Building a Raspberry Pi cluster.

Single-board computers are not suitable for every analysis. They have limited memory and are slower than a contemporary workstation CPU. But they are cheap, they parallelize well for jobs that fit the model, and they make sense for fitting many models to the same dataset, running many iterative simulations, and the rest of the embarrassingly parallel category. A cluster under our own ceiling also lets students experiment with scheduling, MPI, and systems administration in ways that a university HPC does not. The physical cluster is a teaching surface as much as a compute resource.

Most components were sourced from Amazon. With careful shopping or repurposed equipment the costs could be reduced further.

Not included are items already on hand and donated to the project: a 1 TB SSD external hard drive, monitor, keyboard, mouse, an 8-port Gigabit Ethernet switch, and miscellaneous cables.

Power configuration. Six 6-port chargers were purchased for 36 total ports for 26 Pis. After research on power requirements, most power packs run only 4 Pis each to ensure adequate power delivery, especially when USB devices are connected.

• Build a Raspberry Pi cluster, MagPi Magazine
• Five years of Raspberry Pi clusters, G. L. Miller

Assembling the Pi racks was straightforward:

1. Install heatsinks on all Pis.
2. Attach fans to mounting plates, designed for five Pis each.
3. Combine plates into three towers: two with 8 Pis and one with 9 Pis.
4. Mount each Pi on standoffs and stack plates in towers.
5. Connect fans to pins on the Pi below each plate.

The head node is a separate Raspberry Pi 4 (8 GB RAM version) housed in a standard case. This design lets us later replace the head with a more powerful Linux machine from System76 without disrupting worker nodes. At this stage, cabling, disk setup, and software installation remain.

Initial testing compared Ubuntu Server 20.04 with a desktop (LUBUNTU) to the standard 32-bit Raspbian (Buster, based on Debian 10). Ubuntu is more general-purpose, but Raspbian was more responsive with lower RAM overhead and was ultimately chosen for all nodes.

The installation process on the head node follows these steps:

1. Install Raspbian Buster with system updates.
2. Compile R from source (version 4.0.2).
3. Install required system libraries and R packages.
4. Install OpenMPI for distributed computing.

R compilation. Compiling R requires installing multiple dependencies. If the configure script fails, search for the missing library using patterns like "how to install libXXX on raspberry pi". Solutions are typically found in the first few results.

Note. PDF man pages failed to build (texi2any version conflict), but base R and all packages installed successfully.

After head node configuration, use Raspbian's built-in SD card clone utility to create 25 copies, about 20 minutes per card. To enable headless SSH setup, add an empty file named ssh to the boot folder before first boot. Still, connecting a monitor during initial setup is recommended to verify HDMI detection.

HDMI detection issue. If Pis boot without HDMI cables connected, they don't start the X11 windowing system, producing no video output even after plugging in a monitor. Always connect HDMI cables before the first boot.

Configure each node (head first, then workers) by performing these steps on boot:

1. System configuration

Open the Pi configuration application and:

• Change the hostname: head node "stevens"; worker nodes "node1", "node2", and so on.
• Update the default password if desired.
• Select "No" when prompted to reboot. We'll reboot after additional changes.

Naming convention. All lab computers are named after great scientists. The head node is named "stevens" after Nettie Stevens, a pioneer in genetics who wrote the pioneering Studies in Spermatogenesis volumes (1905, 1906) on sex chromosomes.

2. Static IP configuration

Set up static IP addresses so that node identities stay consistent across reboots:

Add these lines to the file (increment the third octet for each node):

For example, head node uses 10.0.0.1, node1 uses 10.0.0.2, and so on, up to 10.0.0.26.

3. Enable SSH

SSH is disabled by default and does not persist after reboot when enabled through the GUI. Enable it from the command line:

Verify SSH status:

4. Reboot and verify

After reboot, verify the hostname change and confirm SSH is running.

Generate SSH keys for passwordless communication between nodes:

Press Enter through all prompts to use defaults. Configure key sharing:

On worker nodes. Copy the head node's key:

On head node. After all nodes are configured, copy the head's key to all workers:

Verification. From the head node, test passwordless SSH access to a worker:

Once all nodes are configured, finalize the physical cabling setup. In hindsight, slightly longer network cables would have provided more flexibility for cable management and rack organization. The current setup works but leaves minimal slack in some cable runs.

Coming soon: distributed computing benchmarks and first job submissions.