Check out the new USENIX Web site. next up previous
Next: Analysis results Up: Experimental methodology Previous: Measurement setup

Measurement impact

Because we propose to use our measurement infrastructure on a production Web site, it is important to evaluate its impact on the server performance and other aspects of its operation. The additional overhead our measurement technique imposes on Web client performance is the retrieval of the transparent image, including the HTTP redirect and extra DNS requests. Because the image is transparent, it does not visually affect the page. Furthermore, the image is small in size--43 bytes--which keeps the added delay to a minimum. We also encourage participants to include the image at the end of the HTML page containing it; therefore, browsers will normally request it last. Thus, the extra latency associated with the image is usually hidden from the user's Web browsing experience. Another advantage of the small size of the image is that when the image is not available for download, it does not affect the visual appearance of the Web page at all.

Our custom HTTP redirector is a single-threaded, non-blocking, 300-line C program. The redirector responds to all Web requests with a ``302 Moved Temporarily'' HTTP redirect to a URL with the client's IP address embedded in it. Due to the small size and overhead of the redirector, we found it to be highly reliable and more responsive than a standard Web server.

To validate the claim of a small increase in latency, we measured a simple Web page with Keynote [2] to compare the download time with and without the embedded calibrating image. Keynote probes are located in 25 cities within the US and 10 cities outside the US. The Web page we measured had a total size of 39 Kbytes including 13 images and was accelerated by a CDN. The increased overhead percentage is therefore higher than we would expect for a regular unaccelerated Web page with more embedded images. Table 1 shows that the increased overhead averages less than 140 ms, which is 10-12% of the total download time.

Table 1: Keynote image overhead measurements
Location Avg download latency (sec) Increased
  without image with image overhead
World wide 1.17 1.31 12%
US 1.04 1.14 10%

We also tested our system to see what would happen in the event of a failure of the redirector, image content server, or DNS server. We found that the impact of failure on the user is minimal. We tested the failure of these three components using Microsoft Internet Explorer (MSIE) 6 and Netscape Navigator 6 and found that those browsers will first load the rest of the Web page and then time out while trying to fetch the image.4 There is no visible change to the Web page or any pop-up error message; however, the Netscape logo or MSIE browser logo will provide visual feedback until the browser times out.

next up previous
Next: Analysis results Up: Experimental methodology Previous: Measurement setup