嵌入式应用中的循环级线程推测并行性分析  

Profiling loop-level speculative parallelism in embedded applications

作　　者：卜得庆 王耀彬[1,2] 李凌 杨洋[3] 程一鸣 刘志勤[1,2] 吴亚东 Bu Deqing;Wang Yaobin;Li Ling;Yang Yang;Cheng Yiming;Liu Zhiqin;Wu Yadong(School of Computer Science & Technology,Southwest University of Science & Technology,Mianyang Sichuan 621010,China;Sichuan Civil-military Integration Institute,Southwest University of Science & Technology,Mianyang Sichuan 621010,China;Sichuan Institute of Computer Sciences,Chengdu 610041 ,China)

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621010 [2]西南科技大学四川省军民融合研究院,四川绵阳621010 [3]四川省计算机研究院,成都610041

出　　处：《计算机应用研究》2019年第9期2691-2695,共5页Application Research of Computers

基　　金：国家自然科学基金面上项目(61672438);国家留学基金委资助项目(CSC201908510040);四川省军民融合研究院开放基金资助项目(2017SCII0213);四川省科技计划资助项目(2019YJ0326);四川省教育厅研究项目(18ZB0603);西南科技大学科研项目(17xn0045,17xn0038,17lzx621,13zx7101,17lzxt10);西南科技大学研究生创新基金资助项目(19ycx0051)

摘　　要：如何有效利用多核提供的丰富晶体管资源对串行程序的执行进行加速是当前研究中的热点问题。线程级推测(thread-level speculation,TLS)技术旨在充分利用多核资源,最大化地开发出串行代码中存在的潜在并行性。目前TLS技术已经在多种串行应用的并行化工作中得到有效利用,但嵌入式应用程序仍未在推测并行化方面进行有效的分析。因此,选取了八个具有代表性的嵌入式应用,对其在循环级推测并行化中的性能提升潜力和运行时特征(数据依赖、线程粒度和并行覆盖率)进行探讨。实验结果表明,利用线程级推测并行化嵌入式应用的加速效果优于指令级并行技术,实验中的最大加速比达到了13. 29;在嵌入式应用领域,该技术可以有效地利用4~8核的计算资源。How to effectively utilize the rich transistor resources provided by multi-core to accelerate the execution of serial programs is currently a hot issue in research.Thread-level speculation(TLS) technology aims at making full use of multi-core resources and maximizing the potentially parallelism in serial code.At present,TLS technology has been effectively utilized to parallelize several serial applications.However,speculative thread level parallelism in embedded applications has not yet been explored thoroughly.Therefore,this paper analyzed eight selected representative embedded applications from their potential parallelism and runtime characteristics(data dependency,thread granularity and parallel coverage) in loop-level speculation.The experimental results show that for embedded applications,the speculative thread level parallelism is better than that in instruction level parallel technology,and the maximum speed up in experiment achieves 13.29.In the field of embedded applications,the technology can effectively utilize resources of 4 to 8 core computing.

关 键 词：线程级推测 多核 嵌入式应用 数据依赖 

分 类 号：TP302.7[自动化与计算机技术—计算机系统结构]