Matthew Lewis

Configuring a Single-board Parallella Cluster For Parallel Execution of Benchmarks

Traditional clusters are very expensive, require a lot of power and take up a lot of space, and as such are only available to large research groups and industry.  Clusters consisting of single-board computers, however, are more affordable for smaller companies and even individuals to own.  Using the Parallella single-board computer option, this research seeks to ascertain to what extent a cluster comprising these single-board computers can be useful for high-performance computing.   For the cluster, the Parallella nodes were set up with relevant frameworks and libraries and a Raspberry Pi controller node, external to the cluster, was set up to perform supporting functions for the cluster, these included: DHCP, NFS and NAT.  For benchmarking, a matrix multiplication program was adapted to use Open MPI, and this was run on the cluster, with a slightly different version run on a single-node and a standard CPU.  The results showed a 1.5× to 2.5× speedup on the cluster relative to the single-node, but the standard CPU had a 4× speedup relative to the cluster.