Abstract: Internet workload is dominated by Web traffic today, which demonstrates its perceived performance by some well-known metrics, such as response time and download rate. The popular approaches to measuring HTTP performance are Web server and client logs, or traffic traces. The performance differences are compared to improve protocol design and implementation by building corresponding pattern to various parameters from different levels existing behind Web browsing, such as the number of connections per page, the connection sizes, etc., or by simulating HTTP processes under different workload or version. But the customers' urgent demand is how to detect service response time and Web page download rate, which provide impactful knowledge to pinpoint the source of HTTP performance problems, and how to detect performance outliers and alarm timely. The performance of HTTP service along the end-to-end path is obtained through the method of active measurement. Furthermore, with the experimental data the composing and distribution of HTTP service response time are analyzed. And the high variation of the download rate is found in multiple accesses to the same Web page, which proves that some popular approaches to performance fault detection, such as mean and standard deviation calculation, do not fully serve the turn. Therefore a better algorithm based on box plots to detect performance outliers of HTTP service is presented.