There are a number of limitations to our EtE monitor architecture. Since EtE monitor extracts HTTP transactions by reconstructing TCP connections from captured network packets, it is unable to obtain HTTP information from encrypted connections. Thus, EtE monitor is not appropriate for sites that encrypt much of their data (e.g., via SSL).
In principle, EtE monitor must capture all traffic entering and exiting a particular site. Thus, our software must typically run on a single web server or a web server cluster with a single entry/exit point where EtE monitor can capture all traffic for this site. If the site ``outsources'' most of its popular content to CDN-based solutions then EtE monitor can only provide the measurement information about the ``rest'' of the content, which is delivered from the original site. For sites using CDN-based solutions, the active probing or page instrumentation techniques are more appropriate solutions to measure the site performance. A similar limitation applies to pages with ``mixed'' content: if a portion of a page (e.g., an embedded image) is served from a remote site, then EtE monitor cannot identify this portion of the page and cannot provide corresponding measurements. In this case, EtE monitor consistently identifies the portion of the page that is stored at the local site, and provides the corresponding measurements and statistics. In many cases, such information is still useful for understanding the performance characteristics of the local site.
The EtE monitor does not capture DNS lookup times. Only active probing techniques are capable of measuring this portion of the response times. Further, for clients behind proxies, EtE monitor can only measure the response times to the proxies instead of to the actual clients.
As discussed in Section 3, the heuristic we use to reconstruct page content may determine incorrect page composition. Although the statistics of access patterns can filter invalid accesses, it works best when the sample size is large enough.
Dynamically generated web pages introduce another issue with our statistical methods. In some cases, there is no consistent content template for a dynamic web page if each access consists of different embedded objects (for example, some pages use a rotated set of images or are personalized for client profiles). In this case, there is a danger that metrics such as the server file hit ratio and the server byte hit ratio introduced in Section 6 may be inaccurate. However, the end-to-end time will be computed correctly for such accesses.
There is an additional problem (typical for server access log analysis of e-commerce sites) about how to aggregate and report the measurement results for dynamic sites where most page accesses are determined by URLs with client customized parameters. For example, an e-commerce site could add some client specific parameters to the end of a common URL path. Thus, each access to this logically same URL has a different URL expression. However, service providers may be able to provide the policy to generate these URLs. With the help of the policy description, EtE monitor is still able to aggregate these URLs and measure server performance.