[dune-pdelab] Accumulation mode for Nonoverlapping AMG

Rebecca de Cuveland rebecca.cuveland at iwr.uni-heidelberg.de
Mon Sep 8 09:55:38 CEST 2014


Hey Marcus,
it's a long time since I last worked on nonoverlapping AMG...
When I finished working on nonoverlapping AMG it was only possible to 
use atOnceAccu or noAccu.
(This is not too well documented, I know, in novlpistlsolverbackend.hh 
there is this one line: //use noAccu or atOnceAccu)

The problem had something to do with the fact, that the owner process of 
a vertex does not always know all neighbours of this vertex.
Therefore in matrixredistribute.hh there exists a function 
completeSparsityPattern that adds missing entries for the nonoverlapping 
case.
Somehow this was problematic for the successiveAccu mode, but I cannot 
remember the exact problem right now, sorry :-(.
At this time atOnceAccu mode was enough for everyone so I did not work 
further on it, but I think one would have to extend 
matrixredistribute.hh from istl to use successiveAccu with 
nonoverlapping AMG.

If you want I can check tomorrow if I can find more in my notes...
Bye,
Rebecca

-- 
Rebecca de Cuveland
Universität Heidelberg
Interdisziplinäres Zentrum für Wissenschaftliches Rechnen
Im Neuenheimer Feld 368, 69120 Heidelberg
Tel: 06221/548890   Fax: 06221/548884
e-mail: <rebecca.cuveland at iwr.uni-heidelberg.de>

On 04.09.2014 11:00, Markus Blatt wrote:
> Hi,
>
>
> is there a special reason (e.g. a bug), why the default accumulation
> mode for the nonoverlapping AMG ist atOnceAccu (if accumulation
> happens, everything is gathered on one process)? For the overlapping
> case we allow the numbers of processes on the coarse level to decrease
> to a smaller number of processes (>=1) and do this successfully.
>
> Markus
>
>
> _______________________________________________
> dune-pdelab mailing list
> dune-pdelab at dune-project.org
> http://lists.dune-project.org/mailman/listinfo/dune-pdelab

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.dune-project.org/pipermail/dune-pdelab/attachments/20140908/dc798a7a/attachment.htm>


More information about the dune-pdelab mailing list