Increasing convergence rate

[Matthew Knepley]

On Thu, Jan 22, 2009 at 11:59 PM, jerome ho <jerome.snho at gmail.com> wrote:
> On Wed, Jan 21, 2009 at 3:14 PM, Jed Brown <jed at 59a2.org> wrote:
>> Most preconditioners are not the same in parallel, including these
>> implementations of AMG.  At a minimum, the smoother is using a block
>> Jacobi version of SOR or ILU.  As you add processes beyond 2, the
>> increase in iteration count is usually very minor.
>>
>> If you are using multiple cores, the per-core floating point
>> performance will also be worse due to the memory bandwidth bottleneck.
>>  That may contribute to the poor parallel performance you are seeing.
>>
>
> Hi
>
> I'm getting strange results. In parallel (on 2 processors), the result
> doesn't to be able to converge further but appears to fluctuate
> between 1e-9 and 1e-8 (after 100+ iterations), when it solves in 8
> iterations on a single machine. I decrease the rtol (from 1e-7) for
> the parallel simulation because I'm getting a 20% result difference.

The default KSP is GMRES(30). Since the parallel preconditioner is weaker,
you might not be able to solve it in 30 iterates. Then this solver can
fluctuate every time you restart. You can try increasing the number of
Krylov vectors kept.

   Matt

> When I split into more (6) processors, it's reporting divergence. Am I
> doing something wrong here? Should I be switching to DMMG method
> instead? The matrix size is about 1mil x 1mil.
>
> Jerome
-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which
their experiments lead.
-- Norbert Wiener