Changes between Version 4 and Version 5 of ticket/1473_METO_ISFCLEAN


Ignore:
Timestamp:
2015-03-11T11:28:34+01:00 (6 years ago)
Author:
mathiot
Comment:

Legend:

Unmodified
Added
Removed
Modified
  • ticket/1473_METO_ISFCLEAN

    v4 v5  
    1414The main issue is the order of the loop when we need to start from mikt(ji,jj). To order the loop jk/jj/ji instead of jj/ji/jk, the first solution is to create a wmask, wumask and wvmask in order to mask all the variable at w point properly. Like this the mask system can be applied at the top as it is done at the bottom. 
    1515 
    16 In some place we also need to change the order between the initialisation of the interior value and surface value (3D loop jj/ji/jk => 2D loop jj/ji with mikt in it) 
     16In some place we also need to change the order between the initialisation of the interior value and surface value (3D loop jj/ji/jk followed by a 2D loop jj/ji with mikt in it) 
    1717 
    18 Duplication of some routine are done in order to keep the same performance than the one before the merge if ln_isfcav is false (hpg_sco => hpg_isf and zps_hde => zps_hde_isf and the old one are restore) 
     18Duplication of some routine are done in order to keep the same performance than the one before the merge if ln_isfcav is false (hpg_sco => hpg_isf and zps_hde => zps_hde_isf and the old one are restored) 
    1919 
    2020svn command: 
     
    3434sqlite[S8]: attempt to write a readonly database 
    3535 
    36 Commit 23/02 : remove all 3 d loop + some missing flag and add 3 masks (wmask, wumask, wvmask). Sette OK and not change in the solution of sette. For ORCA1 (acc), to have the same results we need to force the zgr_isf (bug from the trunk move to this subroutine) and this subroutine is only activated if ln_isfcav. Need also to force zps_hde_isf and hpg_isf as we restore hpg_sco and move the current one to hpg_isf. 
     36Commit 23/02 : remove all 3 d loops + some missing flags and add 3 masks (wmask, wumask, wvmask). Sette OK and not change in the solution of sette. For ORCA1 (acc), to have the same results we need to force the zgr_isf (bug from the trunk move to this subroutine) and this subroutine is only activated if ln_isfcav. Need also to force zps_hde_isf and hpg_isf as we restore hpg_sco and move the current one to hpg_isf. 
    3737  
    3838Commit 02/03 : add some flag and compatibility test + doc. Sette OK and no change in the solution of sette. 
     
    4141How ice shelf cavity is coded need to have e3t(jk) = (gdepw(jk+1)-gdepw(jk)) for the pressure gradient. The start point for the water column is risfdep(ji,jj) and not 0. So, level in depth can be not aligned (ie can have a different gdep3w with the default e3t definition).  
    4242 
    43 Change in runtime: ORCA1-LIM3 (acc)     before : 254.181 s (average) after : 165.250 s (average) 
    44                    ORCA2-LIM2 (mathiot) before/after : -20% of runtime (1y ORCA2 LIM2) 
     43Change in runtime:  
     44* ORCA1-LIM3 (acc)     before : 254.181 s (average) after : 165.250 s (average) 
     45* ORCA2-LIM2 (mathiot) before/after : -20% of runtime (1y ORCA2 LIM2) on BAS local cluster 
     46* ORCA2-LIM2 (mathiot) before/after : similar total runtime (1y ORCA2 LIM2) on ARCHER, however, variability of runtime for a same job is about 20% depending of date of test, this test is maybe bias by disk load (or something similar). Timing output show improvement in the modified routines as zdftke, zdfavn, traldf_iso (...) 
    4547 
    4648----