[MITgcm-devel] stripes in adjoint along tile edges

[Martin Losch]

Hi Dimitris,

thanks, I tried this and it does not seem have any noticable effect. The (adjoint) solution is a little smoother for the very low (1e-20) values, but the striping does not really change between residuals of 1e-8 or 1e-18. I also tried to turn off the cg2d solver by using explicit schemes (implicSurfPress = 0, implicDiv2DFlow = 1) with very short time steps. There the stripes **almost** disappear. That’s my only piece of evidence that something is not right with cg2d, but maybe it’s also somewhere in some MPI-related routines.

Martin

On Nov 27, 2014, at 3:54 PM, Menemenlis, Dimitris (329D) <Dimitris.Menemenlis at jpl.nasa.gov> wrote:

> Not sure if relevant, but I do remember that Ralf Giering was sufficiently concerned about cg2d inaccuracies to write an alternative "direct" solver during his time at JPL.
> 
> One way to test this hypothesis would be to vary (lower or raise) the accuracy (number of iterations) of the cg2d solver and observe whether striping decreases or increases.
> 
>> On Nov 27, 2014, at 6:37 AM, Martin Losch <Martin.Losch at awi.de> wrote:
>> 
>> Hi there,
>> 
>> since there hasn’t been any other statement: Can I assume that no-one on this list has ever seen stripes like this in the adjoint (would be good, because it would mean, that I am making a mistake, that I should be able to find)? Or does the low reply rate mean that no-one has ever looked closely enough to see the stripes? 
>> 
>> I am running a little out of ideas about where they might come from. I have a suspicion, that they are related to cg2d/solve_for_pressure, but no good pieces of evidence, yet.
>> 
>> Martin
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel