MIT's Polaris Touts Making Web Pages 34 Percent Faster
In a paper that will be presented at next USENIX Symposium on Networked Systems Design and Implementation, MIT PhD student Ravi Netravali and others explain their new approach, which is based on two tools:
Dependency analysis is a technique commonly used by browsers to optimize the way they load resources. Before Scout, though, says Netravali, that kind of analysis was carried through based on lexical relationships between HTML tags, which missed many fine-grained dependencies, as it can be seen in the picture below for a real case.
Polaris prioritizes the fetching and evaluation of objects along the dynamic critical path, trying to make parallel use of the client’s CPU and network, and trying to keep the client’s network pipe full, given browser constraints on the maximum number of simultaneous network requests per origin.
The researchers behind Polaris tested their system under a range of network conditions, “with latencies ranging from 25ms to 500ms, and bandwidths ranging from 1Mbps to 25Mbps”, and on 200 popular websites. This showed, they say, a decrease by up to 34 percent at the median and 50% at the 95th percentile. Performance varied significantly across sites, being higher with complex pages and lower with pages making aggressive use of caching.