Saem:
My friend and I did a lot of clustering back about 2 years ago when I was doing a lot of stochastic raytracing. Unfortunately I didn't really deal too much with actually setting up the cluster itself (they were all his machines, and in another state).
If I remember correctly, with beowulf you'll need to use software that is written specifically to use PVM for the load balancing and distribution to take place. This is probably ok for stuff like povray and some of the other distributed apps that have compiled versions to be used with beowulf. Otoh, a better proof-of-concept show may be to use mosix. Mosix is clustering software for linux that basically allows processes to be distributed between systems (and cpus) in the same way that you'd expect SMP to work on a multiprocessor system. Basically you tweak the resource usage values (how important bandwidth vs processor vs memory is to you) and it automatically farms out processes to whatever machine(s) seem to be the most worthy. What's probably more impressive, is that it works with reguluar multithreaded applications (I think this includes both kernel and userspace, but not sure).
Mosix is what we used on the cluster for distributed rendering, and it worked fairly well once we got the parameters set right. It can get kind of bad when it doesn't know what do to about resources (say it runs out of ram on a node) and starts trying to move the process around to different computers, hosing the network. Otoh, overall I was pretty amazed at how well it seemed to work. Pretty neat that it could have seti running on all the cluster nodes, we could tell it to migrate the process off to another node, shut down the system, boot to windows, play halflife for a while, reboot the system to linux, rejoin the cluster, and migrate the thread back to that node as if nothing ever happened.
Nite_Hawk