Publication: Creación de un clúster de computación científica basado en FPGAs de bajo coste y consumo
Loading...
Official URL
Full text at PDC
Publication Date
2017
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
En este trabajo se presenta la construcción de un clúster basado en FPGAs de bajo consumo energético y coste, capaz de ejecutar programas de alta complejidad, en el mismo o en menor tiempo que una estación de trabajo de mucho mayor coste y consumo. En la actualidad ya existen clústeres de este tipo, pero lo que diferencia al nuestro es que se han utilizado placas con FPGAs de bajas prestaciones y que se ha utilizado OpenCL como lenguaje de programación para acelerar la ejecución de los programas. Estas placas son las DE1-SOC de Altera y se caracterizan, aparte de por su bajo coste y consumo, por ser capaces de ejecutar un sistema operativo de base UNIX/Linux en su hard-core, un procesador ARM Cortex-A9 de dos núcleos. Sin embargo, las imágenes de UNIX/Linux disponibles tanto oficiales como no oficiales, presentan problemas de configuración o limitaciones. Debido a esto, se ha generado una imagen personalizada basada en Debian 8 y se ha instalado en ella el software necesario para poder ejecutar códigos escritos en OpenCL y compilados con el Kit de desarrollo de software de Intel para FPGAs. Se ha elegido esta distribución por ser muy utilizada, robusta y actualizada.
Además, se ha realizado una comparativa de los tiempos de ejecución, coste y consumo energético resultado de ejecutar un conjunto de 5 benchmarks, que hemos implementado en C y OpenCL, entre el clúster y una estación de trabajo o Workstation de altas prestaciones. Aunque en algunos casos los tiempos de ejecución de la Workstation han sido menores que los del clúster, el bajo consumo y coste de este último hace que su eficiencia energética sea mucho mejor que la de la Workstation y, por lo tanto, que sea una mejor opción.
This MSc Thesis presents the creation of a cluster based on low power FPGAs, capable of executing high complexity programs, in the same or in a shorter time than a Workstation of much greater cost and power consumption. Currently there are clusters of this type, but what set us apart is that we have used boards with low-end FPGAs in combination with OpenCL as a programming language to accelerate the execution of the programs. These boards are the Altera DE1-SOCs and besides their low power and cost they are characterized for being able to run a UNIX/Linux operating system on their hard- core, a dual core ARM Cortex A9 processor. However, the official and unofficial available UNIX/Linux images possess configuration problems or limitations. Because of this, a customized image based on Debian 8 has been generated and the necessary software has been installed on it to run codes written in OpenCL and compiled with the Intel Software Development Kit for FPGAs. This UNIX/Linux distribution has been chosen because it is very used, robust and updated. Furthermore, we have compared the execution times, power consumption and costs between the cluster and a Workstation, by running 5 benchmarks that we have implemented in C and OpenCL. Although in some cases the Workstation’s execution times have been lower than the cluster's, the low power consumption and cost of the cluster makes it more efficient and, therefore, a better option.
This MSc Thesis presents the creation of a cluster based on low power FPGAs, capable of executing high complexity programs, in the same or in a shorter time than a Workstation of much greater cost and power consumption. Currently there are clusters of this type, but what set us apart is that we have used boards with low-end FPGAs in combination with OpenCL as a programming language to accelerate the execution of the programs. These boards are the Altera DE1-SOCs and besides their low power and cost they are characterized for being able to run a UNIX/Linux operating system on their hard- core, a dual core ARM Cortex A9 processor. However, the official and unofficial available UNIX/Linux images possess configuration problems or limitations. Because of this, a customized image based on Debian 8 has been generated and the necessary software has been installed on it to run codes written in OpenCL and compiled with the Intel Software Development Kit for FPGAs. This UNIX/Linux distribution has been chosen because it is very used, robust and updated. Furthermore, we have compared the execution times, power consumption and costs between the cluster and a Workstation, by running 5 benchmarks that we have implemented in C and OpenCL. Although in some cases the Workstation’s execution times have been lower than the cluster's, the low power consumption and cost of the cluster makes it more efficient and, therefore, a better option.
Description
Máster en Ingeniería Informática, Facultad de Informática, Departamento de Arquitectura de Computadoras y Automática, curso 2016-2017