[Mitgcm-support] Note: January 11, 2001

mitgcm-support at dev.mitgcm.org mitgcm-support at dev.mitgcm.org
Wed Jul 9 15:56:41 EDT 2003


HyperNews notification of new message.  See: 
  http://escher.JPL.NASA.GOV:2000/HyperNews/get/forums/wkmtg/78.html

     =======================================================================     
     
NOTES FROM THE ASSIMILATION MEETING : 11 January 2001

1) K-filter
------------
   - The assimilation code that was lost in December has been
     reconstructed and deposited to CVS; see posting[1].  Furthermore,
     the baroclinic and barotropic codes have been combined to perform
     the assimilation simultaneously.

   - Two baroclinic filter runs that were lost in December have been
     recovered.  Two more barotropic filter runs are being carried
     out.

   - In the process of recovering the code, we discovered that
     previous calculations were using outdated topex data files (dated
     5/11/00 instead of 7/20/00) that had a bug at some data
     locations.

   - It was also discovered that the previous K-filter runs were using
     suboptimal fortran optimization.   Optimization by

  set FOPTIM     = ( '-O3 -OPT:Olimit=0:roundoff=3:div_split=ON:alias=typed')

     gives 40% speedup in code execution.  Now the baroclinic K-filter
     code runs take 4 hours/model-year instead of 8 hours.  In
     comparison, the forward model alone (control) takes 3
     hours/model-year.

====> ACTION : Benny to post summary description of what the
               optimization flags do. 

   - The recovered baroclinic results are comparable to what we had
     before, even after the bug correction of T/P data.  This figure[2]
     compares the dynamic update skill (simulation residual minus
     innovation) of the old result (top two) with recovered results
     (lower two).  kf023 assumes process noise Q to be NCEP wind
     covariance.  kf024 assumes a larger Q the scaling of which is
     from covariance matching.  

   - The negative patch seen in the skill plot above east of the
     Philippines indicates degradation by the filter.  Sea level time
     series are compared at this and other locations in 6-deg bins to
     see what may be the cause of this degradation; 138E-27N[3],
     183E-27N[4], 223E-0N[5].  In these figures, black is T/P, blue is
     simulation, red is kf023 and green is kf024a.  The major
     difference between model and T/P east of the Philippines
     (138E-27N) appears to be a trend in the model; Note black curve
     being lower than the colored curves during 1993-1994 while it is
     opposite during 1996-1997.  Such trend may indicate
     time-correlated process noise which the present K-filter does not
     account for.  Similar trends and degradation were corrected in
     the MOM assimilation by removing trends in computing the
     model-data difference.  

====> ACTION: Assimilation to be repeated with trend removal.

     Such trend is not obvious at other locations where the filter
     does not degrade the model estimate.  In particular, see how the
     filter corrects the blue simulation at 223E-0N. 

   - More model-data comparisons will be done with the assimilated
     results; e.g., XBTs.

====> ACTION: Comparison of K-filter results with XBT data [Dimitris].
              Detlef shows an example of such comparison based on
              upper ocean heat content.

2) Ames Origin 2000
-------------------
   - Ames queue allows time-sharing; i.e., multiple jobs are allowed
     to run simultaneously.   Therefore, our present helper algorithm
     will not work at Ames.   There is also a possibility of
     oversubscription.  

====> ACTION: Zhangfan to see if some other helper algorithm
              can be implemented. 

   - Ames computer is indeed faster than alhena.  Last week's
     assessment was in error.  The summary[6] has been updated. 
     In particular,
           * lomax is large and faster than alhena
           * although the queue is crowded during non-prime-time[7],
             prime-time usage is low and should be taken advantage of
             if the application can be run within 2-hours wall clock. 

3) Others
----------
   - The Division requires us to heighten our awareness and compliance
     of our working hours.  A 45-min lunch time is excluded from the
     8-hour work hours.  Telecomuting (working from home over the
     network instead of working on lab) is not approved.  

====> ACTION: Notify Ichiro of your normal start and end times  
             (e.g., 9am-5:45pm). 

   - Communication needs to be improved within the team to eliminate
     duplication and to achieve a higher level of productivity (e.g.,
     some were not aware of the fortran optimization above that
     speed-up code execution.)   

====> ACTION
       * Each member should summarize at the weekly meeting what he
         has been doing the previous week and to describe what he will
         be doing next.  Each member should think of this before the
         meeting and come prepared. 
       * More postings to the forum should be made.
       * The forum should be upgraded to allow searching strings
         within the text of the postings. 

   - A 50-year run (1950-2000) will be conducted to establish a
     benchmark for a longer assimilation and as a headstart for the
     50-year assimilation (one of the ECCO goals).

====> ACTION: 
        * Dimitris to look into obtaining the necessary 50-year
          forcing. 
        * Zhangfan to put the forcing field on Ames.  Assess whether
          it would be better to archive the fields on the Ames mass
          storage or to rsync the fields from alhena:/tmp whenever the
          fields are necessary. 

   - The present control (c20000630) and K-filter need to be extended
     further into the future. 

=====> ACTION:
         * Zhangfan/Benny to ask Akiko for the data and prepare them
           for assimilation. 
         * Tony will retrieve the forcing field from Charmaine.
           (up to Oct 2000)
         * The T/P data and forcing preparation need to be done on a
           routine basis.  

    - The operating system on three Linux boxes (nautilus, giant,
      stormy) will be upgraded next week.  [Zhangfan]

    - The parallel version of the doubling program will be finished
      this week.  [Zhangfan]

    - Nick is calibrating prior error for assimilation with a wind and
      pressure driven global barotropic model.  Errors are being
      estimated so as to make the innovation sequence as small as
      possible.  The best results are obtained with NCEP pressure
      covariance scaled by 0.1%, which seems too small (if P rms is
      20mb, the error would be 0.6 mb).  The error in pressure is
      believed to be about 2 to 3 mbar.

    - Escher disks are full.  Tony needs ~20 GB.  

=====> ACTION: Everyone to review what he has on guppy and delete what
               is not needed. 

=======================================================================     

-----------------------------------------------------------------------
[1] http://escher.JPL.NASA.GOV:2000/HyperNews/get/forums/cvslog/10.html
[2] http://escher.JPL.NASA.GOV:2000/hosts/escher/escher2/medea/if/forum/t73f5c.ps
[3] http://escher.JPL.NASA.GOV:2000/hosts/escher/escher2/medea/if/forum/t73c6d_1.ps
[4] http://escher.JPL.NASA.GOV:2000/hosts/escher/escher2/medea/if/forum/t73c6d_2.ps
[5] http://escher.JPL.NASA.GOV:2000/hosts/escher/escher2/medea/if/forum/t73c6d_3.ps
[6] http://lochness.jpl.nasa.gov/naso2k.html
[7] http://escher.JPL.NASA.GOV:2000/HyperNews/get/forums/comp/21/3.html



More information about the MITgcm-support mailing list