> > This kind of reasoning goes nowhere. If the rational argument that
> > the precision is better, thus the (new) product is also better
> > doesn't work in this case, how can you ever add new optimizations
> > to the language or, e.g., better precision FPU's?
> For models as complicated as our forecasting model, sometimes the
> only way to ensure a hardware/compiler upgrade has worked it to get
> bit comparibility. For climate work, exact repeatability across many
> years is essential. One of the biggest issues of our C90 - T3E
> upgrade and MPP conversion was the precision changes due to the
> change in number format from Cray to ieee, and the calculation order
> changes required for MPP.
> This disruption can be accepted for such a major hardware and design
> philosophy change, but any suggestion that language changes would
> also affect reproducability would worry us a lot, and disuade us
> from using new compilers or language features. We are introducing
> large Fortran 90 systems into operational use now, but the main
> forecast model is still resolutely f77 for this kind of reason.
I respectfully submit that, if your models are so sensitive to "change
in number format from Cray to ieee, and the calculation order changes
required for MPP," then your algorithms are decidedly unstable.
-jack
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|