Dropbox have a good real-world comparison [1] between BBRv1 and BBRv2. They found the retransmission rate to be 4x lower for v2. It's on their edge network, so would be interesting to see how v2 performs on connections with a high bandwidth delay product.
It's been a while since I caught up on the BBR work, are there any ways to tweak the algorithm for cases where you want 'increased fairness' – ie. deferring to other data streams because the current one is a lower priority? This would probably mean approximating LEDBAT, which is often used in large backgrounded file downloads like system updates.
When should I send my packets to maximize my throughput while not unfairly affecting everyone else's packets?
And the best strategy in such a game with the whole intermediate network hard to model and with plenty of unknowns, tends to be some kind of adversarial neural net.
Clearly running that for every packet is impractical, but it could at least be used to make a set of strategies to choose from.
6 comments
[ 3.3 ms ] story [ 26.5 ms ] threadhttps://datatracker.ietf.org/meeting/104/materials/slides-10...
So v2 now also uses packet loss as a congestion indicator and supports ECN.
[1] https://dropbox.tech/infrastructure/evaluating-bbrv2-on-the-...
When should I send my packets to maximize my throughput while not unfairly affecting everyone else's packets?
And the best strategy in such a game with the whole intermediate network hard to model and with plenty of unknowns, tends to be some kind of adversarial neural net.
Clearly running that for every packet is impractical, but it could at least be used to make a set of strategies to choose from.
https://web.mit.edu/remy/
https://www.usenix.org/conference/nsdi18/presentation/dong
https://dl.acm.org/doi/10.1145/3229543.3229550