muhaha

Notes on BBR, three years later

2026-10-30 · 4 min read

I enabled BBR on every box I own somewhere around 2023 and then mostly stopped thinking about it. Going back through my notes, most of the original reasoning still holds, but a few assumptions did not survive contact with real traffic.

Still true

BBR does not treat packet loss as a congestion signal. On a path with random loss from a policer rather than a full queue, this is the whole point: CUBIC backs off, BBR keeps estimating bandwidth and pushes on. The throughput difference on a lossy long-haul link is not subtle.

net.ipv4.tcp_congestion_control = bbr
net.core.default_qdisc = fq

The fq part matters. Without a pacing-capable qdisc, BBR cannot pace its sends and the benefit shrinks considerably.

Aged badly

The "BBR is unfair to CUBIC flows" complaint turned out to be less dramatic in practice than the papers suggested, at least at my scale. What did bite is buffer bloat on the sending side: BBR will happily fill a deep queue and add latency for everything else sharing the link. If you are also running latency-sensitive traffic, look at the queue discipline before blaming the congestion control.

Values I keep coming back to

net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
net.ipv4.tcp_rmem = 4096 131072 6291456
net.ipv4.tcp_wmem = 4096 16384 6291456
net.ipv4.tcp_mtu_probing = 1

That last one has saved me twice. When a path has an MTU black hole, probing recovers automatically instead of hanging until the connection times out.