Hi there,

Before I even start, I know this is probably going to be a loaded
question, but nevertheless, any help would be appreciated.

Background: I'm working on solution to large scale PDE systems based
on libmesh for discretization and petsc for solution procedures. My
physics objects are derived from EquationSystems in libmesh
datastructure and can contain one or more implicitsystems in it. I use
the Jacobian-Free Krylov scheme for my nonlinear solve and hence the
efficiency of the method is achieved from good preconditioning. I've
tried using ILU to precondition my preconditioner (linearized version)
of the original nonlinear system and this works well for serial runs.
But this does not scale when I run in parallel.

So my questions are

1) Is it is possible to use libmesh objects to perform geometric
multigrid preconditioning and completely avoid the matrix creation
except in the coarsest case ?! I've seen some of Ben's presentation
that mentions using multigrid for stokes problems etc. and so am
curious as to what this would entail ... For example, would I have to
create multiple  equationsystems object since the mesh is not
associated with just the system. Hence for every multi-grid level, I
would have a new instance of the physics system but on a coarser mesh
level. Is this how you would possibly implement this ?!
2) If I can afford to create the matrix in the finest level, can we
use algebraic multigrid procedures to precondition this system in
parallel ? Anyone know how this scales ?
3) Also, I was reading about Prometheus
(http://www.columbia.edu/~ma2325/prom_intro.html) and it looks
promising as both a black-box solver and preconditioner when the
matrix is available and geometric multigrid in 2D (not sure about 3D).
Has anyone in this list used this package with Petsc and/or libMesh ?!

Again, like I said before, there is a lot of things here you can
comment on. Feel free to write your thoughts because I want to get as
much input as possible before choosing anything specific.

Thanks,
Vijay

-------------------------------------------------------------------------
This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
Build the coolest Linux based applications with Moblin SDK & win great prizes
Grand prize is a trip for two to an Open Source event anywhere in the world
http://moblin-contest.org/redirect.php?banner_id=100&url=/
_______________________________________________
Libmesh-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/libmesh-users

Reply via email to