It's called clustering--the grouping together of several lower-end systems to create one larger computing force d'jour. If that sounds like grid computing it's because, in essence, it is. It just avoids the pitfalls of the Internet.
To get a handle on clustering potential we'll need to revisit the distant past of 1999 for a moment. At the LinuxWorld Expo that year, IBM showed a cluster of 17 of its Netfinity servers containing 36 Pentium II processors. It had purchased a copy of Red Hat Linux off the shelf of a Barnes and Nobles, tweaked it a bit, and set the collection of parts to doing a ray-tracing, image-rendering application that inserted an image into a movie. (It wasn't completely a whip-it-up-in-the-garage affair, as IBM also used a 100Mb Ethernet network and hub to connect the servers and a piece of parallel computing software to ensure the system's computations connected properly.)
According to Tom Figgatt, then IBM's e-business manager, it would have taken a DEC VAX somewhere between 10 and 15 minutes to do the required computations in 1989. In 1999, a Cray T3t-900-AC64 could do it in 3 seconds. That same year, IBM's collaboration of hardware running Linux also did the job in 3 seconds. Do I have your undivided attention yet? If not, keep reading.
Fast-forward three years and you'll notice that the Pentium II is long gone, replaced by Intel with a new line of Pentium 4 Xeon processors that support hyper-threading. IBM itself has a Power4 shared-core dual CPU server and markets Intel Xeon-based systems as well. Although only 3 years removed from the Cray comparison, these new processors are, on their own, orders of magnitude faster than the Pentium IIs used in that test. Although I'm unaware of any new tests similar to the one I've just described, the increase in processing power plus the refinements of Linux should demote that Cray T3t-900-AC64 to a very distant "also-ran."
Alas, all is not upbeat. If you're not already a Linux house, you'll most certainly need to add some savvy Linux staff to create and maintain your cluster(s) and that will take a bite of your budget. Java apps can be ported over. Java is currently one of the most-taught programming languages at colleges, and although college interns aren't free, they do fit into the "reduced cost" budget line item.
So let's summarise: Clustering several low-end systems can produce the same computational power available from a so-called supercomputer; making that happen is not necessarily difficult; there are some additional cost factors involved; and you may not need to re-train for new applications.
And the part that will irrevocably rivet your attention on clustering alternatives: IBM estimated the price of the Cray T3t-900-AC64 used in its test at about US$5.5 million. The price of the cluster it put together to compete with the Cray was somewhere in the neighborhood of $150,000. I'm sorry, I forgot to ask you to make sure you were sitting down before you read that, didn't I?
http://openmosix.sourceforge.net/