一种基于深度学习的性能分析框架设计与实现

高性能计算系统的体系结构日益复杂和现有性能分析工具的智能程度不足，导致高性能计算应用的程序性能分析和优化的成本代价日益高昂。所幸，人工智能领域目前取得了重要进展，其中深度学习技术发挥了重要作用，它给性能分析工具的智能化带来了契机。提出一种基于深度学习的程序性能智能分析框架，其核心思想是将程序的性能分析问题抽象成可用机器学习技术描述的分类问题，使用处理器支持的PMU采集分类所需的性能数据并标准化，使用簇评估技术结合簇的实际含义确定性能问题类别，通过稀疏编码自动学习性能数据特征并构建性能问题分类模型。在神威太湖之光超级计算机上实现了程序性能分析框架原型。实验结果表明，该性能分析方法能够直观地指导程序员快速把握当前应用最为突出的性能瓶颈问题，提高应用优化的效率，降低用户调优代码的成本。

A deep learning based program performance analysis framework

The increasing complexity of high performance computer system architecture and inadequate intelligence of the existing performance analysis tools, lead to the increasing cost for performance analysis and optimization of high performance computing applications. Thanks to the deep learning technology, significant progress in the field of artificial intelligence has been made, and it also presents an opportunity for intelligent program performance analysis tools. We propose an intelligent program performance analysis framework, which abstracts program performance analysis into a classification problem via machine learning technology. The performance data, the input of the analysis framework, which is collected via the performance monitoring unit, is standardized in advance; and the categories of performance problems, the output of the analysis framework, are determined by the cluster evaluation and the actual meaning of every cluster. Sparse coding is used to automatically learn the features of performance data and a performance problem classification model is built. Finally, we implement the prototype of the framework on the Sunway TaihuLight system. Experimental results show that this framework can intuitively guide programmers to grasp the most prominent performance bottleneck problem in current applications, effectively improve the performance analysis efficiency, and reduce the cost of code tuning.