Desarrollo de un planificador para la plataforma de cálculo distribuido T-Arenal
Fecha
2018-10-09
Autores
García González, Luis Antonio
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Central “Marta Abreu” de Las Villas
Resumen
En el presente trabajo se discuten diferentes técnicas de planificación de tareas para sistemas distribuidos en ambientes no dedicados así como la adaptación de estas a la plataforma de calculo distribuido T-arenal. Esta plataforma fue diseñada para ejecutar tareas complejas en cuanto al tiempo, sobre ordenadores desplegados en los laboratorios docentes y conectados por una red local. Inicialmente se describe el paradigma de computación voluntaria así como plataformas que lo implementan. Acerca de estas plataformas se mencionan sus características fundamentales y las aplicaciones que se pueden ejecutar en ellas. También se mencionan los algoritmos de planificación de tareas self-scheduling, estrategias desarrolladas inicialmente para la planificación de iteraciones de ciclos paralelos en sistemas homogéneos, donde las iteraciones son independientes entre ellas. Las principales estrategias self-cheduling son implementadas sobre T-arenal, utilizando el algoritmo LINPACK para estimar la capacidad computacional de las estaciones de trabajo y poder determinar la cantidad de tareas a enviar en cada pedido. Finalmente se realizaron pruebas de rendimiento ejecutando dos aplicaciones sobre cada una de las estrategias de planificación implementadas sobre la plataforma. Las pruebas realizadas muestran como los nuevos planificadores reducen el tiempo de ejecución y la cantidad de comunicaciones necesarias para completar las tareas en la plataforma.
The present research discusses different task scheduling techniques for distributed systems in non-dedicated environments as well as their adaptation to distributed platform T-arenal. This platform was designed to execute complex tasks in terms of time on computers deployed in educational laboratories and connected by a local network. Initially, the paradigm of volunteer computing is described as well as the platforms that implement it. About these platforms are mentioned their fundamental characteristics as well as the applications that can be executed in them. Also, the dynamic strategies self-scheduling to allocate a set of tasks are mentioned, strategies initially developed for the scheduling of parallel cycle iterations in homogeneous systems, where iterations are independent of each other. The main self-cheduling strategies are implemented on T-arenal, using LINPACK benchmark to estimate the computational capacity of the workstations and to determine the amount of tasks to send in each order. Finally, performance tests of two applications were performed on each of the scheduling strategies implemented on the platform. The tests show how the new schedulers reduce the execution time and the amount of communications needed to complete the tasks on the platform.
The present research discusses different task scheduling techniques for distributed systems in non-dedicated environments as well as their adaptation to distributed platform T-arenal. This platform was designed to execute complex tasks in terms of time on computers deployed in educational laboratories and connected by a local network. Initially, the paradigm of volunteer computing is described as well as the platforms that implement it. About these platforms are mentioned their fundamental characteristics as well as the applications that can be executed in them. Also, the dynamic strategies self-scheduling to allocate a set of tasks are mentioned, strategies initially developed for the scheduling of parallel cycle iterations in homogeneous systems, where iterations are independent of each other. The main self-cheduling strategies are implemented on T-arenal, using LINPACK benchmark to estimate the computational capacity of the workstations and to determine the amount of tasks to send in each order. Finally, performance tests of two applications were performed on each of the scheduling strategies implemented on the platform. The tests show how the new schedulers reduce the execution time and the amount of communications needed to complete the tasks on the platform.
Descripción
Palabras clave
Planificador, Cálculo Distribuido, Computación Voluntaria