Re: [Libmesh-users] using PCGAMG

[Ataollah M. <am...@ti...>]

On May 9, 2013, at 4:00 PM, Dmitry Karpeev <ka...@mc...> wrote:

> GAMG is primarily an algebraic multigrid preconditioner.
> The geometric part is somewhat weaker and isn't the primary purpose of GAMG, 
> so I would recommend using -pc_gamg_type agg for plain aggregation.
> 

OK that's clear know, I was asking about the low level mesh implementation because I thought the geometric part will need to know something about the coordinates and what not.


> For good performance with elasticity it is important to use smoothed aggregation,
> for example,  -pc_gamg_agg_nsmooths 1
> but for that to be effective, it is critical to supply a "nearnullspace" -- basically, 
> the nullspace of the Neumann operator.  It is the rigid body modes in the case of elasticity.
> If you are able to specify the block size of the matrix, GAMG will try to guess the correct
> nearnullspace for you.  In particular bs=3 will cause GAMG to construct the rigid body modes
> from the coordinates set with PCSetCoordinates. 
> 
> Now, libMesh doesn't have an API for setting the matrix block size (as far as I know).
> You can hack a bit: extract the PETSc objects and do it yourself, but a more natural way is 
> to compute your own rigid body modes in a callback attached to NonlinearSolver::nearnullspace or NonlinearSolver::nearnullspace_object.  A nearnullspace is simply an array of NumericVector that 
> contains a basis for the nearnullspace. libMesh will orthonormalize it for you (PETSc requires 
> that).
> 

Crystal clear. I'll get to it.

Best,
Ata

> Hope this helps.
> Dmitry.
> 
> 
> On Thu, May 9, 2013 at 11:08 AM, Cody Permann <cod...@gm...> wrote:
> On Thu, May 9, 2013 at 9:18 AM, Ataollah Mesgarnejad <am...@ti...
> > wrote:
> 
> >
> > Cody,
> >
> > On May 9, 2013, at 9:29 AM, Cody Permann <cod...@gm...> wrote:
> >
> >
> >
> >
> > On Thu, May 9, 2013 at 8:07 AM, Ataollah Mesgarnejad <
> > am...@ti...> wrote:
> >
> >> Dear all,
> >>
> >> I was wondering if anyone has any experience using gamg with libMesh (I'm
> >> particularly interested in elasticity)?
> >>
> >>
> > It certainly works, however in my very limited experiecne, it's slightly
> > more tricky to use than Hypre as we've found that it requires a few more
> > options.  Gamg is very flexible and part of PETSc so it ties in nicely.
> >
> >
> >
> > That's wonderful. Do I need to add anything in the code itself like block
> > size or ..? Or do you mean the command line options for gamg?
> >
> 
> Just the command line options. If you have specific questions, the petsc
> user list is a great place to ask.
> 
> Cody
> 
> >
> > My main question would be if the geometric part of gamg would work with
> >> the underlying mesh structure in libMesh?
> >>
> >
> > I'm not sure if I'm missing something with this question.  I don't see why
> > this preconditioner would care about the mesh structure, it's abstracted
> > well above that level.  From libMesh's perspective, it's just another
> > preconditioner.
> >
> >
> > OK. I'm not very familiar with low level structure in libMesh so I wasn't
> > sure.
> >
> > Thanks,
> > Ata
> >
> >
> > Cody
> >
> >
> >>
> >> I would appreciate any comments/recommendations,
> >> Ata
> >>
> >> ------------------------------------------------------------------------------
> >> Learn Graph Databases - Download FREE O'Reilly Book
> >> "Graph Databases" is the definitive new guide to graph databases and
> >> their applications. This 200-page book is written by three acclaimed
> >> leaders in the field. The early access version is available now.
> >> Download your free book today! http://p.sf.net/sfu/neotech_d2d_may
> >> _______________________________________________
> >> Libmesh-users mailing list
> >> Lib...@li...
> >> https://lists.sourceforge.net/lists/listinfo/libmesh-users
> >>
> >
> >
> >
> ------------------------------------------------------------------------------
> Learn Graph Databases - Download FREE O'Reilly Book
> "Graph Databases" is the definitive new guide to graph databases and
> their applications. This 200-page book is written by three acclaimed
> leaders in the field. The early access version is available now.
> Download your free book today! http://p.sf.net/sfu/neotech_d2d_may
> _______________________________________________
> Libmesh-users mailing list
> Lib...@li...
> https://lists.sourceforge.net/lists/listinfo/libmesh-users
> 
> 
> 
>