Repository navigation
Proposal: document and optionally enable TCP BBR during Outline server setup, particularly on AWS Lightsail #2947
OskarEichler
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hi,
We found a substantial improvement in ordinary Outline TCP throughput by enabling Linux’s existing BBR congestion-control algorithm on our AWS Lightsail servers.
Would you consider documenting this as a performance-tuning option, and potentially supporting an explicit installer option on compatible hosts?
This is independent of our Hysteria proposal: these measurements used Outline/Shadowsocks over TCP.
Measured improvement
We ran a reversible CUBIC → BBR → CUBIC experiment on our Singapore server, downloading to a Mac in Bangkok.
Each table cell contains four 8 MiB HTTPS downloads. We held the destination IP, URL, payload size and HTTP mode constant. The experiment used the same Prefix-enabled Shadowsocks listener throughout.
Relative to the initial CUBIC block, BBR achieved approximately 2.2× full-VPN throughput and 2.8× proxy-only throughput. Performance fell again after restoring CUBIC.
These were short sequential tests on a variable Internet path, so the figures are not guaranteed multipliers. However, the rollback comparison provides stronger evidence than a single before/after speed test.
Deployment and regional checks
We subsequently enabled BBR persistently on our Singapore, Frankfurt and US East Lightsail instances, using their existing kernel modules:
Smaller regional checks, using three 1 MiB transfers before and after each change, produced:
These small transfers include TCP/TLS startup and long-distance latency; they are supporting observations rather than sustained-capacity measurements.
Configuration and an important deployment detail
The persistent TCP setting was:
net.ipv4.tcp_congestion_control = bbrWe also configured the existing
tcp_bbrmodule to load at startup.Importantly, we recreated the relevant server listeners after applying the setting. Changing the default alone does not update existing listeners: accepted connections inherit the listener’s congestion-control choice. We verified the resulting sockets with
ss -ti. Linux TCP documentation.Persistence files and runtime behavior were checked; reboot persistence was configured but not tested by rebooting.
We preserved the existing
fq_codelconfiguration. Google’s guidance explains that Linux 4.20 and later provide TCP-level pacing without strictly requiring anfqqdisc. BBR pacing guidance.Suggested documentation or installer support
The current
install_server.shdoes not appear to configure BBR.A useful initial scope could be:
Because this changes the host’s TCP default and can affect other services, I suggest making that scope explicit rather than silently overriding existing host policy.
BBR did not eliminate every bottleneck in our testing. Access-network variability and cloud-instance limits still mattered. The setting also does not change native UDP/QUIC congestion control.
Would you be open to adding this guidance, particularly for Lightsail deployments, and considering an optional setup flag? We can provide sanitized measurement results and the reproduction procedure.
All reactions