Publication: Evaluación y optimización de rendimiento y consumo energético de aplicaciones paralelas a nivel de tareas sobre arquitecturas asimétricas
Loading...
Official URL
Full text at PDC
Publication Date
2016
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Las arquitecturas asimétricas, formadas por varios procesadores con el mismo repertorio de instrucciones pero distintas características de rendimiento y consumo, ofrecen muchas posibilidades de optimización del rendimiento y/o el consumo en la ejecución de aplicaciones paralelas. La planificación de tareas sobre dichas arquitecturas de forma que se aprovechen de manera eficiente los distintos recursos, es muy compleja y se suele abordar utilizando modelos de programación paralelos, que permiten al programador especificar el paralelismo de las tareas, y entornos de ejecución que explotan dinámicamente dicho paralelismo. En este trabajo hemos modificado uno de los planificadores de tareas más utilizados en la actualidad para intentar aprovechar todos los recursos al máximo, cuando el rendimiento así lo necesite, o para conseguir la mejor eficiencia energética posible, cuando el consumo sea más prioritario. También se ha utilizado una biblioteca desarrollada específicamente para la arquitectura asimétrica objeto de estudio en la Universidad de Texas, Austin.
Para obtener el máximo rendimiento se han agrupado los núcleos del sistema en dos niveles: hay un cluster simétrico de núcleos virtuales idénticos, cada uno de los cuales está compuesto por un conjunto de núcleos asimétricos. El planificador de tareas asigna trabajo a los núcleos virtuales, de manera idéntica a como lo haría en un sistema multinúcleo simétrico, y la biblioteca se encarga de repartir el trabajo entre los núcleos asimétricos. El trabajo ha consistido en integrar dicha biblioteca con el planificador de tareas. Para mejorar la eficiencia energética se han incluido en el planificador de tareas políticas de explotación de los modos de bajo consumo de la arquitectura y también de apagado o no asignación de carga de trabajo a algunos de los núcleos, que se activan en tiempo de ejecución cuando se detecta que la aplicación no necesita todos los recursos disponibles en la arquitectura.
Asymmetric architectures, composed by multiple processors with the same instruction set but different performance and energy consumption characteristics, provide many optimisation possibilities in performance and/or energy consumption over the execution of parallel applications. Scheduling tasks in this kind of architectures in such a way that all the resources are used efficiently is a difficult task, and it is usually addressed using parallel programming models, allowing programmers to annotate the parallelism for each of the different tasks of the program, and runtime environments which dinamically take advantage of this kind of parallelism. In this work we have modified one of the most used task schedulers nowadays to try to use all the resources efficiently if high performance is needed, or to achieve the best energy efficiency if consumption reduction is the priority. A math library developed specifically for this kind of asymmetric architectures in the University of Texas, Austin, has been used as well. To obtain maximum performance, cores have been grouped into two different levels: a symmetric cluster with identical virtual cores where each of them is composed by a set of asymmetric cores. With this approach, schedulers can allocate tasks to the virtual cores in the same way as they would do in a symmetric multicore system, and the library is responsible for the allocation of tasks to the asymmetric cores. The work carried out consisted of integrating this library with the task scheduler. To improve energy efficiency, new policies have been included into the scheduler. These policies have been designed to take advantage of the low consumption modes of the architecture and to disable or not assign tasks to some of the cores, activating them at run-time if the execution does not need all the available resources in the architecture.
Asymmetric architectures, composed by multiple processors with the same instruction set but different performance and energy consumption characteristics, provide many optimisation possibilities in performance and/or energy consumption over the execution of parallel applications. Scheduling tasks in this kind of architectures in such a way that all the resources are used efficiently is a difficult task, and it is usually addressed using parallel programming models, allowing programmers to annotate the parallelism for each of the different tasks of the program, and runtime environments which dinamically take advantage of this kind of parallelism. In this work we have modified one of the most used task schedulers nowadays to try to use all the resources efficiently if high performance is needed, or to achieve the best energy efficiency if consumption reduction is the priority. A math library developed specifically for this kind of asymmetric architectures in the University of Texas, Austin, has been used as well. To obtain maximum performance, cores have been grouped into two different levels: a symmetric cluster with identical virtual cores where each of them is composed by a set of asymmetric cores. With this approach, schedulers can allocate tasks to the virtual cores in the same way as they would do in a symmetric multicore system, and the library is responsible for the allocation of tasks to the asymmetric cores. The work carried out consisted of integrating this library with the task scheduler. To improve energy efficiency, new policies have been included into the scheduler. These policies have been designed to take advantage of the low consumption modes of the architecture and to disable or not assign tasks to some of the cores, activating them at run-time if the execution does not need all the available resources in the architecture.
Description
Máster en Ingeniería Informática, Facultad de Informática, Departamento de Arquitectura de Computadores y Automática, curso 2015-2016