程序阶段性分析和阶段检测技术

发布时间：2018-06-04 09:42

本文选题：程序分析 + 程序阶段性　；参考：《计算机科学》2015年01期

【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。
[Abstract]:The rapid development of symmetric multiprocessors and the proposed dynamic heterogeneous processors (DHMPs) in recent years provide a new opportunity for performance optimization. One opportunity is to find out the performance bottleneck of each phase of the program. A static program phase analysis method is proposed, that is, by analyzing the structural parameters and calculating the similarity matrix to find out the resource bottleneck in each phase of the program. Another opportunity is to give the time node of dynamic heterogeneous processor reconfiguration. Two dynamic phase detection algorithms, DPDA and HTPD, are proposed. It is found that the phase change can provide a reconfigurable time node for dynamic reconfigurable processor. DPDA algorithm is very effective and the cost of hardware and software implementation is low. However, HTPD algorithm is the first to use statistical method to perform dynamic detection phase so far. Experiments show that compared with BBV, DPDA and HTPD can avoid BBV off-line, dynamic algorithm needs to add extra hardware, and the result is related to compiler, etc. The stability and accuracy of phase partitioning are comparable to those of BBV. DPDA and HTPD algorithms can be directly used in mainstream processors and dynamic heterogeneous processors because they do not rely on additional hardware.
【作者单位】：中国科学技术大学计算机科学与技术学院;
【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
【分类号】：TP332

【相似文献】