The Raspberry Pi family of single board computers transformed education and has become one of the drivers of the Internet of Things revolution. These boards are low-cost, well-featured, and easily available. So what do you get when you take a bunch of Raspberry Pi boards and wire them together? An incredibly cheap and surprisingly high performance parallel computing system that’s not only valuable in education but can also solve some useful, real world problems.
In this round up we have 10 Raspberry Pi clusters ranging from tiny, four-node systems all the way up to 250-node behemoths. These clusters are being used for everything from teaching through to simulating massive IoT networks to creating leading edge art. If you know of any great Raspberry Pi clusters to include in our next roundup, let me know.
Comments? Thoughts? Drop me a line or comment below then follow me on Twitter and Facebook. And sign up for my new newsletter!
Muratovic’s way cool, water-cooled RPI 3 cluster is just one of his designs and his instructions and videos are applicable to building much larger systems. He also covers how to install MPICH, “a high performance and widely portable implementation of the Message Passing Interface (MPI) standard” along with MPI for Python and then run a Python test program to find primes.
The Raspberry Web Server site’s section on setting up a Raspberry Pi Web Server Cluster and load balancing and Raspberry Pi configuration how-tos are great resources for cluster builders and can be applied to creating much larger cluster systems. Other sections on the site include cluster performance tests, web server tuning, and adding more nodes.
Developed by Kim Sung-Taek, the Raspberry PI 2 Cluster (also called the Pi Spark) is based on six Raspberry Pi 2B+ boards and was developed to run as a Hadoop cluster. Sung-Taek provides not only the detailed steps required to build the setup but also the templates for creating the physical framework. The final system was configured using Sung-Taek’s PocketCluster utility which was created to “setup a multi-nodes BigData cluster on OSX.”
Developed by Afkham Azeez, this is the second iteration of his Raspberry Pi Cluster and boasts an aluminum frame that’s built to last, a very sophisticated power distribution system, and a ton of instrumentation to support running benchmarks. Azeez commented, “Once the cluster was up & running, we ran a series of JMeter load tests. Preliminary testing was conducted using an 8 Pi cluster [that’s a single tier of the cluster]. At a concurrency level of 750, the cluster was able to sustain a throughput of 350 TPS at an average latency of close to 1 second.” To see the full cluster in action, Azeez has a video from his presentation at the 2013 WSO2Con conference in London.
According to the project’s page: “We have created a 24-node Raspberry Pi 2 cluster, with each node instrumented for detailed power measurement. This totals 96 cores with 24GB of RAM. / The cluster draws around 92W when under load, and around 70W when idle. / There is a head node (also a Pi2) with a touch-screen interface and two ethernet adapters that controls the ATX power supply (it runs off of standby power), serves DHCP, NFS, and [Ganglia]. We use SLURM for job submission. / It obtains 15.4 GFlops of performance, meaning in June 1993 it would have been #7 on the Top 500 list.”
The RPiCluster Project is definitely one of the coolest looking clusters. Based on the Beowulf cluster this system was built by Joshua Kiepert while at Boise State University: “The RPiCluster project was started in Spring 2013 in response to a need during my PhD dissertation research. My research is currently focused on developing a novel data sharing system for wireless sensor networks to facilitate in-network collaborative processing of sensor data … I began developing a distributed simulation in which each simulation node would behave like a wireless sensor node (along with inherent communications limitations), and as such, interact with all other simulation nodes within a LAN. This approach provided true asynchronous behavior and actual network communication between nodes which enabled better emulation of real wireless sensor network behavior.”
Built by the University of Southampton, the Idris-pi is a 64 node Raspberry Pi Model B cluster in one of the best frameworks ever; a chassis made entirely of Lego! The paper on the system, Iridis-pi: a low-cost, compact demonstration cluster, details the power and performance characteristics, notes that “The total cost of the system excluding network switches was below £2500,” and concludes, “We foresee that, although our cluster architecture is unconventional by today’s standards, many aspects of its design—the use of open source hardware and software, the adoption of low-power processors, and the wider application of flash based storage—will become increasingly mainstream into the future.” Check out their video reviewing the Iridis-pi.
The goal of resin.io‘s Beast v1 was to be a test bed for large scale IoT implementations. resin.io explains that the system consisted of “120 Raspberry Pi, each with a 2.8″ Adafruit PiTFT screen, all mounted on a wood panel in a somewhat eye-pleasing format. Behind the panel, we mounted 10 USB hubs, and 10 Ethernet Switches. A Corsair CX750 power supply unit completes the beast … The build is composed of 5 independent 24-node clusters that are unified only by the joint provision of power and networking and can operate independently if needed. They are held together with hinges, and the arrangement of the parts in the back of each section is such that the entire configuration can be folded into a tube with an almost pentagonal cross-section.”
Obviously not the type to rest on their laurels, the resin.io chaps upped their RPi cluster game to create the Beast v2. resion.io explains: “Weighing in at nearly 150 kg and measuring nearly 2m tall, Beast v2 truly lives up to its name … Like its predecessor, Beast v2 is a huge cluster of Raspberry Pis, 144 to be exact.” But wait ! There’s more: “We learned a lot in the process of building Beast v2 and immediately started on a new and improved design for Beast v3. One thing we realized right away is that, although Beast v2 impressed in size and scale, it didn’t have the elegance of Beast v1. The tiles on Beast v3 will therefore be denser, with less empty space between devices. We want to squeeze in as many devices as possible.” Check their blog for future Beast v3 updates.
Developed by Sam Blanchard, an assistant professor of sculpture at Virginia Tech, “SeeMoreis a project that melds Art and Engineering to distill abstract concepts in computing into a tangible experience that informs viewers that computational thinking does not exclusively exist within the preverbal ‘black box’. This kinetic sculpture showcases the inherent beauty of parallel algorithms through the correlating movements of an animatronic 256-node Raspberry Pi computer cluster.” The video is way cool.
