Just saw a great talk by Alexander Szalay about the work at John Hopkins university to develop a cluster to perform data intensive computing. The resulting cluster just won the HPC Storage Challenge at Supercomputing 2008.
There approach is to reach back to Gene Amdahl’s rules of thumb for computer architecture and apply these to large scale parallel machines. See his article with Gordon Bell and Jim Gray in IEEE Computer 2006 (10.1109/MC.2006.29). It’s a nice way of looking at how to design such machines and it works.