TCP Congestion Control

“Congestion control is TCP’s way of being a good citizen — slowing down when the network is overloaded.”

Overview

Congestion control prevents TCP from overwhelming the network with too much data. Unlike flow control (receiver-based), congestion control is network-based — it adapts the sending rate based on perceived network congestion (packet loss, delay).

Why Congestion Control?

graph TD
    P["No Congestion Control"] --> L1["Network becomes overloaded"]
    L1 --> L2["Router buffers overflow"]
    L2 --> L3["Packets dropped"]
    L3 --> L4["All senders retransmit"]
    L4 --> L5["More congestion!"]
    L5 --> L6["Congestion collapse!"]

In the late 1980s, the Internet experienced congestion collapse — congestion control was added to TCP to prevent this.

Key Concepts

Congestion Window (cwnd)

• cwnd: Sender’s estimate of how much data the network can handle
• ssthresh: Slow start threshold — switch point between slow start and congestion avoidance
• Effective window: min(cwnd, rwnd)

AIMD (Additive Increase, Multiplicative Decrease)

graph LR
    A["Additive Increase<br/>+1 MSS per RTT<br/>(linear growth)"] --> D["Multiplicative Decrease<br/>× 0.5 on loss<br/>(cut in half)"]
    D --> A

• Increase: Add 1 MSS per RTT when no loss (probe for bandwidth)
• Decrease: Halve cwnd when loss detected (back off aggressively)
• Property: AIMD is fair — multiple flows converge to equal share

Congestion Control Phases

graph TD
    SS["Slow Start<br/>(exponential growth)"]
    CA["Congestion Avoidance<br/>(linear growth)"]
    FR["Fast Retransmit<br/>(detect loss via dupACKs)"]
    FRCV["Fast Recovery<br/>(avoid slow start after loss)"]
    
    SS -->|"cwnd ≥ ssthresh"| CA
    CA -->|"3 dupACKs"| FR
    FR --> FRCV
    FRCV --> CA
    CA -->|"Timeout"| SS
    SS -->|"Timeout"| SS

Phase 1: Slow Start

cwnd starts at 1 MSS (or 10 MSS in modern implementations)
Double cwnd every RTT (exponential growth)
Stop when cwnd ≥ ssthresh or loss detected

Phase 2: Congestion Avoidance

Add 1 MSS per RTT (linear growth)
Probe for more bandwidth carefully
Stop when loss detected

Phase 3: Loss Detection

Detailed Congestion Control Timeline

sequenceDiagram
    participant S as Sender
    participant N as Network
    participant R as Receiver
    
    Note over S: cwnd=1, ssthresh=64
    
    S->>N: Send 1 MSS
    N->>R: Deliver
    R->>S: ACK
    Note over S: cwnd=2 (slow start: double)
    
    S->>N: Send 2 MSS
    N->>R: Deliver
    R->>S: 2 ACKs
    Note over S: cwnd=4
    
    S->>N: Send 4 MSS
    N->>R: Deliver
    R->>S: 4 ACKs
    Note over S: cwnd=8
    
    Note over S: cwnd reaches ssthresh=64
    Note over S: Switch to congestion avoidance
    
    S->>N: Send 64 MSS
    N->>R: Deliver (mostly)
    R->>S: 64 ACKs
    Note over S: cwnd=65 (+1 per RTT, linear)
    
    Note over S: Packet loss detected!
    Note over S: ssthresh=cwnd/2=32, cwnd=32

Comparison of Phases

Modern Initial Window

RFC 6928 (2013): Initial cwnd = 10 MSS (was 1-2 MSS historically)

IW = min(10 × MSS, max(2 × MSS, 14600 bytes))

This speeds up short transfers (most web traffic).

Interview Questions

Beginner

Q1: What is congestion control? Congestion control prevents TCP from overwhelming the network. It adapts the sending rate based on network conditions — slowing down when congestion is detected (packet loss) and speeding up when the network is clear. Without it, all senders would blast data, causing congestion collapse.

Q2: What is the difference between flow control and congestion control?

