[MITgcm-devel] Re: argghhh

[Jean-Michel Campin]

Hi Dimitri,

Thanks for this usefull test. I will make the changes in
DIAGNOSTICS_FILL relatively soon and will tell you when
it's done.

Jean-Michel

> On Friday 17 December 2004 11:37 am, you wrote:
> > One thing also, if you notice any efficiency/speed
> > difference with diagnostics versus timeave, I would
> > be interested. I was planning (next week) to change
> > the way DIAGNOSTICS_FILL is written so that it does not
> > use local copies but accumulate directly to the storage array.
> > This might have a small benefit in term of efficiency.
> 
> Jean-Michel, last night, for a little while, I ran two parallel jobs (1/8-deg
> global on 960 CPUs each) using timeave and diagnostics, dumping
> the exact same fields in both cases (ETAN, PHIBOT, UVEL, VVEL, WVEL,
> THETA, SALT, TAUX, TAUY, SFLUX, TFLUX).  On average the integration
> with timeave requires 37:09 per 3-day interval while the integration
> with diagnostics requires 40:17.  That's almost a 10% overall
> performance hit.  Memory footprint is a big cost thing for large
> computations because it impacts cache utilization.  It would be
> great if you were able to make the mods you discuss above.
> 
> Cheers, D.