一种应用代价评估的推测多线程路径预测方法  

A Cost Estimation Based Speculative Path Prediction Method for Speculative Multithreading

作　　者：李远成[1] 赵银亮[1] 阴培培[1] 韩博[1] 

机构地区：[1]西安交通大学计算机科学与技术系,西安710049

出　　处：《西安交通大学学报》2010年第12期22-27,共6页Journal of Xi'an Jiaotong University

基　　金：国家高技术研究发展计划资助项目(2008AA01Z136)

摘　　要：推测多线程技术对于自动并行化非规则程序是有效的,然而基于控制流图和分支预测方法的线程划分方法,不可避免地受到划分路径上存在的控制依赖和数据依赖制约.针对现有的路径预测方法在考虑控制依赖影响的同时却不能有效地综合考虑数据依赖影响的问题,提出一种新的基于代价评估的路径预测方法,通过引入数据依赖模型,综合评估控制和数据依赖两种影响因素,寻求一条具有近似最小推测开销的推测划分路径.实验结果表明,文中提出的路径预测方法能够计算出代价更小的推测划分路径,并取得了更好的加速比性能,总体上系统可以得到2.43%的加速比性能提升.Speculative multithreading（SpMT） technology is an effective mechanism for automatic parallelization of irregular programs.However,just generating speculative threads based on the control flow graph which only contains branch probability information,it is inevitable that there may be excessive constraints resulting from control and data dependence in practice.Therefore,it is very important to understand the trade-offs between different speculative paths.In this paper,by introducing the data dependence model and discussing the trade-offs between different speculative paths,we propose a novel cost estimation based speculative path prediction method which comprehensively takes account of control-and data-dependence.By this method,we attempt to seek a speculative path which has the minimum cost overhead.The experimental results show that there are interesting trade-offs between different speculative paths and we can indeed get better performance.On average,we achieve 2.43% performance improvement.

关 键 词：推测多线程 代价评估模型 路径预测技术 数据依赖模型 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]