes, the only geometry in the problem is a straight line 
> of points from -X to +X, where preferably X = infinity, but numerically of 
> course not.
>>> I'm not sure what you mean here. Do you mean I should run normal newton
>>> line
>>> search snes with SNESSetNPC(snes_of_type_gs)?
>> Yes.
> 
> Will do that once I get back to this next week.
> 
>>> I wish it was so easy: I did try all the snes types before posting the
>>> first
>>> post. Also all KSP types. Except those that need things I don't have, of
>>> course.
>> Not sure what you mean here. To use FAS, you can either use DMDA, or provide
>> interpolation operators between grids.
> 
> Yes, but FAS does not converge either. Neither does any other snes type except 
> ngs, and changing to any other KSP type makes no difference either – though 
> why would it as the KSP seems to converge nicely anyway. I have also tried all 
> snes LS types, to no avail.
> 
> A correction: funnily enough, newtontr claims it converges, but it does not 
> really. It ends up with the same "solution" as almost everything else does and 
> thinks it is a solution (CONVERGED_SNORM_RELATIVE) whereas I know full well it 
> is not a solution, not even close. But I always found newtontr fiddly anyway: 
> it tends to be too trigger happy to shout "convergence" and finding the right 
> parameters to make it not-so-trigger-happy is hard.
> 
> Cheers,
> Juha
> 
> -- 
> 		 -----------------------------------------------
> 		| Juha Jäykkä, juhaj at iki.fi			|
> 		| http://koti.kapsi.fi/~juhaj/			|
> 		 -----------------------------------------------