Estudo do impacto da implementação de um escalonador em uma arquitetura multiprocessada com núcleos heterogêneos

Abstract

Heterogeneous multiprocessor architectures are becoming an increasingly coveted study’s object due to its advantages compared to homogeneous multiprocessor architectures, such as better hardware utilization and lower on-chip system power consumption. Because of possessing cores with different processing capabilities, this architecture requires a methodology that enables threads migration as needed for resources, thereby, enabling the best available cores use without a loss of performance in the system. Through the reconfigurable architecture CREAMS utilization, which enables the simulation of heterogeneous cores, the work’s goal is given by the application study of this architecture’s intrinsic characteristics in migrating threads. So, three schedulers which are based on different characteristics have been developed. Since the system enables a run-time system data processing collection, the schedulers use this data as a metric for shifting tasks across heterogeneous cores during the course of the application. Unlike other proposed studies, this developed schedulers do not require a previous mapping to effect migration. Based on various simulations, the paper presents schedulers and test scenarios impact analysis on running applications performance. In relation to a scheduler that does not make migration use, there was obtained a performance improvement for 89% of the simulated cases. On average, considering all simulations, it was obtained an 18.41% superior performance to those found for a static scheduler. The attribute that evinced the best scheduling based on the results obtained was the instructions’s number executed in the reconfigurable array.