[MITgcm-support] restart from pickup files (cube-sphere)

[Jean-Michel Campin]

Hi Yuan,

As Dimitris mentionned, when 
a) you redistribute tiles between processors 
and/or 
b) change the size of the tiles, 
you shoud expect different truncation errors, specially with global sum.
If you run just for few iterations, the differences should be small,
and you can check this.

If you keep the same number of tiles (case a)(just changing nSx,nSy,nPx,nPy)
then you don't have to change the EXCH2 files, and if you
1) change "#undef GLOBAL_SUM_SEND_RECV" to
   to    "#define GLOBAL_SUM_SEND_RECV" in CPP_EEOPTIONS.h
2) compile with zero optimisation
you should be able to get the same results (but some compiler and/or
 some very sensitive set-up might still produce little differences ;
 you can check the results of our daily testing, all very short test 
 with no optimization, e.g. with g77 & mpi:
   http://mitgcm.org/testing.html 
 to get an idea of which type of set-up is in this case).

If you change the size & number of tiles (case b), and you are not using 
nonHydrostatic, you can try "#define CG2D_SINGLECPU_SUM" in
CPP_EEOPTIONS.h, with is supposed to be slower but has been 
added recently to check this kind of thing (again, with no optimization).

Cheers,
Jean-Michel

On Fri, Mar 28, 2008 at 08:01:19AM -0700, Yuan Lian wrote:
> Hi,
> 
> I recently notice that the simulation will produce different results when
> I switch number of CPUs from pickup files. For example, originally I
> used 12 CPUs to run the model, then I switched to 6 CPUs due to the
> lack of CPU resources. The simulation ran but results were different from
> those of 12-CPU run (I started from same pickup file but used different
> number of CPUs to compare results) .
> 
> My question is, when restart a job from pickup file with different number
> of CPUs, do I need to modify any configuration other than w2_e2setup.F,
> W2_EXCH2_TOPOLOGY.h and SIZE.h? Thanks!
> 
> 
> Cheers,
> Yuan
> _______________________________________________
> MITgcm-support mailing list
> MITgcm-support at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-support