BBR + fq on a cheap VPS: what actually changed
Every VPS tutorial tells you to enable BBR. Few of them show what happens afterwards. I ran iperf3 between a small US West VM and a machine in East Asia twice a day for a week, half the time with CUBIC and half with BBR.
The change
net.core.default_qdisc = fq
net.ipv4.tcp_congestion_control = bbr
Drop those in /etc/sysctl.d/90-bbr.conf and run sysctl --system. Existing connections keep their old algorithm; new ones pick up BBR.
Results
| CUBIC (median) | BBR (median) | |
|---|---|---|
| Throughput, single stream | 38 Mbit/s | 141 Mbit/s |
| Throughput, evening peak | 9 Mbit/s | 63 Mbit/s |
| Retransmits per run | ~1,900 | ~4,700 |
On a long, lossy path the difference is dramatic, because CUBIC treats every lost packet as congestion and backs off hard. BBR models bandwidth and RTT instead, so random loss on a trans-Pacific link does not crater throughput.
The caveats
- BBR retransmits more. On a shared link that is not free; be a good neighbour.
- On short, clean paths (same region) I measured no meaningful difference.
- It only helps the sending side. Uploads to the VM depend on the client's stack.
Verdict: for a small server talking to users on another continent, it is the highest value two-line change I know of.