[MITgcm-devel] testing status

[Jean-Michel Campin]

Hi,

I have been working on the testing part, trying to reduce
the differences between compilers. There is still some
work to be done, specially regarding the testing on 
different platforms (I only did comparison on pentium 4).
But already get few interesting results:

1 ) on faulks with ifort (now tested daily), most of the tests that 
fail can be explained by some missing "_d 0" (or, to be more precise,
the missing _d 0 is enough to explain the differences we get):
 cfc_example: in pkg/cfc 
 dic_example: in pkg/dic (<- I assume, but didn't check this one)
 dome: in dome/code/obcs_calc.F
 fizhi-gridalt-hs: in model/src/getcon.F
 global_ocean_ebm: in pkg/ebm/ebm_atmosphere.F
 lab_sea: KPP
 nat_box: KPP (but not sure of nat_box/code/external_fields_load.F)
 vermix: KPP
 vermix.my82: in pkg/my82/my82_init.F & my82_readparms.F
the only one that is less clear is:
 global_ocean.cs32x15.thsice that fails at level 11, and might be due to
 math function + sensitive set-up (starting from a pickup ...)

2) on itrda with g77 + mpi (2 processors run), all the exp pass nicely, 
except global_ocean.cs32x15, global_ocean.cs32x15.thsice and ideal_2D_oce.
all 3 use more than 2 tiles, so that the results of the global sum 
could make a small difference; they all start from a pickup, which 
make them more sensitive, and in addition, ideal_2D_oce is 2D which 
make it even more sensitive regarding the cg2d_ini_res.

3) still some problems with ifc, specially regarding the exp.
global_with_exf for which "ifc -O0" give strange results.

I have a suggestion:
When checking-in new pieces of code, it could be a good idea to 
put the "_d 0" where it's needed. It's not more difficult like this, 
and it makes the results much more platform/compiler independent.
It also make the testing more reliable, since I saw several small 
bugs that affect the results of the test much less than changing 
a crucial "_d 0".

Jean-Michel