[petsc-users] Convergence/accuracy degradation with increasing number of procs

[Alexander Grayver]

Thanks Matt, that was my guess actually.

On 09.01.2012 17:46, Matthew Knepley wrote:
> On Mon, Jan 9, 2012 at 10:37 AM, Alexander Grayver 
> <agrayver at gfz-potsdam.de <mailto:agrayver at gfz-potsdam.de>> wrote:
>
>     I see this is due to preconditioner, but could you explain shortly
>     what particularly causes this problem? It's not obvious to me.
>
>
> The size of the block being factorized by ILU(0) decreases as p 
> increases. Thus, it is
> a weaker preconditioner. I recommend Yousef Saad's book which explains 
> all this clearly.
>
>    Matt
>
>     Thanks.
>
>     On 09.01.2012 17:36, Matthew Knepley wrote:
>>     On Mon, Jan 9, 2012 at 10:31 AM, Alexander Grayver
>>     <agrayver at gfz-potsdam.de <mailto:agrayver at gfz-potsdam.de>> wrote:
>>
>>         I have tested default GMRES+ILU(0) solver for problem of size
>>         ~200000 and I observed that both convergence rate and
>>         accuracy degrade when I increase number of processes.
>>         E.g., number of iterations almost doubles when going from 1
>>         to 4 processors.
>>
>>         I noticed that ILU(0) is somehow represented by sequential
>>         matrices internally and feel that this might be a reason?
>>
>>
>>     The default is Block Jacobi+ILU(0) and GMRES(30) which will
>>     definitely degrade as the number of
>>     processes is increased. You would need to use an optimal
>>     preconditioner like Multigrid if you want
>>     a constant number of iterates.
>>
>>        Matt
>>
>>         Regards,
>>         Alexander
>>
>>
>>
>>
>>     -- 
>>     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
>
>
>
>
> -- 
> 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

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120109/60f65485/attachment.htm>