程序阶段性分析和阶段检测技术  

Program Phase Analysis and Phase Detection Techniques

作　　者：张海博[1] 安虹[1] 贺松涛 孙涛[1] 王涛[1] 彭毅[1] 程亦超 

机构地区：[1]中国科学技术大学计算机科学与技术学院,合肥230027

出　　处：《计算机科学》2015年第1期71-74,共4页Computer Science

基　　金：国家自然科学基金(60970023);国家973计划项目(2011CB302501);国家863计划项目(2012AA010902;2012AA010901)资助

摘　　要：对称多处理器的飞速发展和近年来提出的动态异构处理器(DHMP)为性能优化提供了新的机遇。一个机遇是找出程序每个阶段的性能瓶颈,提出了静态程序阶段分析方法,即通过分析结构参数和计算相似度矩阵来找出程序每个阶段的资源瓶颈;另一个机遇是给出动态异构处理器重构的时间节点,提出了DPDA和HTPD两种动态阶段检测算法,检测出阶段的变化能够为动态可重构处理器提供重构的时间节点。DPDA算法效果很好且软硬件实现代价小,而HTPD算法是目前为止第一个使用统计学方法进行动态检测阶段的算法。实验表明,与BBV相比,DPDA和HTPD能避免BBV离线、动态算法需添加额外硬件、结果与编译器相关等限制,并且阶段划分的稳定性和正确率与BBV相当。DPDA和HTPD算法由于本身不依赖额外硬件,因此都能直接在主流处理器和动态异构处理器(DHMP)中使用。The rapid development of SMP and new proposed DHMP bring new challenges for program performance optimization.We raised two performance tuning problems and the solutions were given by phase analysis.The first problem is to find theperformance bottlenecks in each phase.We proposed a static phase analysis method,which finds performance bottlenecks in each phases by analyzing architecture features and its similar matrix.The second problem is to give the proper time to reconfigure for DHMP.We proposed dynamic phase detection algorithms,namely DPDA and HTPA.DPDA archives effective performance in a relative low software/hardware cost,and HTPD is the first phase detection algorithm using statistics theory.Our results show that comparing with BBV,DPDA and HTPD can avoid its limitation of offline,plus additional hardware in online algorithm and compiler＇s effect,while they offer a comparable stability and correctness.Since DPDA and HTPD do not relay on additional hardware support,they can be implemented directly in mainstream processors and DHMP.

关 键 词：程序分析 程序阶段性 静态程序分析 阶段检测 

分 类 号：TP368.1[自动化与计算机技术—计算机系统结构]