Possible Large memory problems on commodity computing

Hi guys.

Is it possible to run a large memory problem on a Beowulf cluster, or will a severe bottleneck appear at the RAM available for each node? Or is a Beowulf cluster set up so that it is only the *total* RAM that is important, and not the RAM per node?

With just 8 variables, the problem uses several hundred megabytes of RAM upon initialisation. Soon, this drops down to a few hundred megabytes. At least 32 variables are needed for a quasi-realistic simulation, ideally: hundreds.

If someone could point me to a parallel computing introduction for dummy beginners, I would be most grateful. The guides I've found are either too brief or focus too much on the technical details like setting it up. Yes, I'm interested in how to set one up at home (for the heck of it) but I would like to know how the MPI system works, how jobs get divided, whether a licence for the binary is needed at each node, etc.

E.g. if N simultaneous independent runs get run on N processors, then one needs N licences. But if the parallel system focuses on finishing one task as quickly as possible using all the processors to crunch through the arithmetic, then only one licence would be needed. All the other processors should be seeing is arithmetic calls.

Thanks.