Chapter 19· 9 min read · illustrated

TCP Congestion Control

How TCP avoids melting the network — slow start, AIMD, and ECN

Flow control protected the receiver. Congestion control protects everything in between — the shared routers and links of the network. TCP has no direct view of the network’s state, so it cleverly infers congestion from packet loss and adjusts how fast it sends.

The congestion problem

The links between hosts are shared by everyone. If senders push too much data, router queues fill up and packets get dropped. If every sender just kept blasting, the network would collapse under congestion. Congestion control makes senders self-regulate.

Many senders funnel into one overloaded router whose queue overflows and drops packets. — Tap to enlarge

A second window: cwnd

TCP keeps a second limit called the congestion window (cwnd), separate from the receiver’s rwnd. While rwnd reflects the receiver’s capacity, cwnd reflects the sender’s estimate of the network’s capacity. The sender must respect both.

rwnd (receiver limit) next to cwnd (network limit); the sender obeys both. — Tap to enlarge

Slow start

A new connection does not know the network’s capacity, so it starts small and ramps up fast: slow start doubles cwnd every round-trip (1, 2, 4, 8 …). Despite the name, this is exponential growth — a quick probe upward until the first sign of trouble.

A graph where cwnd doubles each RTT: 1, 2, 4, 8. — Tap to enlarge

Congestion avoidance

Once cwnd passes a threshold, TCP switches to congestion avoidance: it grows the window slowly — roughly +1 segment per round-trip. This additive increase gently probes for more bandwidth without provoking loss.

After a threshold, cwnd rises slowly by +1 per RTT. — Tap to enlarge

Loss is the signal

Since TCP cannot see the network directly, it treats packet loss as evidence of congestion. On loss, it sharply reduces cwnd (typically halving it) to relieve the network, then resumes growing. This is a multiplicative decrease.

A lost packet makes the sender cut cwnd in half. — Tap to enlarge

The AIMD sawtooth

Put together, congestion avoidance gives the classic AIMD pattern — Additive Increase, Multiplicative Decrease. Plotted over time, cwnd looks like a sawtooth: climb slowly, drop sharply on loss, climb again. This is what keeps many TCP flows sharing a link fairly.

A sawtooth graph: slow linear rise, sharp drop on loss, repeating. — Tap to enlarge

ECN — a gentler signal

Waiting for a drop is wasteful. ECN (Explicit Congestion Notification) lets routers mark a packet as "congestion building" instead of dropping it. The receiver echoes the mark back, and the sender slows down before any data is actually lost. It is a mechanism (across IP and TCP), not an algorithm.

A router marks a packet instead of dropping it, so the sender slows before loss. — Tap to enlarge

How much can I send?

The rule: At any moment the sender may have at most MIN(cwnd, rwnd) bytes in flight — the smaller of the network’s limit and the receiver’s limit. Both flow control and congestion control apply at once.

send window = MIN(cwnd, rwnd). — Tap to enlarge