TCP: Congestion Control

Цели урока

• Understand the 1986 congestion collapse and the positive-feedback loop that caused it
• Know the four stages of the classic algorithm: slow start, congestion avoidance, fast retransmit, fast recovery
• Tell apart loss-based (Reno, Cubic) and model-based (BBR) congestion control
• Reason about cwnd, rwnd, ssthresh and which one is bounding throughput right now
• Understand why BBR shines on long fat networks and lossy Wi-Fi

October 1986. The internet hits its first congestion collapse: the LBL to Berkeley link, all 400 yards of it, drops from 32 kbit/s to 40 bit/s - 800x slower. Link 100% busy, useful traffic near zero: everyone retransmits, retransmits trigger more retransmits. In 1988 Van Jacobson publishes the SIGCOMM paper introducing slow start and congestion avoidance, and the internet survives.

• **Video calls:** BBR helps Zoom/Meet work even with Wi-Fi losses
• **Downloads:** Cubic efficiently uses gigabit channels
• **Games:** Low cwnd = low latency, but less throughput

Предварительные знания

• TCP: Flow Control

Why Congestion Control Is Needed

**Congestion Control** - a mechanism that prevents a sender from overloading the network. Unlike flow control (protecting the receiver), congestion control protects routers and communication channels between them.

TCP uses **implicit feedback**: a packet loss or growing delay means congestion. There is no explicit signal from routers (unlike ECN). The sender draws its own conclusions.

**Flow Control vs Congestion Control:** Flow control - rwnd (receive window) from the receiver. Congestion control - cwnd (congestion window) at the sender. Actual window = min(rwnd, cwnd).

Slow Start

**Slow Start** - the initial phase of a TCP connection. The name is misleading: growth is actually exponential! The window doubles every RTT until a loss occurs or the threshold (ssthresh) is reached.

**IW (Initial Window):** Previously - 1 MSS. Now RFC 6928 recommends IW=10 MSS (~14 KB). This speeds up loading of small web pages that can be transferred in 1 RTT.

AIMD - Additive Increase, Multiplicative Decrease

**AIMD** - the main congestion control algorithm after slow start. Additive Increase: window grows linearly (+1 MSS per RTT). Multiplicative Decrease: on loss, window is halved.

**Triple Duplicate ACK:** 3 identical ACKs in a row = Fast Retransmit + Fast Recovery. cwnd is halved, but not reset to 1. Timeout (a more serious problem) resets cwnd to 1 and restarts Slow Start.

Congestion Window (cwnd)

**cwnd** (Congestion Window) - the congestion window maintained by the sender. Unlike rwnd (from the receiver), cwnd is the sender's own estimate of network state. Effective window = min(cwnd, rwnd).

**ssthresh (Slow Start Threshold):** The boundary between Slow Start and Congestion Avoidance. On loss: ssthresh = cwnd/2. The next Slow Start will continue only up to ssthresh.

TCP Variants: Reno, Cubic, BBR

Classic AIMD (TCP Reno) is not the only algorithm. Modern variants: **TCP Cubic** (Linux default), **BBR** (Google) - optimized for different scenarios.

**BBR (Bottleneck Bandwidth and RTT):** Instead of reacting to losses, BBR models the network: measures max bandwidth and min RTT. Keeps inflight = BDP. Works better with random losses (Wi-Fi).

Key Ideas

• **Congestion Control** - protecting the network from overload
• **Slow Start** - exponential cwnd growth up to the threshold
• **AIMD** - linear growth, multiplicative decrease on loss
• **cwnd** - congestion window; effective = min(cwnd, rwnd)

Related Topics

Congestion control is at the heart of TCP performance:

• TCP Flow Control — Another constraint - protecting the receiver
• QUIC — Congestion control at the application level
• Buffer Bloat — When large buffers break congestion signals

Вопросы для размышления

• Why is timeout worse than triple duplicate ACK?
• How does BBR determine bandwidth without losses?
• Why do two TCP flows with different RTTs unfairly share a channel?

Связанные уроки

• alg-20-greedy