[MITgcm-devel] more jfnk

[Jean-Michel Campin]

Hi Martin,

This sounds good to me. 
When this is implemented in LSR then we probably want to 
have an option to be able to reproduce old results.

Cheers,
Jean-Michel

On Tue, Feb 12, 2013 at 04:52:44PM +0100, Martin Losch wrote:
> Hi Jean-Michel,
> 
> I have implemented an improved (?) version of the LSR-preconditioner. Actually we can even think of improving seaice_lsr in the same manner: We run LSR on 1-Olx+2:sNx+Olx-2, 1-Oly+2:sNy+Oly-2 and throw away the solution on the overlaps during the exchange. This would be an additive Schwarz method (lot of literature about that) and it seem to improve convergence in parallel runs. I can probably even get rid of the extra unclean exchange in seaice_preconditioner.F
> 
> For now I would like to keep the reference experiment, but change the default to using the additive Schwarz method together with only one exchange in seaice_preconditioner.F. This would mean that I'll have to modify offline_exf_seaice/input.dyn_jfnk/data.seaice and also set a CPP FLAG in SEAICE_OPTIONS.h
> Can you agree with this, or would you rather keep the default, and change it later, once the method has been "proven" to be better in many setups?
> 
> Martin
> 
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel