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

[Greg Lindahl]

> i would like to emphasize that in supercomputing sites the people are
> interested not only for computing power (though it is very important)
> but for available software. In traditional fields where supercomputers are
> used (like modeling for various reactors, weather forecast) there are
> also a number of programs and applications which can not be moved easy to
> Linux cluster.

Generalizations are always dangerous.

Weather forecasting is a field in which MPI codes are seen to be the wave of
the future. Big US weather forecasting sites, for example, include NCEP (IBM
SP), AFWA (IBM SP), GFDL (currently has an RFP for which 100% of the codes
use MPI), and FSL, which owns an AlphaLinux Myrinet cluster from my
employer. Large European sites own machines which are also programmed using
MPI.

Once you're using MPI, it's a matter of getting the right balance of
communication bandwidth, latency, and processor power. This can be done
today.

> One of the way to speed up the process is to realize the real
> supercomputing site on the base of the Linux cluster and demonstrate the
> traditional application software packages which might be attractive for
> end users in conventional supercomputing sites. It would be collaborative
> project in between cluster suppliers, software developers, universities,
> etc.

Many supercomputer sites use IBM SP machines, which are clusters, so there
isn't that much to prove. The Forecast Systems Lab (FSL) machine is a
production Linux cluster with good uptime. Whenever you fly in the US, the
FSL machine is the backup machine providing the weather forecast for your
flight.

-- greg