up connections, use a slow initial rate of data exchange, and recover slowly from packet loss. In contrast, a network that uses persistent connections and optimizes parameters for efficiency (given knowledge of current network conditions) can significantly improve performance by reducing the number of round-trips needed to deliver the same set of data.

Optimization 2: Find better routes. In addition to reducing the number of round-trips needed, we would also like to reduce the time needed for each round-trip—each journey across the Internet. At first blush, this does not seem possible. All Internet data must be routed by BGP and must travel over numerous autonomous networks.

BGP is simple and scalable but not very efficient or robust. By leveraging a highly distributed network—one that offers potential intermediary servers on many different networks—you can actually speed up uncacheable communications by 30% to 50% or more, by using routes that are faster and much less congested. You can also achieve much greater communications reli-

ability by finding alternate routes when the default routes break.

Optimization 3: Prefetch embedded content. You can do a number of additional things at the application layer to improve Web application responsiveness for end users. One is to prefetch embedded content: while an edge server is delivering an HTML page to an end user, it can also parse the HTML and retrieve all embedded content before it is requested by the end user’s browser.

The effectiveness of this optimization relies on having servers near end users, so that users perceive a level of application responsiveness akin to that of an application being delivered directly from a nearby server, even though, in fact, some of the embedded content is being fetched from the origin server across the long-haul Internet. Prefetching by forward caches, for example, does not provide this performance benefit because the prefetched content must still travel over the middle mile before reaching the end user. Also, note that unlike link prefetching (which can also be done), embedded

content prefetching does not expend extra bandwidth resources and does not request extraneous objects that may not be requested by the end user.

With current trends toward highly personalized applications and user-generated content, there’s been growth in either uncacheable or long-tail (that is, not likely to be in cache) embedded content. In these situations, prefetching makes a huge difference in the us-er-perceived responsiveness of a Web application.

Optimization 4: Assemble pages at the edge. The next three optimizations involve reducing the amount of content that needs to travel over the middle mile. One approach is to cache page fragments at edge servers and dynamically assemble them at the edge in response to end-user requests. Pages can be personalized (at the edge) based on characteristics including the end user’s location, connection speed, cookie values, and so forth. Assembling the page at the edge not only offloads the origin server, but also results in much lower latency to the end user, as the middle mile is avoided.

 

feBRuaRY 2009 | vol. 52 | No. 2 | CommunICatIons of the aCm

49