基于性能预测的推测多线程循环选择方法

发布时间：2018-04-18 05:21

本文选题：并行处理 + 线程级推测　；参考：《电子与信息学报》2014年11期

【摘要】：线程级推测(Thread-Level Speculation,TLS)是多核上一种加速串行程序的线程级自动并行化技术。循环具有规则的结构并在运行时占有大量的执行时间,因此循环是挖掘并行性的理想对象。然而,选择哪些循环并行才能提高程序的加速比是一个很难决定的问题。为了解决该问题,该文提出一种基于性能预测的循环选择方法。基于输入训练集获取程序预执行的剖析信息,同时结合各种推测因素,构建了循环结构的性能预测模型。预测结果定量评估了循环推测并行的加速比并决定该循环在运行时是否适合并行。实验结果表明,该文提出的方法能有效地预测循环并行时所蕴含的并行性,并依据预测结果准确地选择具有并行收益的循环推测并行,最终Olden基准测试集加速比性能平均提升了12.34%。
[Abstract]:Thread-Level SpeculationTLS (Thread-Level SpeculationTLS) is a thread-level automatic parallelization technique for accelerating serial programs on multiple cores.Loop has a regular structure and takes up a lot of execution time at run time, so loop is an ideal object for mining parallelism.However, it is difficult to determine which loop parallelism can improve the speedup of programs.In order to solve this problem, this paper presents a method of loop selection based on performance prediction.Based on the input training set, the analysis information of program pre-execution is obtained. At the same time, the performance prediction model of cyclic structure is constructed by combining various conjecture factors.The predicted results quantitatively evaluate the speedup of the loop conjecture parallelism and determine whether the loop is suitable for parallelism at run time.The experimental results show that the proposed method can effectively predict the parallelism of cyclic parallelism, and accurately select the circular conjecture parallelism with parallel income according to the prediction results.The final Olden benchmark set speedup performance improved by an average of 12.34.
【作者单位】：西安交通大学计算机科学与技术系;
【基金】：国家自然科学基金(61173040) 国家“863”计划项目(2012AA011003) 博士学科点专项科研基金(20130201110012)资助课题
【分类号】：TP332

【参考文献】