[petsc-users] Fwd: Any changes in ML usage between 3.1-p8 -> 3.3-p6?

[Mark F. Adams]

The only thing that I can think of is a change in icc or ml.c (the interface code).  I have looked over the logs for clues and do not see any more red flags.  So I would look at the logs:

https://bitbucket.org/petsc/petsc/history-node/cee652cbe889/src/ksp/pc/impls/ml/ml.c?page=2

and the similar one for icc (the history button when you have a file open).  See if you can see anything that could change the semantics.

Some of these changes are required to work with the rest of PETSc so doing a bisection search might be hard.  git does have a bisection utility (http://www.youtube.com/watch?v=X_pQfoaRuhA) that can help in doing this.  You could probably revert to the old version of these two files,  add any of the changes back in that are required for syntax (e.g., PetscTruth --> PetscBool) and get the code to compile and see if you get your old semantics back.  If so then add changes back (bisect might help to speed this up) and search for the change that broke you.

On Apr 23, 2013, at 2:57 PM, Jozsef Bakosi <jbakosi at lanl.gov> wrote:

> Hi Mark and Jed,
> 
> Thanks for your suggestions. I tried setting the zero pivot and shift as well,
> and those do not seem to help. See the attached output.
> 
>> On 04.18.2013 10:34, Jed Brown wrote:
>> 
>> As Mark says, use
>> 
>> -mg_levels_ksp_type chebyshev -mg_levels_pc_type jacobi
> 
> As Mark Christon said earlier, this is not really an option for us because it is
> a solution to problem that is different than what we are trying to solve.
> 
> The problem is not that we would like to have this particular problem converge,
> as BCGS or Hypre handles this problem just fine. What we are trying to track
> down is what causes the change in PETSc's behavior going between these two
> versions, while keeping the application and ML source code the same, as that
> might point out an error on our end, e.g. using an undocumented option, or
> relying on a default that has changed, etc. The zeropivot and the shift,
> pointing to their diff was a good catch, but it seems like there might be
> something else as well that causes a different behavior.
> 
> Thanks,
> Jozsef
> <out.txt>