• Flow control: Prevents overwhelming the receiver (based on rwnd)
• Congestion control: Prevents overwhelming the network (based on cwnd)
• Both limit the sender: effective window = min(cwnd, rwnd)

Q3: What is AIMD? AIMD (Additive Increase, Multiplicative Decrease) is the classic congestion control strategy: increase cwnd by 1 MSS per RTT (slow, linear growth) and decrease cwnd by half on loss (fast, multiplicative decrease). This is fair — multiple flows converge to equal bandwidth sharing.

Intermediate

Q4: Explain the difference between slow start and congestion avoidance.

• Slow start: cwnd doubles every RTT (exponential growth). Used at connection start or after timeout. Aggressive but safe when cwnd is small.
• Congestion avoidance: cwnd increases by 1 MSS per RTT (linear growth). Used when cwnd reaches ssthresh. Conservative probing for more bandwidth.

Q5: How does TCP detect congestion? TCP infers congestion from: (1) Packet loss: Detected by duplicate ACKs (3 dupACKs = fast retransmit) or retransmission timeout. (2) Increased delay: Some algorithms (BBR, Vegas) use RTT increases as congestion signal. (3) ECN: Explicit Congestion Notification — routers mark packets instead of dropping.

Q6: What happens when a timeout occurs? On timeout: (1) ssthresh = cwnd / 2, (2) cwnd resets to 1 MSS (or initial window), (3) Enter slow start, (4) Retransmit lost segment. This is aggressive — it takes many RTTs to recover. Fast retransmit/fast recovery is preferred over timeouts.

Advanced / FAANG-Level

Q7: Why is AIMD considered “fair”? AIMD converges to fairness because: (1) Additive increase grows all flows at the same rate (+1 MSS/RTT), (2) Multiplicative decrease cuts proportionally — a flow with more bandwidth loses more, (3) Over time, competing flows converge to equal share. Mathematically: if two flows share a link, they converge to the line x=y in the bandwidth allocation graph.

Q8: How does ECN (Explicit Congestion Notification) improve congestion control? ECN signals congestion before packet loss:

1. Sender sets ECN-capable bits in IP header
2. Router experiencing congestion marks packets (sets CE bit)
3. Receiver echoes congestion back to sender (ECE flag in TCP ACK)
4. Sender reduces cwnd and sets CWR flag Benefits: No packet loss, no retransmission, lower latency, better for real-time applications.

Q9: Design a congestion control algorithm for data center networks. Data center requirements: ultra-low latency, high throughput, bursty traffic, incast (many-to-one).

Design (like DCTCP):

1. ECN marking at low threshold: Mark at 20% buffer occupancy (not 100%)
2. Fine-grained cwnd adjustment: Reduce cwnd proportionally to marked packets (not 50%)
3. Maintain large window: Only reduce by small amount
4. Fast convergence: React to congestion in 1 RTT
5. No slow start after loss: Maintain high cwnd
6. Pacing: Smooth sending rate to avoid micro-bursts

Common Mistakes

1. ❌ Confusing cwnd with rwnd — cwnd is network-based, rwnd is receiver-based
2. ❌ Thinking slow start is “slow” — it’s exponential growth, faster than congestion avoidance
3. ❌ Forgetting that timeout is more expensive than 3 dupACKs — timeout resets to slow start
4. ❌ Assuming all TCP flows get equal bandwidth — depends on RTT and algorithm
5. ❌ Not considering congestion control in application design — it affects throughput

Summary

• Congestion control prevents network overload using cwnd (congestion window)
• AIMD: Additive increase, multiplicative decrease — fair and stable
• Slow start: Exponential growth (cwnd doubles per RTT)
• Congestion avoidance: Linear growth (cwnd + 1 per RTT)
• Loss detection: 3 dupACKs (fast retransmit) or timeout (slow start)
• Modern algorithms: Reno, CUBIC, BBR — evolved from the basic framework

Cross-References

• Slow Start — Exponential growth phase
• Congestion Avoidance — Linear growth phase
• Fast Retransmit — Loss detection via dupACKs
• Fast Recovery — Avoiding slow start after loss
• TCP Reno — Classic congestion control
• TCP CUBIC — Linux default
• TCP BBR — Google’s bandwidth-based algorithm

Cross References

• Slow Start
• Congestion Avoidance
• Flow Control
• Load Balancing
• BBR