KSPCR with PCILU generally slower with v2.3.3 than with 2.3.0

[Etienne PERCHAT]

Hi Barry,

Using -ksp_truemonitor does not change anything. I have the same number of iterations in the 3 cases (v2.3.3p8 + true_monitor, v2.3.3, v2.3.0) and the same final residual norm but with v2.3.0 the code is slightly faster.

I have to stress that I also monitor convergence with KSPMonitorSet in order to store the "best residual" and the "best solution" (I did not test if it is still useful in 2.3.3). 

Etienne

-----Message d'origine-----
De : owner-petsc-users at mcs.anl.gov [mailto:owner-petsc-users at mcs.anl.gov] De la part de Barry Smith
Envoyé : jeudi 5 juin 2008 20:28
À : petsc-users at mcs.anl.gov
Objet : Re: KSPCR with PCILU generally slower with v2.3.3 than with 2.3.0


    This could also be the result of changes in the way we test for  
convergence.
Try with -ksp_truemonitor and see if it is using the same number of  
iterations
to converge.

    Barry

On Jun 5, 2008, at 5:35 AM, Etienne PERCHAT wrote:

> Hello,
>
> I'm solving MPIBAIJI matrices using KSPCR preconditioned by a  
> PCBJACOBI
> between sub domains and PCILU with a fill ratio of 1 on each sub  
> domain.
>
> Solving exactly the same linear systems, I notice that the solution is
> slower using v2.3.3p8 than with v2.3.0.
> The behavior is not very clear: v2.3.3p8 is generally slower (10 % or
> more) but on some systems it happen to be really faster (may be 40 %).
>
> Unhappily in my case at the end I'm always slower ...
>
> Did somebody noticed such behavior? Is there changes if I upgrade to
> current 2.3.3p13 version (I've seen a fix:" fix error with using
> mpirowbs & icc & ksp_view") ?
>
> Thanks,
> Etienne Perchat
>
>