[MITgcm-devel] Re: [MITgcm-support] optimization problem with obcs control

[Martin Losch]

Hi Suneet,

so far I have never done a multiprocessor optimization with open  
boundaries, so I do not have the code that need. I am afraid, that  
you'll have to do the changes yourself. Good luck!

Martin

On 16 May 2008, at 05:55, suneet dwivedi wrote:

> Hi Martin,
> Thankyou very much for your reply and for sending me the lsopt and  
> optim directories. I looked at them carefully and tried to  
> incorporate your changes in my model setup. However, the problem  
> was somewhere else. Yesterday I figured it out and now I am able to  
> successfully run the optimization/assimilation experiment (with  
> obcs package included).
>
> You may be probably aware of that in one of the recent emails that  
> I sent to the mitgcm help, I mentioned that the "use  
> singlecpuio=true" does not work for obcs control vectors with  
> multiprocessor runs i.e. the obcs control files are written/ 
> scattered on all the cluster nodes even with singlecpuio=true for  
> mitgcm mpi runs. In reply to my email concerning this problem, Matt  
> sent me few mdsio slice routines, which actually do this job for  
> mpi runs. The problem was in one of those files where "nPy" value  
> was hardcoded equal to 1. The only change that I made in the file  
> is to generalize it in such a way so as to work for any nPy rather  
> than a fixed value 1. Now my optimization routines are working fine.
>
> Another problem which I noticed with those slice routines is that  
> they are written to work for northern boundary only so that as soon  
> as I try to ALLOW S/E/W control, I get the same error as I was  
> getting earlier. Matt told me that no one has yet implemented the  
> singlecpuio changes for the other boundaries.
>
> May I request you to kindly send me any such mdsio slice routines  
> that you may be having which worked for you while using all the  
> open boundaries when running the model on multi processors,  
> otherwise, I will have to myself change those routines to make them  
> work for other boundaries (which will be quite difficult, I  
> guess!!!!).
> Hoping for your reply,
> Thanks again and with best regards,
> Suneet
>
>
> On Wed, 14 May 2008 Martin Losch wrote :
> >Suneet,
> >
> >off the list, because I am sending you 2MB file with this email.   
> Please find attached an old version of the lsopt and optim   
> directories, that did work for me with obcs; the directories  
> include  the compiled object files etc, but you can probably delete  
> them. Many  changes  have been made to the code since then, but  
> maybe the  modifications that I made back then to make it work will  
> help you  (find out by cvs diff). Unfortunately I cannot remember,  
> what the  modification where and if they were severe, but honestly,  
> I don't  think so. As far as I remember, I mostly had to struggle  
> with format  issues (big endian/little endian, etc.). I had though  
> that I had sent  you that stuff already, but I guess I didn't.
> >
> >Martin
> >
> >
> >On 13 May 2008, at 21:59, Suneet Dwivedi wrote:
> >
> >>Hi everyone,
> >>Can anybody of you please tell me why my optim.x run doesn't read
> >>"nWetobcsnGlo" even though I defined obcs N control? How to resolve
> >>this problem? When I went through optim_readdata.F, I found that  
> when
> >>OBCS N CONTROL is allowed, I have to add OBCS mask information into
> >>the header section for optimization. What does it mean? I guess obcs
> >>mask informations are already defined in ctrl.h via ctrl_pack/unpack
> >>routines. In addition to this I also tried to include OBCS.h in the
> >>relevant *.F files in optim folder, but nothing of these are  
> working.
> >>Everytime I try to run the optim.x , I get the following error
> >>message:
> >>   
> ---------------------------------------------------------------------- 
>  ----------------------------------
> >>forrtl: severe (67): input statement requires too much data, unit  
> 20,
> >>file /tmp/mitgcm_so_suneet/JHU_ctrl _MIT_CE_000.opt0000
> >>-------------------------------------------------------------------- 
> -- ----------------------------------
> >>It is pertinent to mention here that optimization/assimilation run
> >>successfully completes when obcs is not defined. Is this a problem
> >>with mitgcm assimilation experiment when obcs controls are defined?
> >>Please help me sort out this problem.
> >>Hoping for reply,
> >>Cheers,
> >>Suneet
> >>_______________________________________________
> >>MITgcm-support mailing list
> >>MITgcm-support at mitgcm.org
> >>http://mitgcm.org/mailman/listinfo/mitgcm-support
> >
>
>
>