Guy's Pod » Blog Archive » HTTP Pipelining – Not So Fast…(Nor Slow!)

Summary

HTTP Pipelining, while it should help in theory, doesn’t seem to make browsing the web faster. The reasons vary, but simply put, the real web got in the way of a protocol optimization… Websites today are complicated beasts, network proxies don’t behave as they should, and accelerating in this context is not a simple feat. Subbu Allamar recently reached a similar a similar conclusion from his own set of tests.

One response to that could be to just use SPDY, with true multiplexing. However, SPDY requires SSL (which has performance costs), and does not address all bottlenecks either. Either way, in my opinion, these protocol optimizations will have little effect until we start building optimizations that accelerate across domains, and address front-end bottlenecks. Websites are not going to use less domains – domain sharding is still a good optimization on old browsers, and 3rd party widgets are being used more and more.  Similarly, front-end bottlenecks will stick around until all web software is made perfect…

Making pipelining more effective won’t be easy, but I can offer two suggestions:

1. Support piping requests to multiple domains on one connection. SPDY already does this for hosts that share an IP and certificate. If we can figure out how to validate a server is “allowed” to serve content for several domains, we can mitigate the impact of sharding and better leverage CDNs, which host multiple domains.
2. Use Pipelining more aggressively, leveraging hints from the server. Much of the reason browsers “hold back” on using pipelining is not knowing what the server can handle. If the server could explicitly tell the browser to pipe more, browsers could do just that. Such hinting can be done using headers, using Mark Nottingham‘s “hinting” draft spec, or in other ways. Akamai made some HTTP/2.0 proposals of a similar nature.

Whatever the solution, we need pipelining implementors (and the entire team working on HTTP/2.0) new solutions to “think big”, or they’ll fall short.

If you’re interested in the full results of the tests, ping me via twitter (@guypod).