On Wed, Apr 1, 2009 at 9:46 AM, Angus Hendrick <[email protected]> wrote:
> I have also found that Trilinos is about 4x faster for a 200x50 domain.
> Also, in a for some problems where the standard solver gives bad results,
> the Trilinos solvers give accurate results.  I believe these are "stiff"
> problems, but I haven't actually looked at the matrix.

We've also had the same issues with pysparse, but haven't fixed them
or categorized them yet. In fact I have been experimenting with a pure
python implementation of the krylov solvers using this branch

   <https://matforge.org/fipy/browser/branches/pykrylov>

and these solvers

   <http://github.com/dpo/pykrylov/tree/master>

This may become the default system if it isn't too slow as it would
make installation way easier.

> While it doesn't help for solving large domains, you can gain speed
> increases on multi-core machines for embarrassingly parallel problems (i.e,
> multiple separate cases) using IPython.

I am currently working on a full parallel version of fipy but for
grids only. Essentially the grid is divided up in a trivial way with
guard cells and separate fipy jobs run on each of the subgrids. A
special class assembles the global matrix. It should work very well
with good scaling. Unstructured grids will be difficult and we
probably won't get into that in the near future. I'll keep you posted
regarding this and when we have a branch (currently it isn't even on a
branch of fipy, just in a working copy) and something committed.

-- 
Daniel Wheeler

Reply via email to