Beyond the Data Center: Distributed Computing on a Raspberry Pi 5 Cluster

Abstract

Distributed computing clusters are commonly used to provide scalable computation and large memory capacity for demanding workloads. In recent years, single-board computers have become increasingly capable and power-efficient, making them an attractive low-cost alternative for building small-scale distributed systems. How ever, creating such clusters in a way that is scalable, practical, and user-friendly remains challenging due to limited hardware resources and the need for lightweight management and monitoring solutions. This thesis investigates how a distributed computing cluster built from single-board computers can be made practical through lightweight orchestration and purpose built observability tooling. A central contribution is a custom telemetry system de signed for resource-constrained nodes, where existing monitoring solutions impose unnecessary I/O on storage-limited hardware and offer limited control over which metrics are collected and how frequently they are reported. The system collects, transmits and visualizes hardware and performance metrics in real time through a custom web based interface while imposing no measurable impact on workload performance. To evaluate the system, a Raspberry Pi 5 cluster was constructed using Kubernetes for orchestration. Three workloads were deployed to stress different dimensions of the cluster: matrix multiplication for parallel compute throughput, distributed pass word recovery for CPU-intensive data parallelism, and split large language model in ference for distributed memory capacity. The results show that the cluster achieved significant performance improvements compared to single-node execution, particu larly for highly parallelizable workloads. The system also demonstrated good power efficiency and highlighted the advantages of distributed memory for running larger LLMs. However, the limited computational performance of individual Raspberry Pi nodes means that many devices are required to approach the performance of a conventional high-performance machine. Overall, the work demonstrates that single-board computer clusters can provide a flexible and energy-efficient platform for distributed computing, especially when combined with lightweight orchestration and observability tools.