The caching of dynamic content has been the previous decade’s most popular approach to scaling the delivery of Web content. While the technology to support such scaling has evolved to be more efficient, flexible, and configurable, the fundamental principles have remained the same. The content of a static or dynamic Web site is stored remotely on an intermediary server, often termed an edge caching server, which intercepts the calls before passing the requests to the intended Web site. Before the edge caching server contacts the intended Web site, it checks to see if it has a copy of the requested content. If the server has a copy of the content, it returns it to the requestor without contacting the Web site. During this series of events, or asynchronously, the caching server checks to make sure it has a fresh copy of the content. The goal is to prevent the originating Web site’s servers from having to process each request, especially if it is repeatedly returning the same results. The benefits of caching increase if the site is dynamically generating content that is highly requested.

In this paper, Lam and Ngan intelligently articulate a complimentary and apparently effective concept called pre-fetching. For readers working with Web applications, platforms, deployments, and system designs, this paper is a must-read, offering an alternative approach to improving dynamic Web site performance.

Lam and Ngan describe their method as a proactive caching scheme, where the generation and storage of the resulting Web content is done prior to a request for the content. To be effective, the system would need to accurately predict which content could benefit from pre-fetching. This differs slightly from normal caching schemes, where caches are filled as requests are intercepted. A quick review of the test results shows that there is a dramatic difference in these approaches, so the prediction algorithm generating the pages is critical.

The devil is in the details. What actually needs to be evaluated to optimize this temporal pre-fetching solution? Temporal pre-fetching is made up of four key steps. First, the system needs to generate pre-fetching requests. Then, it needs to calculate the average delay in generating the page for the pre-fetch request. The system then calculates the priority of the page, where priority represents the usability or traffic rank for the particular page. Finally, the requests are processed. The astute reader will see that the calculation of usability is a critical factor. If the priority is not calculated correctly, the system will pre-generate pages at will, with very little predictive intent. Lam and Ngan predict the usability of a page by capturing access statistics, including how long the page stays active, the delay incurred in generating the page, the average inter-request time, and the access probability in the context of all of the other pages.

Lam and Ngan use a generic auction system to obtain a baseline and compare traditional caching and temporal pre-fetching. While the results show a marked improvement over caching, it would be interesting to see if varying the type of load would yield different results. For example, the test server is under load for only ten minutes. This short period of time might not reveal side effects of this approach; maybe resource issues only surface after 30 minutes. Another example is in how they automate the selection of click-throughs, where the link in the middle of the page has the highest probability. Is it possible the system favors such a consistent load? It would be interesting to record real traffic data, and play it back through the same system, to eliminate such a high volume of very consistent user behavior.

While readers should approach this work with enthusiasm, it should be balanced with healthy skepticism, as the data presented is not enough to draw universal claims. Lam and Ngan appear to have identified an approach that enables systems to more efficiently deliver content. This design could be applied to other data caching systems demonstrating similar transaction behavior.