Performance Evaluation of the HimMUC Cluster

Abstract

[EN] In a time when the quantity of information that requires to be collected, stored and processed keeps growing, a niche has appeared for highperformance computers that are affordable to build, use, and maintain. Single-board computers are a cost-effective and energy-efficient means of reaching relatively high computational performances, and their popularity is growing in many areas of application. One of these applications is computer clusters, given the potentially high degree of parallelism that can be achieved at a low cost. This thesis deals with an empirical analysis of the performance of the HimMUC cluster, a computer cluster built out of said devices, in order to quantify just how scalable and overall promising such a solution can be. This has been measured by the means of the HPL benchmark, which also served as a reference for the renowned Top500 list. As a result of an in-depth tuning process, the maximum performance reached with 40 nodes was approximately 231.377 GFLOPS, about 40% of the theoretical peak. Better results could be reached with low-level optimizations and especially by reducing the bottlenecks that currently limit performance, namely the total available RAM and the interconnection network. The results confirm our initial hypothesis and prove that there is, in fact, potential for single-board computers in the field of HPC, at least for less demanding workloads, and that there is still room for future growth given the rapid evolution of this technology.

[ES] En el ámbito de la Computación de Altas Prestaciones y según el Top500, la gran mayoría de los supercomputadores relevantes actualmente son clústeres. Por otro lado, los computadores de placa reducida (SBC) son utilizados para todo tipo de aplicaciones en un amplio abanico de campos, incluyendo el tan relevante actualmente Internet de las Cosas. Gracias a la miniaturización y el cumplimiento de la Ley de Moore hasta ahora, es posible sostener el rendimiento de los supercomputadores del pasado en la palma de una mano. El Clúster HimMUC de la Cátedra de Arquitectura de Computadores y Sistemas Paralelos del Departamento de Informática de la Universidad Técnica de Munich está compuesto por este tipo de dispositivos: 40 Raspberry Pi 3 y 40 ODroid-C2. Este Trabajo de Fin de Grado explora el proceso de evaluación y optimización del rendimiento de dicho clúster, y analiza su escalabilidad y viabilidad general, teniendo en cuenta los costes tecnológicos asociados.