[Trilinos-Users] Amesos and fill-reducing orderings

[Erik Boman]

Hi Jonas.

1) Most direct solvers will do some type of reordering by default, but
this is usually invisible to the user. Typically, you can ignore the
ordering issue and get decent performance, but in some cases, ordering
can make a big difference and you may want to optimize the ordering
yourself.

2) Isorropia can indeed compute a permutation vector to reorder a sparse
matrix. This is currently supported by the Isorropia::Epetra::Orderer
class, see
http://trilinos.sandia.gov/packages/docs/r10.0/packages/isorropia/doc/html/classIsorropia_1_1Epetra_1_1Orderer.html

Right now, Isorropia relies on third party libraries (Metis, ParMetis,
Scotch) but native ordering schemes are under development and will be
released soon.

3) There is currently no standard way in Amesos to provide row or column
permutations (ordering) to external (TPL) solvers. This is a known
deficiency, but Amesos has not been actively developed for quite a
while. My suggestion for now, is to use Isorropia to produce a
permutation vector and then give that directly to the TPL solver as a
parameter (note: only some solvers support this feature).

Will you work in serial or in parallel? Isorropia works in both cases,
but please note that in parallel Isorropia will produce a distributed
permutation, while many solvers require the permutation vector to be on
a single processor (centralized).

Regards,
Erik


On Wed, 2009-12-02 at 07:36 -0700, Jonas Thies wrote:
> Hi,
> 
> I'm wondering about the recommended way to solve tricky linear systems
> with Amesos. Mine work pretty well with the MUMPS interface because
> MUMPS uses a reordering strategy like metis/pord etc, whereas seemingly
> the other solvers don't (?). Should I compute a reordering with some
> other package first (isorropia maybe?), what's the easiest interface to
> use? I want to put a master student to work on something, but he doesn't
> have Trilinos (or even C++) experience, so something not involving Thyra
> and the like would be nice.
> 
> Thanks a lot,
> Jonas
> 
> _______________________________________________
> Trilinos-Users mailing list
> Trilinos-Users at software.sandia.gov
> http://software.sandia.gov/mailman/listinfo/trilinos-users