Chapter 10: TCP Congestion Control

---

Metadata Card

Attribute	Content
Volume	Vol 4 — Computer Networking
Chapter	Chapter 10: TCP Congestion Control
Prerequisites	Chapter 3 (TCP Deep Dive: sliding window, flow control)
Next	Chapter 6 (Modern Protocols: WebSocket & gRPC)
Theory Depth	(4/5)
Python Relevance	≈80%; congestion window simulation, algorithm evolution implementation
Core Models	AIMD, slow start, congestion avoidance, fast recovery, CUBIC, BBR
Code	~180 lines

---

Your Progress

"The finest regulation mechanism of the post road — TCP congestion control. When too many spell messages jam the post road, all TCP transmission spells must coordinate to slow down, rather than all increasing mana output together. Otherwise, the entire post road will collapse."

Flow control handles "the receiver beacon tower can't keep up." Congestion control handles something completely different: the post road itself is too crowded.

AIMD — The Foundation

• Additive Increase (AI): Each RTT, congestion window + 1 MSS (no packet loss)
• Multiplicative Decrease (MD): On packet loss, congestion window × ½

Slow Start (It's Not Slow)

Initial cwnd = 10 MSS (RFC 6928). Every ACK received, cwnd += 1 MSS.

Initial cwnd = 10
RTT 1: send 10 packets → 10 ACKs → cwnd = 20
RTT 2: send 20 packets → 20 ACKs → cwnd = 40
RTT 3: send 40 packets → 40 ACKs → cwnd = 80

Exponential growth (2^N) — doubles every RTT. Ends when cwnd >= ssthresh or packet loss detected.

Congestion Avoidance

After ssthresh, switch to linear growth: each RTT, cwnd += 1 MSS.

AIMD full cycle: Slow start (exponential) → Congestion avoidance (linear +1/RTT) → Loss → MD (×½) → Repeat.

Fast Retransmit & Fast Recovery

Three duplicate ACKs → almost certainly a packet lost → retransmit immediately (don't wait for timeout).

Tahoe: cwnd = 1 (wasteful, restart slow start) Reno: cwnd = cwnd/2 + 3 (preserves experience)

Algorithm Evolution

Algorithm	Year	Key Innovation
Tahoe	1988	First congestion control: slow start, congestion avoidance, fast retransmit
Reno	1990	Added fast recovery
NewReno	1999	Better multi-packet loss handling
BIC	2004	Binary search for optimal window
CUBIC	2006	Linux default: cubic function growth, RTT-fair, high-speed friendly
BBR	2016	Google's paradigm shift: measure bandwidth & RTT instead of using loss as signal

CUBIC

Window growth is a cubic function: W(t) = C*(t - K)³ + W_max

• Near W_max: grows slowly (stable)
• Far from W_max: grows quickly (recovers bandwidth fast)
• RTT-fair: growth is time-based, not ACK-count-based

BBR — Bottleneck Bandwidth and Round-trip

BBR directly measures two physical quantities:

1. BtlBw: Bottleneck bandwidth (narrowest segment)
2. RTprop: Minimum RTT without queuing

Send amount per RTT = BtlBw × RTprop (pipe capacity = bandwidth × latency).

No longer uses packet loss as congestion signal — essential for wireless/mobile networks where random loss is common.

Bufferbloat & CoDel

Bufferbloat: Routers with huge buffers. When buffers are too large, RTT can balloon from 20ms to 500ms without any actual link congestion — just excessive queuing.

CoDel (Controlled Delay, 2012): Actively drops packets that have been in the queue too long (>5ms). No configuration parameters needed — it tracks minimum sojourn time.

Traveler's Notes

Congestion control is where TCP becomes truly elegant. The AIMD sawtooth shape — linear growth until loss, then halving — is a decentralized, globally convergent algorithm that requires no central coordinator. Each sender independently detects loss and makes a local decision, yet the entire system converges to fair bandwidth sharing.