[MITgcm-devel] changing exch-1

[Jean-Michel Campin]

Hi Dimitris,

I think it would be better to re-do one iteration you already done
(onless you don't have time to do it) so that we can be sure. 
I think it's worth to try, since you have this tiling problem. 
And no need to do more than 1 iteration. 

I only did very short test (1 time step, using lab_sea/input_ad.noseaice
and switching on EXCH2), but I was seeing differences at tile edges
in the forcing sensitivity (which make sense since it was in exf),
same as you listed, and not in Precip.
But don't know how this propagate.

Cheers,
Jean-Michel

PS: there was also the AD global sum issue (which relates to 
MPI domain decomposition) which has been fixed, but no idea if this
could be the cause of tile-edge discontinuity.

On Tue, May 04, 2010 at 04:16:26PM -0700, Menemenlis, Dimitris (3248) wrote:
> Jean-Michel, what kind of test do we need to run and what do you expect the difference to be?
> 
> Would checking out updated MITgcm code and running one more iteration of CS510 adjoint suffice?
> 
> As you may remember, one issue we have with CS510 adjoint solution is the presence of
> tile boundaries in the adjoint sensitivities and adjustment fields for uwind, vwind, and atemp?
> Could your AD-exch problem be the cause of this tile-edge issue?
> 
> Dimitris
> 
> On May 4, 2010, at 2:07 PM, Jean-Michel Campin wrote:
> 
> > And finally, the recent problems I found in AD-exch (one did not
> > show up in my tests, but the other with exch2 & exf is likely
> > to affect the CS-510 AD runs: Dimitris, can you check ?)
> > tend to indicate that a more transparent coding could help;
> > And since there are still few specific exch (B-grid, D-grid,
> > SM ) which need an AD version, it could be a good time to 
> > clean-up the AD-exch1 stuff (before adding the missing ones).
> 
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel