[Trilinos-Users] Best direct solvers

Heroux, Michael A maherou at sandia.gov
Fri Oct 16 14:52:21 MDT 2009


Tridiagonal systems are special in that there is no fill-in.  Although using Epetra to remap the data is OK, if I understand the situation, KLU would not be most efficient to solve the resulting independent tridiagonal systems. You would be better off with a specialized algorithm, e.g., the so-called Thomas algorithm.  If nothing else is available, LAPACK has tridiagonal solvers that are reasonable.

Mike

On 10/16/09 3:08 PM, "Erik Boman" <egboman at sandia.gov> wrote:

Hi John,

If I understand correctly, you have lots of tridiagonal systems
corresponding to the x direction that can all be solved independently.
In this case, I would suggest to remap your data (Epetra import/export)
such that you can use a serial direct solver (KLU) on each tridiagonal
system. You could perhaps view this as a domain decomposition approach
and use Ifpack with Amesos/KLU as the subdomain solver.

An alternative is to form one huge global (distributed) matrix and call
Amesos with a 3rd party solver (SuperLU, MUMPS, etc.) but I think that
would be less efficient.

Erik


On Fri, 2009-10-16 at 13:08 -0600, John R. Cary wrote:
> Hi Eric,
>
> In this regard, suppose my matrix corresponded to differences in only one
> direction (e.g., only x, not y, z).  On each line, I could do a direct
> tridiagonal
> solve if I were in serial.  What would be best for parallel in this case?
>
> Thanks...John
>
> Erik Boman wrote:
> > Jatin,
> >
> > Why do you need the inverse? If you wish to solve Ax=b, Trilinos has
> > many solvers, both iterative (Ifpack, ML) and direct (Amesos).
> >
> > Regards, Erik
> >
> > On Fri, 2009-10-16 at 10:08 -0600, jatin ralhan wrote:
> >
> >> Hi,
> >>
> >> Does any one know whether it's possible to get the inverse of
> >> Epetra_CrsMatrix (Matrix is Sparse Symmetric with filled diagonal
> >> elements).
> >>
> >> Thanks in advance for your help.
> >>
> >> Jatin
> >>
> >>
> >>
> >> ______________________________________________________________________
> >> Add whatever you love to the Yahoo! India homepage. Try now!
> >>
> >
> > _______________________________________________
> > 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

-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://software.sandia.gov/pipermail/trilinos-users/attachments/20091016/2e118091/attachment.html 


More information about the Trilinos-Users mailing list