[petsc-users] PetscDSSetJacobianPreconditioner causing DIVERGED_LINE_SEARCH for multi-field problem
Sander Arens
Sander.Arens at UGent.be
Sun Apr 3 06:55:09 CDT 2016
Any thoughts on this, Matt?
Thanks,
Sander
On 21 March 2016 at 09:49, Sander Arens <Sander.Arens at ugent.be> wrote:
> Matt,
>
> I think the problem is in DMCreateSubDM_Section_Private. The block size
> can be counted there (very similar as in DMCreateMatrix_Plex) and then
> passed down to the IS. I can create a pull request for this if you'd like
> (or should I just add this to my other pull request for the Neumann bc's)?
>
> On 3 March 2016 at 18:02, Sander Arens <Sander.Arens at ugent.be> wrote:
>
>> Ok, thanks for clearing that up!
>>
>> On 3 March 2016 at 18:00, Matthew Knepley <knepley at gmail.com> wrote:
>>
>>> On Thu, Mar 3, 2016 at 10:48 AM, Sander Arens <Sander.Arens at ugent.be>
>>> wrote:
>>>
>>>> So is the block size here refering to the dimension of the
>>>> near-nullspace or has it something to do with the coarsening?
>>>>
>>>> I would have thought to also see the block size in this part:
>>>>
>>>
>>> Dang, that is a problem. It should have the correct block size when FS
>>> pulls it out of the overall matrix. This should be
>>> passed down using the block size of the IS. Something has been broken
>>> here. I will put it on my list of things to look at.
>>>
>>> Thanks,
>>>
>>> Matt
>>>
>>>
>>>> Mat Object: 1 MPI processes
>>>> type: seqaij
>>>> rows=1701, cols=1701
>>>> total: nonzeros=109359, allocated nonzeros=109359
>>>> total number of mallocs used during MatSetValues
>>>> calls =0
>>>> has attached near null space
>>>> using I-node routines: found 567 nodes, limit used
>>>> is 5
>>>>
>>>> Thanks,
>>>> Sander
>>>>
>>>>
>>>> On 3 March 2016 at 17:29, Matthew Knepley <knepley at gmail.com> wrote:
>>>>
>>>>> You can see the block size in the output
>>>>>
>>>>> Mat Object: 1 MPI processes
>>>>> type: seqaij
>>>>> rows=12, cols=12, bs=6
>>>>> total: nonzeros=144, allocated nonzeros=144
>>>>> total number of mallocs used during MatSetValues
>>>>> calls =0
>>>>> using I-node routines: found 3 nodes, limit used
>>>>> is 5
>>>>> linear system matrix = precond matrix:
>>>>> Mat Object: 1 MPI processes
>>>>> type: seqaij
>>>>> rows=12, cols=12, bs=6
>>>>> total: nonzeros=144, allocated nonzeros=144
>>>>> total number of mallocs used during MatSetValues calls
>>>>> =0
>>>>> using I-node routines: found 3 nodes, limit used is 5
>>>>>
>>>>> Thanks,
>>>>>
>>>>> Matt
>>>>>
>>>>> On Thu, Mar 3, 2016 at 10:28 AM, Sander Arens <Sander.Arens at ugent.be>
>>>>> wrote:
>>>>>
>>>>>> Sure, here it is.
>>>>>>
>>>>>> Thanks,
>>>>>> Sander
>>>>>>
>>>>>> On 3 March 2016 at 15:33, Matthew Knepley <knepley at gmail.com> wrote:
>>>>>>
>>>>>>> On Thu, Mar 3, 2016 at 8:05 AM, Sander Arens <Sander.Arens at ugent.be>
>>>>>>> wrote:
>>>>>>>
>>>>>>>> Yes, the matrix is created by one of the assembly functions in
>>>>>>>> Plex, so that answers my question. I was confused by this because I
>>>>>>>> couldn't see this information when using -snes_view.
>>>>>>>>
>>>>>>>
>>>>>>> Hmm, then something is wrong because the block size should be
>>>>>>> printed for the matrix at the end of the solver in -snes_view, Can you
>>>>>>> send that to me?
>>>>>>>
>>>>>>> Thanks,
>>>>>>>
>>>>>>> Matt
>>>>>>>
>>>>>>>
>>>>>>>> Thanks for the helpful replies,
>>>>>>>> Sander
>>>>>>>>
>>>>>>>> On 3 March 2016 at 14:52, Matthew Knepley <knepley at gmail.com>
>>>>>>>> wrote:
>>>>>>>>
>>>>>>>>> On Thu, Mar 3, 2016 at 7:49 AM, Sander Arens <
>>>>>>>>> Sander.Arens at ugent.be> wrote:
>>>>>>>>>
>>>>>>>>>> And how can I do this? Because when I look at all the options
>>>>>>>>>> with -help I can strangely enough only find
>>>>>>>>>> -fieldsplit_pressure_mat_block_size and not
>>>>>>>>>> -fieldsplit_displacement_mat_block_size.
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Maybe I am missing something here. The matrix from which you
>>>>>>>>> calculate the preconditioner using GAMG must be created somewhere. Why
>>>>>>>>> is the block size not specified at creation time? If the matrix is
>>>>>>>>> created by one of the assembly functions in Plex, and the PetscSection has
>>>>>>>>> a number of field components, then the block size will already be
>>>>>>>>> set.
