[petsc-dev] features of GAMG

[Mark F. Adams]

On Dec 25, 2011, at 3:07 PM, Jed Brown wrote:

> On Sun, Dec 25, 2011 at 14:02, Mark F. Adams <mark.adams at columbia.edu> wrote:
> I've looked at ksp/ksp/examples/tutorial/ex56 and the modification that I make to emin and emax seem to be essential and very useful, respectively.  But I am seeing something funny with Cheb.
> 
> Are you comparing to "-mg_levels_ksp_chebychev_estimate_eigenvalues 0,0.1,0,1.1" or do you mean relative to some naive Cheby configuration?

I set the KSP to cheby as the default, then do a KSPGetOptions, and then if its (still) cheby I set the min and max eigenvalues manually.

>  
> 
> Jed:  I'm seeing very little difference in convergence rates on ex56 when I use a first, second or fourth order Cheb smoother.  I've seen some cases where the final residual was slightly larger with a higher order smoother.  I think there is something wrong here ...
> 
> Can you show me an example? I have certainly seen it make a different for other problems.

OK, well we need to think about this then.  Maybe we can take a look at it at Brown in January.  There must be some bad interaction in the way that I'm using it.  The code is around 662 in gamg.c.

>  
> 
> Also, I thought we had added something like -pc_mg_smooth to set the number of smoothing steps when you are using the same smoother for the up and down smoother.  I can not see a command for this.  I did '-help | grep smooth', and I tried a few things.  So to test this you need to manually fix the smoothing steps at about line 662 in gamg.c.
> 
> We would normally use something like -mg_levels_ksp_max_it 3. Or are you talking about applying first-order Cheby multiple times?


Damn, I knew this, I forgot.

Thanks,
Mark



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20111226/171427a6/attachment.html>