4 ms·
Delay-gradient congestion control
- Animats 11y agoThis is useful only for long-lived connections. One of the big problems with web connections is that they're so short-lived. TCP never sees enough data to construct a good bandwidth/delay/jitter model of the channel. (Video is a different case entirely - long connections, delay isn't a big problem but loss and delay jitter are.) HTTP2 forces long TCP connections, which gives TCP a better chance to characterize the channel. HTTP2 only helps if all the traffic is to one site, though. This delay-gradient congestion control will work best in that case. They should pitch this to Google. (The optimal case for HTTP2 is a world where every URL looks like "https://cloud.google.com/6e0790d661f75c65b10dc95293f79630" https://cloud.google.com/6e0790d661f75c65b10dc95293f79630". One CDN to rule them all. I hope that's not the endgame for the Web.)
- wtallis 11y agoIt's an interesting and potentially useful tool, but it definitely isn't a solution to anything. In addition to not being effective when used under HTTP as deployed on the current web, it doesn't engage its smarter bits on networks with effective AQM, it doesn't compete well in a mixed environment with some traffic being generated by traditional TCP, and LEDBAT already exists. Every situation that this could help with still needs a real fix elsewhere in the stack (AQM below, HTTP2 above).