I'm no expert on beowulf clusters, but as I understand it they can be setup to behave as if they were one large computer. The method of having one application use memory and cpu from each node in the cluster requires the application be written using things like MPI, PVM and MPICH. Blender does not make use of these so would require a rather in depth code alteration to take advantage of this setup.
However blender does have a built in network render available. Basically you setup one machine as a master that tracks the rendering tasks, and several computers as slaves to do the work. Then you have a client on your desktop that sends the rendering task to the master and gets the result back. BlenderCookie has a tutorial covering the basics. While the video is citizen only the main info is detailed on the page. The main drawback of network render is lack of support for calculating simulations using this setup. The other consideration is it is intended for animations, each node renders one complete frame.
You can also setup your own render farm. Using CLI options or python scripting you can have each computer render images. With a little effort you can create script that positions cameras to break a final image into separate renders that can then be stitched together.