another getdp problem

Christophe Geuzaine Christophe.Geuzaine at ulg.ac.be
Mon Sep 25 13:34:15 CEST 2000


Samuel Kvasnica wrote:
> 
> Hi Christophe,
> 
> I thought I wouldn't bother you any more with that magnetostatics, but I've encountered some
> strange behaviour.
> In the problem I've sent you last time, I made some minor changes to geometry - I just divided one
> surface, so new one has been created. Well, everything works ok with basic basis functions but if
> I add the second coeficient for HBF functions, it converges suddenly very slowly, after ~35000
> iterations.
> Before it took ~10000. Is this normal ? Nodes and elements count is not very different (~1%).
> Could it be related someway to solver parameters ?
> 

Ow, 10000 iterations? I usually apply a preconditionner to converge with
100-200 its. So I think you should use a stronger preconditionner (i.e.
a better approximation of the inverse of the matrix). For example, you
can select (in the SOLVER.PAR file) an ILUTP preconditioner
(Preconditioner 2) allow more non-zero elements in the ILU (Nb_Fill
100).

> best regards,
> 
> Sam
> 
> P.S. for processing of computed data (smoothing, visualization etc..) I finally
> 
> decided to use OpenDX, that's really great and free !
> 

Is it from IBM?

-- 
Christophe Geuzaine

Tel: +32-(0)4-366.37.10    mailto:Christophe.Geuzaine at ulg.ac.be
Fax: +32-(0)4-366.29.10    http://www.geuz.org