体系结构级GPU功耗建模及软件低功耗优化方法研究

项目名称： 体系结构级GPU功耗建模及软件低功耗优化方法研究

项目编号： No.61300029

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 王卓薇

作者单位： 广东工业大学

项目金额： 26万元

中文摘要： 随着处理器功耗不断增大，功耗问题逐渐成为高性能计算机系统设计与实现的首要问题。低功耗优化技术将提升系统性能的同时防止功耗过快增长，并提高系统的可靠性及可用性。 本课题开展体系结构级GPU功耗模型及软件低功耗优化方法研究，面向CPU-GPU异构并行系统，以降低系统能耗，提升系统效能为目标，拟重点研究计算与存储划分的动态功耗模型；研究基于温度实时感知的静态功耗模型；并在该模型基础上，研究面向程序执行特征的关键路径查询动态功耗优化方法；研究基于任务均衡划分的多维GPU静态功耗优化方法；设计并实现功耗模型及功耗优化方法验证系统。 本课题将在体系结构级GPU功耗建模、基于关键路径分析的动态功耗优化算法、基于自适应任务划分的静态功耗优化算法，以及功耗模型及功耗优化验证理论与方法等方面形成创新性成果，为GPU低功耗优化研究提供理论依据与技术支撑。

中文关键词： 图形处理器；体系结构级功耗建模；软件低功耗；关键路径查询；任务均衡划分

英文摘要： As the processor's power consumption continuously increases, power consumption has become the most critical problem during the design and implementation of high performance computer(HPC)system. Low power optimization technology will enhance the system performance while preventing the excessive growth of power consumption, and improve system reliability and availability. This project proposes to carry out the researches on architecture-level power model and low power software optimizations for CPU-GPU heterogeneous parallel system. We are committed to lower the system power and improve the power efficiency. This project intends to focus on the key technologies as follows: the dynamic power model based on division of computation and memory, the static power model based on real-time temperature perception, the dynamic power optimization method for program execution characteristics based on efficient critical path analysis, the static power optimization method for multi-GPU environment based on task distribution, as well as the verification system of the power model and optimization, etc. The primary intellectual merit of the project includes the architectural-level GPU power model, dynamic power optimization algorithm based on critical path analysis, static power optimization algorithm based on self-adaption task p

英文关键词： Graphics processing unit；Architecture-level power model；Low power software；Critical path querying；Task balance scheduling