On Jan 23, 2008 10:17 AM, Vijay M <[EMAIL PROTECTED]> wrote:
> I understand your point of view but even Petsc can work with just residuals
> to obtain the Jacobian matrix (numerical Jacobian). But when the dof's are
> in the order of millions in a 3D case and I wouldn't even bother finding the
> numerical Jacobian and then eventually solving it. Doing it matrix-free with
> good preconditioners to reduce linear iterations can provide great CPU
> improvements.

Absolutely agree... NOX can use a jacobian computation as well... I
was just pointing out one feature of it.  Preconditioning of an
automatic jacobian in a matrix free solver is definitely a problem....

> Transparency here could affect performance. Just my 2 cents.

Definitely agree again... I just want to make sure that there exists a
transparent way to do this so we have the choice.  The way it is now
it looks great to me... I just didn't want too much Petsc-isms
sneaking into the interface.

Derek

>
> Vijay
>
>
> -----Original Message-----
> From: Derek Gaston [mailto:[EMAIL PROTECTED]
> Sent: Wednesday, January 23, 2008 10:40 AM
> To: Benjamin Kirk
> Cc: Vijay M; [email protected]
> Subject: Re: [Libmesh-devel] [Libmesh-users] Support for Matrix-free
> algorithms
>
> This is an interesting development...
>
> One thing I would warn against is tying ourselves too close to Petsc.
> I am currently waiting on a project to get green lighted where I will
> be tying Trilinos into libMesh... where we will have access to other
> matrix free solvers (like NOX, which can work with just residuals...
> doing finite differenced jacobians).  So coming up with a good
> interface where you actually don't interact with Petsc primitives
> would be ideal.  In most other places in the code you guys have done
> well at this exact thing... making it pretty easy to tie in a new set
> of linear/non-linear/transient solvers... no reason to go backwards
> now!
>
> So far, what you have here looks great!
>
> Just some food for thought...
>
> Derek
>
> On Jan 23, 2008 8:39 AM, Benjamin Kirk <[EMAIL PROTECTED]> wrote:
> > Here is a really, really raw example, the comments are not clear right
> now,
> > but I wanted to keep you informed.  This requires the latest svn branch to
> > work.
> >
> > Unpack it in the ./examples directory and run make.
> >
> > Run it as
> >
> > $ ./ex19-dbg -snes_view -r 4
> > for a successive approximation which will converge linearly, and
> >
> > $ ./ex19-dbg -snes_view -r 4 -snes_mf_operator
> > for a matrix-free approach in the iterative solver which will converge
> > quadratically.
> >
> > > Anyway, I do have a question regarding using PETSc object with LibMesh.
> I
> > > have been trying to use Petsc objects Mats, Vecs and SNES solver with
> > > Libmesh but the one thing I cant seem figure out is how to set the
> > > nonlinear_solver public attribute of say a NonlinearImplicitSystem
> object to
> > > a PETSc SNES object which I have created and initialized separately.
> Since,
> > > the SNES object used in the wrapper PetscNonlinearSolver is private, I
> don¹t
> > > understand how this can be done.
> > >
> > > Have I missed something and taken a completely wrong path on this ? I
> would
> > > very much appreciate any comments that you can provide to help me out
> here.
> >
> > The user interface is totally up for discussion since I am the only one
> who
> > as exercised it to date.  (I am sure Roy will have some comments!)  It
> seems
> > to me the right approach will be to add a method which gives the user
> access
> > to the SNES object?  From there the KSP, Mat, Vec, PC, etc... can be
> > accessed.  This would be similar to the approach used in the
> > PetscLinearSolver.
> >
> > -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
> >
> >
>
> No virus found in this incoming message.
> Checked by AVG Free Edition.
> Version: 7.5.516 / Virus Database: 269.19.9/1239 - Release Date: 1/23/2008
> 10:24 AM
>
>
> No virus found in this outgoing message.
> Checked by AVG Free Edition.
> Version: 7.5.516 / Virus Database: 269.19.9/1239 - Release Date: 1/23/2008
> 10:24 AM
>
>
>

-------------------------------------------------------------------------
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