>>>>>>>>>
>>>>>>>>> Thanks,
>>>>>>>>>
>>>>>>>>> Matt
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>> Thanks,
>>>>>>>>>> Sander
>>>>>>>>>>
>>>>>>>>>> On 3 March 2016 at 14:21, Matthew Knepley <knepley at gmail.com>
>>>>>>>>>> wrote:
>>>>>>>>>>
>>>>>>>>>>> On Thu, Mar 3, 2016 at 6:20 AM, Sander Arens <
>>>>>>>>>>> Sander.Arens at ugent.be> wrote:
>>>>>>>>>>>
>>>>>>>>>>>> Ok, I forgot to call SNESSetJacobian(snes, J, P, NULL, NULL)
>>>>>>>>>>>> with J != P, which caused to write the mass matrix into the (otherwise
>>>>>>>>>>>> zero) (1,1) block of the Jacobian and which was the reason for the
>>>>>>>>>>>> linesearch to fail.
>>>>>>>>>>>> However, after fixing that and trying to solve it with
>>>>>>>>>>>> FieldSplit with LU factorization for the (0,0) block it failed because
>>>>>>>>>>>> there were zero pivots for all rows.
>>>>>>>>>>>>
>>>>>>>>>>>> Anyway, I found out that attaching the mass matrix to the
>>>>>>>>>>>> Lagrange multiplier field also worked.
>>>>>>>>>>>>
>>>>>>>>>>>> Another related question for my elasticity problem: after
>>>>>>>>>>>> creating the rigid body modes with DMPlexCreateRigidBody and attaching it
>>>>>>>>>>>> to the displacement field, does the matrix block size of the (0,0) block
>>>>>>>>>>>> still have to be set for good performance with gamg? If so, how can I do
>>>>>>>>>>>> this?
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> Yes, it should be enough to set the block size of the
>>>>>>>>>>> preconditioner matrix.
>>>>>>>>>>>
>>>>>>>>>>> Matt
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>> Thanks,
>>>>>>>>>>>> Sander
>>>>>>>>>>>>
>>>>>>>>>>>> On 2 March 2016 at 12:25, Matthew Knepley <knepley at gmail.com>
>>>>>>>>>>>> wrote:
>>>>>>>>>>>>
>>>>>>>>>>>>> On Wed, Mar 2, 2016 at 5:13 AM, Sander Arens <
>>>>>>>>>>>>> Sander.Arens at ugent.be> wrote:
>>>>>>>>>>>>>
>>>>>>>>>>>>>> Hi,
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> I'm trying to set a mass matrix preconditioner for the Schur
>>>>>>>>>>>>>> complement of an incompressible finite elasticity problem. I tried using
>>>>>>>>>>>>>> the command PetscDSSetJacobianPreconditioner(prob, 1, 1, g0_pre_mass_pp,
>>>>>>>>>>>>>> NULL, NULL, NULL) (field 1 is the Lagrange multiplier field).
>>>>>>>>>>>>>> However, this causes a DIVERGED_LINE_SEARCH due to to Nan or
>>>>>>>>>>>>>> Inf in the function evaluation after Newton iteration 1. (Btw, I'm using
>>>>>>>>>>>>>> the next branch).
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Is this because I didn't use PetscDSSetJacobianPreconditioner
>>>>>>>>>>>>>> for the other blocks (which uses the Jacobian itself for preconditioning)?
>>>>>>>>>>>>>> If so, how can I tell Petsc to use the Jacobian for those blocks?
>>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> 1) I put that code in very recently, and do not even have
>>>>>>>>>>>>> sufficient test, so it may be buggy
>>>>>>>>>>>>>
>>>>>>>>>>>>> 2) If you are using FieldSplit, you can control which blocks
>>>>>>>>>>>>> come from A and which come from the preconditioner P
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> http://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/PC/PCFieldSplitSetDiagUseAmat.html#PCFieldSplitSetDiagUseAmat
>>>>>>>>>>>>>
>>>>>>>>>>>>> http://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/PC/PCFieldSplitSetOffDiagUseAmat.html#PCFieldSplitSetOffDiagUseAmat
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>> I guess when using PetscDSSetJacobianPreconditioner the
>>>>>>>>>>>>>> preconditioner is recomputed at every Newton step, so for a constant mass
>>>>>>>>>>>>>> matrix this might not be ideal. How can I avoid recomputing this at every
>>>>>>>>>>>>>> Newton iteration?
>>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> Maybe we need another flag like
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> http://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/SNES/SNESSetLagPreconditioner.html
>>>>>>>>>>>>>
>>>>>>>>>>>>> or we need to expand
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> http://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/SNES/SNESSetLagJacobian.html
>>>>>>>>>>>>>
>>>>>>>>>>>>> to separately cover the preconditioner matrix. However, both
>>>>>>>>>>>>> matrices are computed by one call so this would
>>>>>>>>>>>>> involve interface changes to user code, which we do not like
>>>>>>>>>>>>> to do. Right now it seems like a small optimization.
>>>>>>>>>>>>> I would want to wait and see whether it would really be
>>>>>>>>>>>>> maningful.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Thanks,
>>>>>>>>>>>>>
>>>>>>>>>>>>> Matt
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>> Thanks,
>>>>>>>>>>>>>> Sander
>>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> --
>>>>>>>>>>>>> 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
>>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> --
>>>>>>>>>>> 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
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> --
>>>>>>>>> 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
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> --
>>>>>>> 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
>>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> 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
>>>>>
>>>>
>>>>
>>>
>>>
>>> --
>>> 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/20160403/e9d2a849/attachment-0001.html>
More information about the petsc-users
mailing list