[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: adding SNESSetLinearSolve()


On Nov 1, 2007, at 10:16 AM, Matthew Knepley wrote:

Actually, making the regularization parameter independent for each
process is much more efficient. Gene Golub had a poster on this at
the last SIAM CS&E meeting.

  Matt

On Nov 1, 2007 9:24 AM, Lisandro Dalcin <dalcinl@xxxxxxxxx> wrote:
On 10/31/07, Barry Smith <bsmith@xxxxxxxxxxx> wrote:
Lisandro,
A followup to our previous discussion. It sounds to me like you
are "actually" solving an n+1 unknown nonlinear problem where the
"special" unknown is kept secret from SNES and managed somehow by the
application code?

That's exactly the case. Furthermore, this 'special' unknown in just a
regularization parameter who tends to zero as the nonlinear solution
is reached. Unfortunatelly, this unknow is coupled with all other
degree of freedom, thus generating a full dense row and a dense column
in the Jacobian matrix. But fortunatelly, the special unknown is just
a single scalar, thus computing the schur complement is feasible, but
requires two linear solves with the other 'sparse' block of the
Jacobian.

The point is that your ComputeJacobian routine would NEVER compute this Jacobian
with the final row/column that are dense. You would use a shell matrix for the full beast and
in side it store the n by n sparse beast in the usual manner, the shell PC would then
use the two parts of the shell matrix to do its thing with the two solves.

I admit I do not have a handle on what it means to move the Eisenstat-Walker test
inside the outer-most KSP to the two inner linear solves. Either theoretically or
practically in code.



You can guess how I feel about this :-).

Yes, of course. I agree that PETSc API must be consistent and clean.
But I also feel that some time I need more features. Please remember I
use PETSc exclusively from Python. And then is so easy to manage
complicated application setups. But at some point I need more
low-level support from PETSc to make it working.

For example, I would love to have SNESSetPresolve/SNESSetPostSolve and
SNESSetPreStep/SNESSetPostStep, and perhaps a
SNESSetPreLinearSolve/SNESSetPostLinearSolve. Of course, this make the
API grow with features that are not frecuently needed.

I'd like to understand why they are ever needed and if they can be done
with current constructs?

   Barry



PETSc/SNES is suppose
to be good enough to allow you to feed it the ENTIRE nonlinear problem
in a way that would allow as efficient solution as if you "handled the
special unknown" specially. 

Even for my particular case? Can I pass-over the issue with the full
denses rows and columns?

In particular for this problem the intended
solution is to use the PETSc DMComposite object via DMCompositeCreate()
You may want to look at this construct to see if it is suitable
for your friends needs. And what we need to add (note that though DMComposite
can be used with DMMG it does not have to be, it can be used directly
with a SNES also. 

I'll definitely take a loop ASAP.

Regards,

--
Lisandro Dalcín
---------------
Centro Internacional de Métodos Computacionales en Ingeniería (CIMEC)
Instituto de Desarrollo Tecnológico para la Industria Química (INTEC)
Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET)
PTLC - Güemes 3450, (3000) Santa Fe, Argentina
Tel/Fax: +54-(0)342-451.1594





--
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