[Trilinos-Users] Epetra Matrices for ML from MetisPartitioned Matrices

[Williams, Alan B]

The Epetra_FECrsMatrix is intended to handle the scenario where
finite-element data (e.g. element-stiffness submatrices) is assembled
into the matrix object, a submatrix at a time, without worrying about
whether a given element is connected to boundary nodes or not.
After data assembly is complete, the method 'globalAssemble()' is
called, and that method moves shared data onto owning processors (owning
processors determined by the non-overlapping Epetra_Map that was passed
to the matrix constructor).

Alan
 

> -----Original Message-----
> From: trilinos-users-bounces at software.sandia.gov 
> [mailto:trilinos-users-bounces at software.sandia.gov] On Behalf 
> Of gjl7 at et.byu.edu
> Sent: Monday, May 15, 2006 4:08 PM
> To: Marzio Sala
> Cc: trilinos-users at software.sandia.gov
> Subject: Re: [Trilinos-Users] Epetra Matrices for ML from 
> MetisPartitioned Matrices
> 
> As for using Epetra_RowMatrix derived objects, since the the 
> matrices are already set up in a compressed sparse row 
> format, it is easy to create and use Epetra_CrsMatrix objects.
> 
> However, after looking at the online development 
> documentation of the Epetra_FECrsMatrix, I'm not seeing how 
> to "unify" the rows corresponging to boundary nodes.  Is that 
> the correct location and I'm just not seeing it or is there a 
> better place to look?
> 
> Greg
> 
> 
> > Hi Greg,
> >
> > the ML preconditioning class works with Epetra_RowMatrix 
> derived objects.
> > It
> > requires that each row is *entirely* owned by exactly one 
> processor. 
> > If you can reproduce this situation, it is quite easy to wrap your 
> > distributed matrix format as an Epetra_RowMatrix, then call 
> ML. The ML 
> > user's guide describes how local and global indices are 
> managed by ML.
> >
> > If you partition the grid elements, then probably you first need to 
> > "unify"
> > the rows corresponding to boundary nodes, then call ML. 
> Classes like 
> > Epetra_FECrsMatrix can make this step easy.
> >
> > -Marzio
> >
> >
> > On 5/10/06, gjl7 at et.byu.edu <gjl7 at et.byu.edu> wrote:
> >>
> >> We are linking Trilinos with another code.  The finite 
> element mesh 
> >> has already been partitioned using metis and the local 
> matrices are 
> >> already set up in a compressed sparse row format and 
> distributed on 
> >> each processor.  Now, we want to use ML to solve these 
> matrices, with 
> >> ML realizing these matrices are coupled due to node overlap.  Is 
> >> there an example showing how to set up the Epetra matrices so ML 
> >> realizes these matrices are linked by the overlapping nodes?  If 
> >> there is no documented example, what would you recommend?  Thanks.
> >>
> >> Greg Larson
> >>
> >>
> >>
> >> _______________________________________________
> >> Trilinos-Users mailing list
> >> Trilinos-Users at software.sandia.gov
> >> http://software.sandia.gov/mailman/listinfo/trilinos-users
> >>
> >>
> >
> 
> 
> 
> 
> _______________________________________________
> Trilinos-Users mailing list
> Trilinos-Users at software.sandia.gov
> http://software.sandia.gov/mailman/listinfo/trilinos-users
>