A Low-Cost Single-Board Computer Cluster: Architecture, Design and Use Cases

Abstract

In the field of high performance computing (HPC), supercomputers are used to solve computationally large problems. However, barriers have been informally identified regarding the usage of HPC systems. Not every potential user can gain access to a supercomputer nor can it be quickly bought since such a machine comes at a large total cost of ownership. Existing systems are shared among multiple users, leading to problems. Additionally, they are usually remotely accessed, raising privacy and security concerns. Finally, the usage of the system itself can be a threshold since its usage is more complicated than a machine typically owned by a user.

In order to solve some of the issues in HPC, a computer cluster consisting of low-cost single-board computers was developed. Through evaluation of two real-world problem cases, the cluster was found to scale equivalently to a full-edged cluster for programs that have a large parallel fraction. The performance of the parallel regions is very reasonable considering the cost of the cluster, although sequential code severely diminishes the advantages gained through parallelism. Several methods for user friendliness were deployed and were informally confirmed to reduce the usage threshold. The cluster does not replace existing HPC machines, rather, it bridges the gap between user systems and full-edged clusters. It is complementary to both, since for large sequential parts, the user's own machines are the better choice. For raw performance, an existing HPC machine should still be used.