[MITgcm-devel] seaice code beyond checkoint61j blows for ECCO-GODAE

[Martin Losch]

Hong,

I can't tell you how sorry I am about the trouble. I was so confident  
that the new discretization was so much better (and nice in terms of  
written code). And I still think that we should move away from the old  
discretization as quickly as possilble, so that I would appreciate any  
information that leads to debugging the new one. How fast does it blow  
up? does it start somewhere locally? Problems with CS-grid may have a  
totally different origin that problems with lat/lon-grids.

For a test, could you also try to run CS-grids with the new code and  
SEAICEuseMetricTerms=.false.  It's very well possible, that I made a  
mistake there.

For CS-grids, the weird flag SEAICE_OLD_AND_BAD_DISCRETIZATION should  
recover the old results (I wonder if I should have named it  
SEAICE_OLD_AND_STABLE_DISCRETIZATION). It does so for  
global_ocean.cs32x15/input.icedyn. Remember that with this  
discretization, the metric terms are turned off of CS-grids.
However, there was a bug in the no-slip boundary conditions (off by  
default) that I corrected, so if you use SEAICE_noslip=.true., you'll  
not be able to recover the old results.

For lat/lon grids, the old discretization (checkpoint61j) was really,  
really wrong (metric terms). I tried to fix that for checkpoint61k,  
but I have the same problems with 61k as I had with 61j: the solver  
occasionally does no converge and the model blows up as a consequence;  
there is probably still a problem in the symmetry of the discretized  
operator under certain circumstance. The fact that the ECCO-GODAE runs  
did not blow up with the old code can only be a (un-)fortunate  
coincidence (sometimes it took my runs a couple of decades that have  
the problem).
For our three different lat/lon configurations (2deg, half degree,  
quarter degree), the problems only went away with the new finite- 
volume discretization.

Please send me more details about the problems (and I won't be able to  
look at that closely until after the EGU, ie Apr25).

Martin

On Apr 14, 2009, at 6:04 PM, Hong Zhang wrote:

> On Tue, 2009-04-14 at 03:49 -0700, Martin Losch wrote:
>>> I do not
>>> know if Dimitris has gotten around to test this on CS-or curvilinear
>>> grids.
>>>
>>
> Hi
> We also got the problem on CS:
> got all seaice fields NaN both on levl 28 and level 50 config.
> We updated the code on April 7
> and want to fix the potential blowup (sensitivity) on CS-corner
> with Patrick's check-in (see :
>> Modified Files:
>>     mom_calc_visc.F
>> Log Message:
>> Overlaps had been forgotten in calculating ijk keys
>> (spotted by jmc using gfortran with check-bounds)
> Have to revert?
>
> cheers,
> hong
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel