Illinois Genetic Algorithms Laboratory

A year and half ago, we built a 90-node diskless PC cluster, similar to (and inspired by) the Beawolf cluster in John Koza’s lab for running all our GEC experiments. We mostly use the cluster to run multiple GA runs to collect statistics, and also for testing the effect of different parameters on the scalability of the algorithms (For example, see Martin’s blog). Since then, quite a few people have asked me a how to build such a cluster, what components to buy, etc.

So, I decided to put it all here in the blog (note that the component costs are rough estimates, and Pricewatch and Newegg are good sources to check prices):

• Motherboard (ASUS A7N8X-X or similar which has built-in LAN with PXE boot): $70
• If the motherboard doesn’t come with PXE-boot, you have to buy a network card with PXE-boot such as 3-COM 3C905CX-TX-M (which costs about $25).
• Processor (Athlon XP 2800 Barton core): $100
• Go with retail box which comes with fan+hitsink, and 3 year warranty
• Memory (At least 512MB DDRPC2700): $100
• I usually prefer Kingston, Crucial, or Corsair
• Case + Power supply (~350 Watts should be fine): $50
• Montior + keyboard + mouse + video card for the server: $180
• Software: Gentoo Linux, Open Mosix

Don’t forget, the new motherboards and processors drain a lot of power, so you need about 1.5 Amps per node. That means they produce a lot of heat, but nothing that a few well placed fans and couple of ACs can’t fix.

Total estimated cost of building the cluster: 500 (server) + 320 * (# nodes).

Limitations:

• The memory that a program can use is limited by the system memory (no swap as there is no disk!)
• Open Mosix doesn’t migrate Java processes (at least not yet)
• I/O may be a bottleneck