Narrowing down possible causes of performance anomaly in web applications

Performance anomalies are a critical problem for commercial web applications. To help administrators discover their cause, many monitoring/analyzing systems have been developed, but their practical use is hindered by their limited scope and/or significant overheads; we cannot attach them all the time to the entire application. In this paper, we propose a front-end narrower that enables us to narrow down possible causes of performance anomalies. Our narrower is much less obtrusive but provides only hints about the cause of performance anomalies. Using control charts on response times of each request type, our "front-end" narrower indicates suspicious components or processes that may contain the root causes of the anomaly. After that, the administrator chooses an appropriate tool as a "back-end" and attaches it to the suspicious components or processes to trace down the original cause. The results of our case studies are encouraging: we succeeded in correcting some performance anomalies (including an unknown one) in default RUBiS.