[MITgcm-devel] changing exch-1

[Menemenlis, Dimitris (3248)]

Jean-Michel, we already discussed this with Matt yesterday
(off the list to avoid multi-mail chatter).
Hong and I will look into this some more and get back.
I attach yesterday's discussion.  Dimitris

On May 4, 2010, at 9:17 PM, Jean-Michel Campin wrote:

> Dimitris,
> OK, I understand. Matt has an other suggestion, but I am not
> too familiar with how things evolve regarding the sensitivity,
> so let you discuss this with him.
> Jean-Michel

> From: Matthew Mazloff <mmazloff at ucsd.edu>
> Date: May 4, 2010 8:38:09 PM PDT
> 
> My issue was rather benign too...basically the sensitivities were zero  
> on either the x or the y tile edges..so it didn't hurt the  
> optimization...but looked funny
> 
> also, oddly, i think it didnt really show up in adxx_uwind....only in  
> ADJustress...so i too let the issue slide
> 
> anyway, not sure if that issue still exists...maybe Patrick knows
> 
> please let me know how it turns out...
> 
> -matt
> 
> On May 4, 2010, at 8:24 PM, Menemenlis, Dimitris (3248) wrote:
> 
>> I will talk with Hong tomorrow and let you know how it goes.
>> For some reason this tiling does not appear to degrade the solution
>> (or to get worse with time) so we haven't taken as close a look as
>> we should.  D.
>> 
>> On May 4, 2010, at 8:17 PM, Matthew Mazloff wrote:
>> 
>>> Hi Dimitris,
>>> 
>>> I have seen tiling in wind.  I think it showed in adj(u,v)stress  
>>> and I
>>> thought it may be an A vs. C-grid thing with the exchanges....but not
>>> sure...don't really remember
>>> 
>>> anyway, it showed up right away...very obvious after 5 days....so I
>>> thought maybe you could diagnose it in just a few days of run time
>>> with an adjoint state dump....at least maybe in the derivatives
>>> 
>>> However I have never seen tiling in atemp and I don't use the CS so
>>> maybe this is a completely different problem....
>>> 
>>> -Matt
>> 
>>> 
>>> On May 4, 2010, at 8:06 PM, Menemenlis, Dimitris (3248) wrote:
>>> 
>>>> Hong could you take a look at some of the earlier adjoint sensitivity
>>>> fields to see what is shortest integration period that shows the  
>>>> tiling.
>>>> 
>>>> Matt, I haven't had time to look at this issue in detail.
>>>> It's kind of subtle and not readily visible except in certain  
>>>> adjoint fields.
>>>> For example, we (sort of) see the tiling in uwind, vwind, and atemp,
>>>> but not in precip, shortwave, and humidity.
>>>> 
>>>> D.
>>>> 
>>>> On May 4, 2010, at 7:58 PM, Matthew Mazloff wrote:
>>>> 
>>>>> Dimitris, can't you just run a few days backward to see the issue?