BBRv2

[JavaPerformance]

Hi all,

My current project relies heavily on Quinn. Thanks for a great library.

Because of a specific WAN bulk-transfer need, I have been experimenting with BBRv2 support in a fork. The prototype currently includes:

• selectable BBRv2 alongside the existing BBR path
• round-bounded loss classification
• inflight_hi / inflight_lo and bw_hi / bw_lo bounds
• packet-derived delivery-rate sampling
• first-class ECN/loss signal plumbing
• a guarded experimental flag for the more aggressive inflight_hi shrink path
• 29 focused BBR unit tests

The branch is here if useful:

https://github.com/JavaPerformance/quinn/tree/bbrv2-prototype

I am not assuming this is PR-ready. Before spending time turning it into upstreamable patches, I wanted to ask about the preferred direction.

One specific API question: current Quinn main exposes ECN through is_ecn: bool on Controller::on_congestion_event. In the prototype I used a typed congestion event enum internally/to the controller because it made BBRv2 loss-vs-ECN handling clearer.
Would you prefer any BBRv2 work to build on the existing is_ecn API, or would a richer typed congestion event be worth discussing?

If there is interest, I am happy to split this into smaller PRs, probably starting with either the congestion-event API question or the sent-packet/rate-sampling pieces.

No pressure either way. If this is out of scope for Quinn right now, I am also happy to keep it fork-local.

Paul

[djc]

Thanks for working on this! It might be useful to have a look at

• feat(bbr3) Implemented bbr3's latest version based on the RFC #2481

which is pretty advanced at this stage.

Given the existence of (1) version 3 of BBR and (2) an actual implementation of that standard for Quinn, I am not sure we'll want to merge a BBRv2 implementation. However, I think our traits should be able to support v2, so I'd at least be open to reviewing changes to the trait definitions. Though I think there's some of that in the aforementioned PR too, so should probably align with that to get to consensus trait definition that can serve all the use cases (assuming you're still interested in v2 given the existence of a v3 implementation!).

[JavaPerformance]

Thank you for the reply.

I should have researched and found the ongoing BBRv3 work before being so presumptuous — next time I'll look before I leap :)

#2481 looks excellent. I'll pivot to using it.

I'll put a bow on our BBRv2 branch and keep it parked in our fork. Happy to share notes if it's ever useful.

Paul