Congestion Control in TCP

Congestion control ensures the sender doesn't overwhelm the network (routers, links) — unlike flow control which protects the receiver.

The Problem

Many Senders ---- too much data ----> Router
                (buffer overflow → packets dropped → network collapse)

Congestion Window (cwnd)

• Sender-side limit on unacknowledged data in flight.
• Sender controlled (unlike rwnd which is receiver controlled).
• Not advertised — sender calculates it internally.

How TCP Detects Congestion

Signal	Meaning	Severity
Packet loss (timeout)	Severe congestion	High
3 Duplicate ACKs	Mild congestion	Low

• Timeout — no ACK received in time → severe
• 3 Dup ACKs — receiver got out-of-order packet, keeps ACKing last good one → mild

TCP Congestion Control Algorithms

1. Slow Start

• Begins when connection starts or after timeout.
• cwnd starts at 1 MSS (Maximum Segment Size).
• Doubles every RTT (exponential growth).
• Continues until ssthresh (slow start threshold) is reached.
• Despite the name, slow start is actually fast (exponential). "Slow" refers to starting from 1 MSS instead of full speed.

2. Congestion Avoidance

• After reaching ssthresh.
• cwnd grows linearly — increases by 1 MSS per RTT.
• Probes network carefully for available bandwidth.

3. Fast Retransmit

• On receiving 3 duplicate ACKs.
• Sender immediately retransmits lost packet.
• Doesn't wait for timeout (timeout is slow).

4. Fast Recovery

• After Fast Retransmit (3 dup ACKs).
• Does NOT go back to Slow Start.
• Instead: ssthresh = cwnd / 2 ; cwnd = ssthresh + 3 ; continue with Congestion Avoidance.

Slow Start Example

RTT 1: cwnd = 1 MSS
RTT 2: cwnd = 2 MSS
RTT 3: cwnd = 4 MSS
RTT 4: cwnd = 8 MSS  ← hits ssthresh → switch to Congestion Avoidance

Congestion Avoidance Example

RTT 1: cwnd = 8 MSS
RTT 2: cwnd = 9 MSS
RTT 3: cwnd = 10 MSS
RTT 4: cwnd = 11 MSS  ← congestion detected!

Fast Retransmit Example

Sender receives: ACK5, ACK5, ACK5 (3 dup ACKs)
→ Immediately resend packet 6 without waiting for timeout

TCP Tahoe vs TCP Reno

Event	TCP Tahoe	TCP Reno
Timeout	cwnd=1, go to Slow Start	cwnd=1, go to Slow Start
3 Dup ACKs	cwnd=1, go to Slow Start	Fast Recovery (cwnd = ssthresh+3)

• Tahoe — treats all loss the same (harsh)
• Reno — distinguishes between mild and severe congestion (smarter)
• Modern TCP uses Reno or newer (CUBIC, BBR)

cwnd over Time (Conceptual Sawtooth)

cwnd
 |            /
 |           /   Congestion
 |          /    Avoidance
 |         /   /
 |        /  --/----x  (3 dup ACKs)
 |       /  /      \
 |      /  / ssthresh \----/
 |     /  /             \
 |    /SS/                \  SS again (timeout)
 |---------------------------------------→ time

SS = Slow Start (exponential)
/  = Congestion Avoidance (linear)
x  = congestion detected

AIMD (Additive Increase Multiplicative Decrease)

• Additive Increase — grow cwnd by 1 MSS per RTT (slow, careful).
• Multiplicative Decrease — cut cwnd by half on congestion (aggressive reduction).
• This creates the sawtooth pattern typical of TCP throughput.

No congestion → cwnd + 1 MSS (per RTT)
Congestion    → cwnd / 2

Modern Congestion Control

Algorithm	Used By	Approach
CUBIC	Linux default	Cubic function for window growth
BBR	Google	Measures bandwidth & RTT directly
QUIC	HTTP/3	UDP-based, own congestion control

Congestion Control vs Flow Control (Summary)

	Congestion Control	Flow Control
Protects	Network	Receiver
Window	cwnd (sender)	rwnd (receiver)
Controlled by	Sender	Receiver
Signal	Packet loss / dup ACKs	rwnd in ACK
Algorithm	Slow start, AIMD	Sliding window