Defined as "a Beowulf clusterof Raspberry Pi devices"
It's not useful except as an academic exercise, as, despite the fact that the RPi has excellent GPU capabilities, it's CPU is mediocre and it has very limited IO capabilities (which are critical to any cluster). This is because ethernet and USB share same bandwidth and so sustaining high bitrate transfers between pis is likely to be very difficult.
You can make a processing cluster with anything that has a CPU and a network interface, and it will be effective for a range of applications that lend easily to loosely-coupled parallel tasks. For learning about parallel computing and the problems involved, R-Pi could be suitable, and you have always got 64 R-Pis which can be re-used afterwards. Thus, in a school or university environment, it might be doable.
For practical applications, the R-pi is a cheap, but not the best MIPS/$. Putting together 64 R-Pis might cost around £2,500, so you might think twice about spending that much.