Accelerating last-mile web performance with popularity-based prefetching

As broadband speeds increase, latency is becoming a bottleneck for many applications—especially for Web downloads. Latency affects many aspects ofWeb page load time, from DNS lookups to the time to complete a three-way TCP handshake; it also contributes to the time it takes to transfer the Web objects for a page. Previous work has shown that much of this latency can occur in the last mile [2]. Although some performance bottlenecks can be mitigated by increasing downstream throughput (e.g., by purchasing a higher service plan), in many cases, latency introduces performance bottlenecks, particularly for connections with higher throughput. To mitigate latency bottlenecks in the last mile, we have implemented a system that performs DNS prefetching and TCP connection caching to the Web sites that devices inside a home visit most frequently, a technique we call popularity-based prefetching. Many devices and applications already perform DNS prefetching and maintain persistent TCP connections, but most prefetching is predictive based on the content of the page, rather than on past site popularity. We evaluate the optimizations using a simulator that we drive from traffic traces that we collected from five homes in the BISmark testbed [1]. We find that performing DNS prefetching and TCP connection caching for the twenty most popular sites inside the home can double DNS and connection cache hit rates.