Escalonamento adaptativo de tarefas em arquiteturas híbridas

Abstract

The constant growing volume of data coupled with the need for more processingpower in differents areas of knowledge, has stimulated the emergence of new computing architectures. We witnessed the popularization of the use of efficient coprocessors, such as GPUs and MICs for machines equipped with powerful multicore processors, constituting hybrid architectures. In this new context, it becomes necessary applications of differents scenarios are able to explore coordinated and efficient all processing units (processors andcoprocessors) available in a hybrid architecture, taking advantagethe most of their processing capabilities. Runtime environments have been proposed in order to exploit these resource as much as possible by offering a variety of methods for dynamically scheduling tasks on different process units (PUs). These schedulers determine which PU is better suited for executing each task, based on caracteristics of tasks and avaiable resources. In our study, therefore, present astudy on dynamic task scheduling in hybrid architectures,proposing new strategies for efficient scheduling and exploring differentapplication scenarios and hybrid systems with various processing units.