TCP 혼잡 제어 튜닝

TCP congestion control decides how fast a sender may transmit without overwhelming the network.

It governs the trade-off between throughput and latency, and the right algorithm can dramatically change performance.

The congestion window (cwnd) limits how many unacknowledged bytes are in flight. Algorithms grow and shrink cwnd in response to ACKs and loss signals.

Too small wastes bandwidth; too large causes loss and bufferbloat.

Two families exist:

• Loss-based (CUBIC, Reno) grow until packets drop
• Latency/model-based (BBR) estimate bandwidth and RTT to avoid filling buffers

BBR often wins on long-distance, lossy links.

See which algorithms the kernel has loaded.

CUBIC is the Linux default. Confirm what is active.

Load the BBR module, then set it as the default. BBR pairs best with the fq queue discipline.

Runtime sysctl changes are lost on reboot. Persist them in /etc/sysctl.d/.

Congestion control can only fill a window the buffers allow. For high bandwidth-delay-product paths, raise the autotuning maximums.

Optimal window size = bandwidth x round-trip time. A 1 Gbps link with 100 ms RTT needs about 12.5 MB of buffer to stay full.

Undersized buffers cap throughput regardless of the algorithm.

Inspect live socket state, cwnd, and RTT with ss to confirm tuning is taking effect.

Benchmark before and after with a tool like iperf3 across a real path. Always measure both throughput and latency, since aggressive tuning can trade one for the other.

You can now tune TCP congestion control:

• Understand cwnd and the bandwidth-delay product
• Compare loss-based (CUBIC) vs model-based (BBR)
• Enable BBR with the fq qdisc and persist it
• Size tcp_rmem/tcp_wmem buffers correctly
• Verify with ss -tin and benchmark with iperf3

This extends your kernel parameter and benchmarking lessons.