Beowulfs can compete with Supercomputers [was Beowulf: A theorical approach]

[Greg Lindahl]

> :-) Indeed, but the reason why this one is wrong is easy:  Yes, it is
> hard to change codes from a shared memory threaded model (e.g. OpenMP)
> to a distributed memory model.  However, the field of weather
> forecasting is probably one of the few remaining where programmer time
> is neglegible when compared to other costs, like the global observation
> system (including satellites).

This is true. However, it's not always so hard to parallelize codes. For
example, if you have a stencil code or a spectral weather code, the SMS
system from FSL can be used to parallelize a serial version in just a few
programmer days of effort. Check out:

http://www-ad.fsl.noaa.gov/ac/sms.html

SMS is a good example of a domain-specific system that provides the benefits
that HPF promised, at the promised cost.

I would encourage folks to examine systems like SMS very carefully. It has
limitations -- limited F90 support -- but if your code does what it does
well, you're in great shape. My PPM code would parallelize fairly easily
using it.

-- greg