[MITgcm-devel] verification experiment seaice_obcs

[Dimitris Menemenlis]

> thanks for reminding me. I noticed now that the IO is all single  
> precision, could that be a reason for the sensitivity between  
> platforms?

Good question.  This is also the case for the lab_sea experiment,  
whose solution is also very sensitive between platforms.
Does anyone on the devel list have experience with cross-platform  
single-precision I/O sensitivity?  Is it worth changing I/O to double- 
precision in the lab_sea and seaice_obcs experiments?

> I remember that there was a restart problem with seaice, but as far  
> as I remember it is related to EVP:  we need to store sigma12, but  
> sigma 12 is defined on Z-points, so we'd need the field (1:nx+1,1:ny 
> +1), but there's no simple way of creating a pickup file like this  
> (except via netcdf). For the LSR solver there was no problem and  
> there isn't any problem in lab_sea with EVP, because we fixed the  
> topography, see here:
> <http://dev.mitgcm.org/cgi-bin/viewcvs.cgi/MITgcm/verification/lab_sea/input.hb87/?only_with_tag=MAIN 
> >
> with this text
> "lat-long grid & no-slip BC in seaice-dynamics produce different  
> sigma12 at j=1 & j=Ny+1 : this case is not treated correctly in  
> restart process. Fixed here by closing the domain at the Northern  
> boundary."

Interestingly, seaice_obcs passes tst_2+2 if land is added at  
SouthWest edge but problem is probably different from what you  
describe above since seaice_obcs uses LSR not EVP.

D.