When you think of the Raspberry Pi, words like "small," "affordable," and "flexible" come to mind, right? And you probably don't then add "supercomputer?" It turns out, however, that those attributes of the Pi can be quite useful in the field of high-performance computing, which I personally associate with beastly machines made by Cray or IBM. The concept behind Bitscope's Cluster Module is simple: take a large group of Raspberry Pis and put them to work in tandem, creating a system that combines great power with scalable design. This approach grabbed the attention of no less than Los Alamos National Lab, which has faced the problem of needing to develop software for supercomputers while not having available time, as they're tied down with other research, as well as the fact that these machines are rather expensive to run. They hope that the Raspberry Pi cluster concept could prove a timely solution.
Says Bruce Tulloch, CEO of Bitscope: "Our Bitscope Blade is the key building block we used for the cluster. We originally built the Blade boards as a power mounting solution for Pi ... [and] as soon as we built them we had people asking “Can I stack these, can I put them together?” So we created the Blade rack series, small clusters of 10, 20, maybe 40 nodes, usually. But then people would start buying several of those to build bigger clusters, and finally SICORP approached us with Gary Grider's request for a larger cluster. And of course we said “Oh yeah we can do that, how big do you want?” and then they asked for "10, maybe 50 thousand nodes?" (laughs) And we said “We can do that.”
The idea of combining a cluster of Raspberry Pis in this way had been around for some time, but in a more limited and less practical form. Explains Tulloch: "In the past, people have toyed with the idea of using Pi with clusters, mounted on cardboard or plywood or something along those lines, and people have seen it as a bit of a toy. But when you actually mount a Pi and power it reliably in the way Bitscope Blade does, it gives you a solution that we’ve found is remarkably resilient and reliable. So the Pi actually is an industrial quality product in a lot of applications. Now that Los Alamos is planning to scale it out in the way they are we'll certainly find out if that’s true!"
Bitscope brought its Cluster Module to the Supercomputing 2017 conference in Denver and found a good deal of interest in their concept. "I think what we found when we exhibited at the show," says Tulloch, "is that everyone loves the Pi. So when they saw our cluster at the show they all said “Wow, look at that!” But then of course the next question is “Why??” (laughs). But almost every university student who came up to the stand, you could see their mouths watering at the prospect of having access to a hundred nodes or two hundred nodes to do their course work or their post doc research, exploring things having to do with architectures relating to scalability." And the appeal of the Cluster concept isn't limited to education, it has potential for private sector companies as well. As Tulloch explains: "Another area that we had a lot of inquiries about at this show was distributed computing, think more cloud services and so on. We see things like function as a service rather than software as a service being rolled out on [a system of] very low-powered nodes but in large numbers. For highly distributed computing, even though the Pi has a small amount of memory it’s not an issue in those applications."
Though a dense cluster of Raspberry Pis might seem like an odd concept, in practice it actually relates well to a common tool for designers. Explains Tulloch: "I think a term that SICORP came up with [to describe the Cluster Modules] at the show says it really well, which is a 'high-performance computing breadboard.' Effectively, what we’re doing for people building clusters is giving them a tool in the same way that electronics designers have breadboards, to put together their ideas at low cost, low power, very quickly, and in a flexible way." The Pi clusters may not be able to compete with the Los Alamos Trinity or Crossroads supercomputers in terms of computational power, but they provide a more cost-effective way to test development code without needing the use of a massively expensive and energy-intensive machine like those.