Online Client-Side Bottleneck Identification on HTTP Server Infrastructures

摘要

Ensuring short response times is a major concern for all web site administrators. To keep these times under control, they usually resort to monitoring tools that collect a large spectrum of system metrics, such as CPU and memory occupation, network traffic, number of processes, etc. Despite providing a reasonably accurate picture of the server, the times that really matter are those experienced by the user. However, not surprisingly, system administrators will usually not have access to these end-to-end figures, due to their lack of control over web browsers. To overcome this problem, we follow the opposite approach of monitoring a site based on times collected from browsers. We use two browser-side metrics for this: i) the time it takes for the first byte of the response to reach the user (request time) and ii) the time it takes for the entire response to arrive (response time). We conjecture that an appropriate choice of the resources to control, more precisely, one or two URLs, suffices to detect CPU, network and I/O bottlenecks. In support of this conjecture, we run periodical evaluations of request and response times on some very popular web sites to detect bottlenecks. Our experiments suggest that collecting data from the browsers can indeed contribute for better monitoring tools that provide a deeper understanding of the system, thus helping to maintain faster, more interactive web sites.