Building a Parallella board cluster
Michael Johan Kruger · 2015
Currently, high performance computing is largely restricted to well-funded research groups. This project aimed to create a high performance cluster using the cheap and energyefficient 18-core Parallella boards. Four boards were connected over a network as a cluster and basic test programs were run using MPI. Experimental results show that the Epiphany chip performs very well compared with other energy-efficient chips such as the Cortex A9 ARM with a 11× speedup. Similar performance is achieved by the cluster of four Parallella boards against an Intel i5 3570 running a single thread. The Epiphany however, sees a drop in speed when attempting complex arithmetic operations compared with the other processors owing to the lack of hardware support. It is possible to achieve high performance using low-powered Parallella boards as long as the user is aware of the Epiphany chip’s weaknesses and avoids these. ACM Computing Classification System Classification •Computer systems organization → Multicore architectures; Heterogeneous (hybrid) systems; Interconnection architectures; •Networks→Network on chip; •Hardware → Impact on the environment;