> I agree about not tying LibMesh too much to PETSc but since there is a
> PetscNonlinearSolver class, I definitely want to use its full potential when
> solving nonlinear problems. And hence exposure to the actual snes object
> either through a property or function is necessary to have better control of
> how the nonlinear system is being solved. And that was my original request !
> There could always be a class for TrilinosNonlinearSolver or
> ...NonlinearSolver that could solve the system with Newton or FPI scheme
> from the corresponding packages but nevertheless, user needs control of the
> base objects. Transparency here could affect performance. Just my 2 cents.

The approach will undoubtedly parallel the linear solver class.  Everything
that is common & that can be abstracted will be; and it should be accessed
through the underlying NonlinearSolver class.

I've got no issue exposing the underlying PETSc or Trilinos objects in *an
unsupported way*.  John has done this in the past with the
PetscLinearSolver: he used a specific preconditioner through the PC object.
I doubt that will make it into any of the examples, though.

-Ben  


-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft
Defy all challenges. Microsoft(R) Visual Studio 2008.
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
_______________________________________________
Libmesh-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/libmesh-devel

Reply via email to