| With the maturation of the World Wide Web, a variety of technologies have been created to optimize content delivery, such as caching, content delivery networks, and prefetching. While these techniques can be effective, it remains unclear how and when they are best used. This dissertation examines current Web performance practices and optimizations and provides a tool to improve Web performance by facilitating informed optimization of Web site content delivery to the end user.; This dissertation first examines Web site performance from a user perspective by characterizing whole-page performance. While many performance optimizations are very effective on individual objects, they may not be effective across the whole page. Similarly, other optimizations may only be effective at the page level. Using our Medusa proxy, we find that the optimization that has the largest impact is parallel connections. Other performance optimizations only provide incremental benefit. In some cases this is due to limited opportunity for use, and, in these cases, clients would see better performance if these optimizations were more heavily used.; This dissertation then examines Web site performance from a Web site perspective by analyzing the properties of a wide variety of web servers, focusing on content delivery. We find that our workload contains a much higher degree of uncacheable responses than previously found and that current practices impede Web site performance. We also performed a simulation analysis to estimate the potential performance benefits of content delivery networks (CDNs) for these Web sites. Our results indicate that Web sites benefit from the use of a CDN, but that CDNs can decrease the average request rate more than the peak request rate.; Lastly, this dissertation presents the design and implementation of the Cassandra Toolkit, a tool that identifies performance problems and estimates the effect of performance optimizations on a Web site (e.g., a cache or CDN). We also validate the Cassandra Toolkit against a real optimization by comparing the Cassandra Toolkit's cache analysis with a real cache on actual Web site workloads. Finally, we show how the Cassandra Toolkit can optimize several real Web sites in different ways. |
