云环境下性能监测数据预处理方法研究

On the Preprocessing of PMU Data for Cloud Computing

基于性能监测单元(performance monitoring unit,PMU)的分析技术对深入理解程序运行特征十分重要,也是性能优化的重要依据.在云计算环境下,由于机器数目庞大,PMU起着更为重要的作用.然而性能计数器却存在监测效率低,数据质量不高的问题.从两方面对这一问题进行了研究:缺失值的填充和异常值的替换.对缺失值,采用回归方法来进行补全;对异常值,采用局部滤波的方法进行过滤.提出了基于知识库的数据预处理方法(KBDP)将监测效率提升了2-5倍,综合分析了使用PMU对程序执行造成的可能影响,处理后的结果与参考值的相似度高于80%,最高达95%.

PMU （Performance monitoring unit） based profiling is of great importance to understand the characteristics of program behaviors, and it is therefore utterly important for performance optimization. In cloud computing, PMU becomes even more important. However, the quality of the profiled is poor （e.g., missing values and outliers） when high profiling speed is needed （e.g., multiplexing is employed）. This paper proposes a novel approach to improve the PMU data quality. For the missing values, we build a regression model to fill in them. For the outliers, we construct a local filter to replace them. Finally, we proposed a Knowledge Based Data Preprocessing framework （KBDP） to integrate the proposed approaches together. Our experimental results show that the profiling speed is improved by a factor of 2x and up to 5x. In addition, the similarity between the processed data and the real measurements achieves 80% at least and up to 95%.