Almost all of the fastest parallel supercomputers in the world utilize the Beowulf cluster architecture, which distributes processing tasks over a local network to a multitude of individual nodes. The idea, of course, is that those nodes are readily-available and relatively inexpensive, which makes it possible to build a supercomputer from off-the-shelf parts. Matt Trask decided to build a Beowfulf cluster with Raspberry Pi single-board computers, and created a very unique servo-based processor visualization system for it called Pi VizuWall.
As Trask explains, Beowulf clusters rely on a system called MPI (Message Passing Interface) that divides a program up and distributes it to the individual nodes for processing. Unfortunately, there are very few MPI programmers in the world. That means that manufacturers don’t produce many parallel supercomputers, because there aren’t enough experts in the industry to support them. That, in turn, means that very few new programmers are trained, because the equipment isn’t easy to gain access to. That’s a vicious cycle that needs to be addressed.
The solution, as Trask sees it, is to construct low-cost Beowulf clusters that can be installed in universities so that students have the opportunity to learn MPI programming. At just $35 per computer, the Raspberry Pi 3 Model B+ is perfect for the job. To prove the concept, Trask built a Beowulf cluster containing 12 Raspberry Pis called Pi VizuWall. It was given that name because each Raspberry Pi is mounted on a 3D-printed hinge and actuated by a servo. A lightweight program runs on each Raspberry Pi to monitor processor usage, and that determines the servo position. The result is that students can easily visualize how their MPI programs are affecting distribution within the cluster.