[petsc-users] Ghost point communication for Red-Black Gauss-Siedel

[Matthew Knepley]

On Wed, Feb 15, 2012 at 9:14 PM, Sylvain Barbot <sylbar.vainbot at gmail.com>wrote:

> > R-B has crappy cache performance since you alternately pull data but
> ignore
> > half of it.
>
> Understood.
>
> > Why not just use the GMG in PETSc? Its very flexible, scalable,
> > and efficient.
>
> I tried to use Petsc's implementation of multigrid, cf thread "V-cycle
> multigrid with matrix shells". My concern is performance. I'd like the
> whole procedure to use matrix shells, or "matrix-free" matrices,
> because my stencil can have up to 21 off-diagonal terms. Petsc 3.1 did
> not allow this functionality. I do not know what Petsc 3.2 has to
> offer in that regard.
>

So you want to calculate the action to avoid the memory bandwidth limit? If
so,
you can still use MG by specifying the coarse operators as MatShells,
rather than
using Galerkin. You might want to tweak the interpolator later if it is
inadequate.

   Matt


> I have a working multi-grid method that uses the low functionality of
> Petsc. I would obviously prefer if I could use something more
> advanced.
>
> Cheers,
> Sylvain
>



-- 
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/20120215/502dc0e7/attachment.htm>