CPU-GPU系统中基于剖分的全局性能优化方法  

Profiling Based Optimization Method for CPU-GPU Heterogeneous Parallel Processing System

作　　者：张保[1] 董小社[1] 白秀秀[1] 曹海军[1] 刘超[1] 梅一多[1] 

机构地区：[1]西安交通大学电子与信息工程学院,西安710049

出　　处：《西安交通大学学报》2012年第2期17-23,共7页Journal of Xi'an Jiaotong University

基　　金：国家高技术研究发展计划资助项目(2009AA01A135;2009AA01Z108);中央高校基本科研业务费专项资金资助项目(08142007)

摘　　要：针对将应用移植到CPU-GPU异构并行系统上时优化策略各自分散、没有一个全局的指导思想的问题,提出了一种基于剖分的全局性能优化方法.该方法由优化策略库、剖分工具库和策略配置模块组成.优化策略库将应用移植到异构并行系统上的性能优化过程划分为访存级、内核加速级和数据划分级3级优化;针对3级优化剖分工具库提供了3级剖分机制,通过运行时的剖分技术获取剖分信息;策略配置模块根据所获取的信息指导用户在每级优化中选择合适的优化策略.实验证明,基于剖分的全局性能优化方法可以明确地指导将应用移植到CPU-GPU异构并行系统上的全局优化过程,利用该优化方法后,以矩阵相乘和傅里叶变换为例的应用性能提升明显,最终性能相对于访存级优化最高可提高30%左右.A profiling based optimization method for CPU-GPU heterogeneous parallel processing system is proposed to address the problem that the present optimization strategies get sectional thus failed to guide a global optimization.It is composed of the optimization strategy library,the profiling tool library,and the strategy deploy module,and the optimization strategy library divides the performance promotion process into a three-level optimization,including the memory-access level,the kernel-speedup level,and the data-partition level.The profiling tool library realizes three-level profiling mechanisms towards three-level optimizations to obtain application information,and the strategy deploy module guides users to choose an adaptive strategy with the information obtained by profiling tool library.Experimental results show that the proposed one is able to guide the optimization process of applications transplanted to heterogeneous parallel system.The performance for matrix multiplication and fast Fourier transform are improved obviously,and the final performance is heightened by 30% compared with the memory-level optimization.

关 键 词：CPU-GPU异构并行系统 全局优化 3级优化 3级剖分 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]