中文摘要：

针对由通用微处理器和专用加速部件构成的异构并行系统,提出结合通信感知的并行任务划分和动态电压频率调节技术的异构系统能耗优化方法,该方法旨在将并行任务图划分并映射在异构处理单元,在满足性能约束的条件下最小化系统能耗.在目前典型异构并行系统中,主处理器与加速部件大都通过系统总线连接,必然引入不可忽略的通信开销,因此通信感知的任务划分技术是该问题的关键.提出了基于整数线性规划的静态最优能耗优化方法和基于遗传算法的动态能耗优化方法.并通过一个典型科学计算应用验证了本文方法的有效性.

英文摘要：

Targeted for the heterogeneous parallel system consisting of general purpose processor and application-specific accelerator,this paper proposes an energy optimization method of coordinating communication-aware task partition and voltage scaling.The main ideal is partitioning a parallel task graph and mapping each task onto heterogeneous processing units to minimize the energy consumption within performance constraint.In current heterogeneous parallel systems,the host processor and accelerator are connected via system bus,which incurs in un-neglected overhead;therefore,communication-aware task partition method is the key problem.This paper proposes an ILP-based static optimal energy optimization method and the genetic algorithm based dynamic energy optimization method.Finally,we validate the efficiency of the proposed methods with a typical scientific application.