Cluster is a very general term. I have a hard time putting words to the different types so I'll just explain them. (I generally reffer to the different types as single and multiple processes, but this is a ugly way to say it.)
distributes tasks. Say you want to convert a WAV to a MP3 and this takes you 1 minute. If you had 10 clustered computers running OpenMosix, and you did this on the head node, it would still take you 1 minute. Now if you converted 10 WAV's to 10 MP3's it would take you 1 minute, still, because OpenMosix would put each WAV to MP3 conversion on each of the nodes. This is really cool and is very easy to implement. Check out ClusterKnoppix, which of course, runs from a CD, is a knoppix distro, and runs OpenMosix. Analyze Quantin is ClusterKnoppix based and comes with some spiffy apps that generally requre clusters to compute such as some of the cool things done at CERN (the particle accelerator in Switzerland.) Don't be afraid though, you can actually tackle some calculations with some old computers! I would call this a multiple process cluster.
Now there are so-called single process clusters, which I think are usually reffered to as parallel clusters, but this can be confusing, because in the above example, our programs were running parallel to eachother. In a parallel cluster, the source code of the program has been parallelized. You can't put a WAV to MP3 conversion on 10 computers and have it go 10 times faster. Why? Because the WAV to MP3 program has been designed to work on 1 cpu. If you parallalized it, it would mean that your audio conversion program acutally splits the work done by the processor, sends it out, has it processed, brings it back to the head node and puts it togeather as the MP3. This is actually a difficult thing to do, since parallalizing code can be a very complex task. Generally you have to write a program from scratch. This is done alot by corporations of government entities, for example, you have one bitch of a math problem to solve, such as calculating pi. You can't run a pi calculation on a multiple process cluster like the one in the first example. Each stage of the pi calculation generally builds upon the other. (except on parallized pi programs.) You could run it on an OpenMosix cluster, but it wouldn't go any faster. Plus, you would have to have a program that could modularize the stages of the calculations. But you could run a WAV to MP3 and calculate pi on OpenMosix and that would go faster, because you would have more processing power to go around.
Hope this helps.