基于SaaS软件运行日志分析的软件性能问题的在线识别和诊断方法

在运维阶段识别和诊断SaaS软件性能问题时，现有的记录日志机制缺乏有效性，日志数据缺乏精确性，日志数据分析缺乏科学性。针对上述问题，本项目提出一种支持SaaS软件性能分析的日志记录方法，实现性能相关日志信息的获取、规范表达和高效存取。提出一种对日志数据对象进行聚类分析的谱聚类方法，以适应日志数据对象在数据空间中几何形状的多样性。在谱聚类方法中还针对日志数据的分布特点，采用基于K近邻局部密度的方法构造谱聚类算法中用到的相似矩阵,以体现日志数据在性能问题分析方面的相似性。通过这种谱聚类方法可以从日志大数据中，有效地把与性能问题相关的日志数据找出来，并通过进一步分析这些数据就可以识别和诊断出性能问题。本项目研究的日志数据谱聚类方法在分析诊断软件性能问题研究方面具有一定的创新性。本项目研究成果将为SaaS软件的运维管理，提供一种实时高效和可伸缩的软件性能问题在线识别和诊断方法，具有重要学术意义。

The existing log recording mechanism faces three critical challenges when it is used to identify and diagnose the performance problem of SaaS software: the ineffective mechanism, the inaccurate log data and the ad hoc methodology for log analysis. In this proposal, we hope to develop a reasonable mechanism for log data analysis to support the performance analysis of SaaS software, including the effective retrieval of performance related log data, the representation of the specification and efficient data storage and access. Based on the spectral clustering methodology, we intensively address the following aspects: 1) the tailored spectral clustering algorithm by considering the shape diversity problem of the log data; 2) K-NN density estimation based algorithm to build the affinity matrix; 3) an effective retrieval and analysis model to identify the relevant log records from the clustering results; 4) a diagnosis algorithm to associate the underlying performance issues with the extracted relevant log record. The main novelty is that we exploit the possibility of using spectral clustering techniques over log data for software performance analysis and diagnosis. This proposal will make significant contributions to the development of a real-time effective and flexible framework of online performance problems identification and diagnosis for SaaS software maintenance.