Implement BBR

[rafibaum]

I've recently been working on a version of BBR for quiche for a research project. It's broadly based on information found in the draft BBR RFC, the Linux kernel implementation, and Google's implementation in their (similarly named) QUIC library.

This version of BBR implements the core BBR algorithm, the packet conservation loss handling mechanism, and the long-term sampling functionality which so far has only been implemented in the Linux kernel implementation and isn't detailed in the RFC or the QUIC implementation. I've had to make a number of adaptations to the algorithm to get it to work in quiche and there's still some features missing:

• No ack aggregation handling.
• Packet conservation adapted since we're not using the TCP state machine anymore, adapted from how Google handles this.
• Under constant packet loss (1%) bitrate seems to collapse. May be a delivery rate estimation issue but needs investigating.
• No unit testing

I'm submitting this draft now to see if you'd be receptive to getting BBR merged in and what changes you'd want to see here before starting code review. As part of my research, I've done a bit of evaluation and I've seen typical BBR behaviour where the throughput is higher than what CUBIC usually gets you but it incurs higher packet loss rates.

Closes #514

[rafibaum]

Please excuse the double pull request. I accidentally based the old one on a WIP branch and did some refactoring on my fork. Since the last PR I've done a commenting pass explaining the implementation and added some attribution for the referenced code bases. At this point the last thing I'd like to add is some testing.

[LPardue]

Thanks for the PR Rafi, it will take some time to review but we appreciate your proposrd contribution and patience.

[rafibaum]

I've done some more evaluation of this implementation for my work and unfortunately it exhibits very unstable behaviour. When long-term sampling and packet conservation is enabled, BBR tends to enter into a feedback loop of ever-decreasing sending rates which grinds transfers to a halt. I won't be doing further work on this PR so I'm going to close it, but anyone is welcome to use it if they find it useful.

[LPardue]

Thanks for the update @rafibaum. Sad to hear there are some issues but this sounded like some good work in the process of testing anyhow. We'll keep your findings in mind should this PR (or other BBR proposals) come up again.