Detecting performance anomaly with correlation analysis for Internetware  

作　　者：WANG Tao WEI Jun QIN Feng ZHANG WenBo ZHONG Hua HUANG Tao 

机构地区：[1]State Key Laboratory of Computer Science [2]Institute of Software,Chinese Academy of Sciences [3]University of Chinese Academy of Sciences [4]The Ohio State University

出　　处：《Science China(Information Sciences)》2013年第8期50-64,共15页中国科学（信息科学）（英文版）

基　　金：supported by National Grand Fundamental Research Program of China (Grant No. 2009CB320704);National High-tech R&D Program of China (863 Program) (Grant No. 2012AA011204);National Natural Science Foundation of China (Grant No. 61173004)

摘　　要：Internetware has become an emerging software paradigm to provide Internet services. The perfor- mance anomaly of Internetware services not only affects user experience, but also causes severe economic loss to service providers. Diagnosing performance anomalies has become one of the keys to improving the quality of service （QoS） of Internetware. Existing approaches create a system model to predict performance. Then, the prediction from the model is compared with the observation; a significant deviation may signal the occur- rence of a performance anomaly. However, these approaches require domain knowledge and parameterization efforts. Moreover, dynamic workloads affect the accuracy of performance prediction. To address these issues, we propose a correlation analysis based approach to detecting the performance anomaly for Internetware. We use kernel canonical correlation analysis （KCCA） to model the correlation between workloads and performance based on monitoring data. Furthermore, we detect anomalous correlation coefficients by XmR control charts, which detect the anomalous coefficient and trend without a priori knowledge. Finally, we adopt a feature se- lection method （Relief） to locate the anomalous metrics. We evaluated our approach on a testbed running the TPC-W industry-standard benchmark. The experimental results show that our approach is able to capture the performance anomaly, and locate the metrics relating to the cause of anomaly.Internetware has become an emerging software paradigm to provide Internet services. The perfor- mance anomaly of Internetware services not only affects user experience, but also causes severe economic loss to service providers. Diagnosing performance anomalies has become one of the keys to improving the quality of service （QoS） of Internetware. Existing approaches create a system model to predict performance. Then, the prediction from the model is compared with the observation; a significant deviation may signal the occur- rence of a performance anomaly. However, these approaches require domain knowledge and parameterization efforts. Moreover, dynamic workloads affect the accuracy of performance prediction. To address these issues, we propose a correlation analysis based approach to detecting the performance anomaly for Internetware. We use kernel canonical correlation analysis （KCCA） to model the correlation between workloads and performance based on monitoring data. Furthermore, we detect anomalous correlation coefficients by XmR control charts, which detect the anomalous coefficient and trend without a priori knowledge. Finally, we adopt a feature se- lection method （Relief） to locate the anomalous metrics. We evaluated our approach on a testbed running the TPC-W industry-standard benchmark. The experimental results show that our approach is able to capture the performance anomaly, and locate the metrics relating to the cause of anomaly.

关 键 词：performance anomaly anomaly detection INTERNETWARE system metrics kernel canonical correla-tion analysis 

分 类 号：TP393.09[自动化与计算机技术—计算机应用技术]