[MITgcm-devel] pkg/seaice pseudo time stepping for LSOR

[Patrick Heimbach]

Hi Martin,

I could increase to M=100, but would be huge.
I'll change so as to put M(.NE.N) only for adjoint.
Two things:
* could we try to put pseudo loop inside seaice_lsr?
   It may (but not sure yet) give more options to trick adjoint.
* I just forgot 2nd point. Maybe it will come back...

-p.

On May 27, 2009, at 4:04 AM, Martin Losch wrote:

> Hi Patrick,
>
> I didn't see, that you already fixed it with M=2. That's perfectly  
> OK for me. Thanks,
>
> Martin
>
> On May 27, 2009, at 9:10 AM, Martin Losch wrote:
>
>> Hi Patrick,
>>
>> I was afraid of that. Unfortunately it's hard to say what  
>> iteration count is to be expected, basically I want to play around  
>> with this. In the recent paper by Lemieux and Tremblay (http:// 
>> www.agu.org/pubs/crossref/2009/2008JC005017.shtml), they find that  
>> their model reaches convergence only after 10,500 iterations. I  
>> think that that number is far too much for the generally small  
>> time steps that we are usually using nowadays, but I cannot claim  
>> that 100 or 1000 are better guesses.
>>
>> Can you set a hard limit at say 100, that is only active for  
>> adjoint computations? The (absolutely feasible) alternative is  
>> that I just keep my local copy of this parameter. Go for 100.
>>
>> Martin
>>
>> On May 27, 2009, at 12:02 AM, Patrick Heimbach wrote:
>>
>>>
>>> Hi Martin,
>>>
>>> what are max. expected number of pseudo-timesteps.
>>> As you probably guess, your changes are producing some
>>> relatively ugly recomp. of seaice_lsr within adseaice_dynsolver.
>>> For the storing we'll need a hard limit (to be used as
>>> common block size).
>>> What's a good number?
>>>
>>> Cheers
>>> -p.
>>>
>>> On May 25, 2009, at 5:50 AM, Martin Losch wrote:
>>>
>>>> Hi Patrick,
>>>> I have checked in code to do pseudo time stepping for LSOR.  
>>>> Basically, the two calls to seaice_lsr in seaice_dynsolver have  
>>>> been replaced by a loop with a default of 2 iterations, and the  
>>>> copy/interpolation between time levels has been moved to  
>>>> seaice_lsr.
>>>>
>>>> A potential problem for the adjoint are the store directives for  
>>>> uicec and vicec in the 'predictor time step'. I have moved them  
>>>> to seaice_lsr along with the rest of the code and I have  
>>>> deliberately left out the "(icall-1)*ncklev_1 part of the key  
>>>> that is used for uice/vice, because the storing is only applied  
>>>> for icall==1. I am not sure if this will work properly and  
>>>> apologize beforehand for all problems that this may cause.
>>>>
>>>> Martin
>>>>
>>>> _______________________________________________
>>>> MITgcm-devel mailing list
>>>> MITgcm-devel at mitgcm.org
>>>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
>>>
>>> ---
>>> Patrick Heimbach | heimbach at mit.edu | http://www.mit.edu/~heimbach
>>> MIT | EAPS 54-1518 | 77 Massachusetts Ave | Cambridge MA 02139 USA
>>> FON +1-617-253-5259 | FAX +1-617-253-4464 | SKYPE patrick.heimbach
>>>
>>>
>>> _______________________________________________
>>> MITgcm-devel mailing list
>>> MITgcm-devel at mitgcm.org
>>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
>>
>> _______________________________________________
>> MITgcm-devel mailing list
>> MITgcm-devel at mitgcm.org
>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
>
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel

---
Patrick Heimbach | heimbach at mit.edu | http://www.mit.edu/~heimbach
MIT | EAPS 54-1518 | 77 Massachusetts Ave | Cambridge MA 02139 USA
FON +1-617-253-5259 | FAX +1-617-253-4464 | SKYPE patrick.heimbach