You can keep the OS X as is. Load current JVM on all of them and connect them to a network.

Then write a program that uses Hazelcast (http://www.hazelcast.com/) that gives you a pseudo-distributed JVM cluster. It actually gives distributed Maps and Queues.

Then mount a bunch of Jetlang (http://code.google.com/p/jetlang/) Fibers into the distributed Map. The Jetlang fibers provide thread-less concurrency.

Then develop a simple function-independent task framework, to which you can pass a function and a parameter map as two arguments, which calls a Callable and returns an Object.

Now create a bunch of tasks, mount them into fibers and mount the fibers into the distributed map.

Finally write a simple countdown latch based task runner, you can safely run four per core, and run them on each mac, sit and watch.

Be careful, you can easily destroy a few database servers..

Apart from that, I find using older, slower computers an interesting exercise. What appears instantaneous on my netbook takes an observable time on my iMac G3. That usually allows me to guesstimate what was used to implement the function and to get a feel of the bottlenecks in the app. It's useful if you don't feel like profiling the app.