Computerphile
June 15, 2026
TCP Congestion Control: AIMD and Slow Start Explained
TL;DR
TCP uses Additive Increase Multiplicative Decrease (AIMD) and Slow Start mechanisms to dynamically adjust data transmission rates based on network congestion signals detected through packet loss and timeouts.
Key Takeaways
- 1TCP detects congestion through two main signals: triple duplicate acknowledgements (indicating isolated packet loss) and timeouts (indicating severe congestion)
- 2AIMD (Additive Increase Multiplicative Decrease) gradually increases the congestion window by one packet per successful round trip, but cuts it in half upon detecting loss
- 3Slow Start uses exponential growth (doubling window size each round) at connection initiation to quickly reach the network's capacity before switching to AIMD
- 4Round Trip Time (RTT) is measured and used to set adaptive timeout thresholds, allowing the same algorithm to work across vastly different network latencies
- 5Real-world TCP behavior is far more complex than theoretical models due to network mesh topologies, competing traffic, and TLS encryption overhead
- 6TCP's header space constraints and need for backward compatibility with middleboxes limit how much the protocol can be modified despite being nearly 50 years old
Notable Quotes
“We call it slow start because it's tremendously quick. I know. I didn't name it.”
— Presenter
“The real internet is really really much more complicated than the sketches we draw out in the classroom.”
— Presenter
“Despite its age, it's doing a very, very, very good and complicated job.”
— Presenter
Chapters
1. TCP Basics and Packet Acknowledgement
Review of sequence numbers, acknowledgement numbers, and cumulative acknowledgements. Introduction to the concept of packets in flight and the challenge of not knowing network size or RTT in advance.
2. Congestion Detection Signals
Explanation of how TCP infers congestion through lost packets. Triple duplicate ACKs indicate isolated loss; timeouts indicate severe congestion. Discussion of why TCP cannot send detailed loss information in packet headers.
3. AIMD: Additive Increase Multiplicative Decrease
Core congestion control algorithm: window increases by one packet per successful round trip (additive increase) but halves upon loss detection (multiplicative decrease), creating a sawtooth pattern.
4. Slow Start and Exponential Growth
Initial connection phase where the congestion window doubles each round (1→2→4→8→16 packets) to rapidly reach network capacity. Named counterintuitively; called slow start because it's the mechanism to escape truly slow single-packet-at-a-time transmission.
5. TCP Flavors and RTT-Adaptive Timeouts
Different TCP variants (Cubic, Vegas, etc.) implement different responses to congestion signals. Timeout thresholds adapt to measured RTT, allowing the same algorithm to function across data centers, home networks, and satellite links with vastly different latencies.
6. Real-World TCP Behavior vs. Theory
Analysis of actual packet captures shows TCP behavior is far messier than textbook diagrams. Network mesh topologies concatenate packets, TLS adds overhead, and competing traffic creates jagged, complex patterns rather than clean sawtooth curves.
7. Observed Slow Start in Real Traffic
Demonstration of beautiful exponential slow start visible in real Ubuntu download data, reaching ~10 MB/s before transitioning to AIMD. Shows that theoretical mechanisms do work in practice despite complexity.
8. Constraints and Evolution
TCP header space is fixed and limited; middleboxes reject non-standard flag usage. Protocol has changed minimally since the 1980s due to compatibility requirements, yet it continues to perform well across diverse modern networks.
Key People & Entities
- Computerphile
- Educational video series covering computer science topics
- Shawn
- Recipient of GitHub code for analyzing network traces
- Cubic
- Most widely used TCP congestion control flavor, employed by Windows, Mac, and Linux
Glossary
- AIMD
- Additive Increase Multiplicative Decrease—TCP congestion control algorithm that increases window size by one packet per successful round trip and halves it upon detecting loss
- Slow Start
- Initial exponential growth phase of TCP connection where the congestion window doubles each round trip until first loss is detected
- Congestion Window (CWND)
- Sender-maintained estimate of the number of bytes the network can handle, adjusted based on congestion signals
- Round Trip Time (RTT)
- Time measured for a packet to travel from sender to receiver and acknowledgement to return; used to set adaptive timeout thresholds
- Packets in Flight
- Packets that have been sent but not yet acknowledged by the receiver
- Triple Duplicate ACK
- Receipt of the same acknowledgement number four times, indicating a specific packet was lost but subsequent packets arrived
- Cumulative Acknowledgement
- Single ACK confirming receipt of all packets up to a specific sequence number, rather than acknowledging each packet individually
- RFC 5681
- Internet standard document defining TCP congestion control mechanisms