Performance Benchmarking a Raspberry Pi Cluster

08/01/2014 - 9:45am
Mesa Lab Main Seminar Room

Justin Moore portrait

Justin Moore, CISL Summer Extern
(Salish Kootenai College)

The Raspberry Pi computer offers a low-cost, entry-level Linux platform for education and entertainment purposes. Combining several Raspberry Pi computers together provides students access to a small Linux cluster with a parallel computing environment analogous in many ways to the largest supercomputers in the world today, again at very low cost. Research was conducted over the summer with the objective to benchmark the performance of a cluster of four Raspberry Pi computers using the standard High Performance LINPACK (HPL) benchmark, calling both the generic Basic Linear Algebra Subroutines (BLAS) and the Automatically Tuned Linear Algebra Software (ATLAS) library for performing matrix-matrix multiplies. The results were also compared to a hand-coded parallel version developed by me to determine the maximum achievable GFLOPS on the Raspberry Pi cluster via these different approaches. I observed that the performance of the Raspberry Pi cluster was dependent on parameters like clock speed, available memory, problem size, and the cache-blocking factor. The results were used to compare the Raspberry Pi cluster’s FLOPS/Watt and FLOPS/$ performance metrics with the Yellowstone supercomputer.

Video Presentation