3 7 TCP Congestion Control

Summary

This video script delves into the intricacies of TCP congestion control, a cornerstone for maintaining internet stability. It explores classic TCP, where senders increase their rate until packet loss occurs, prompting a rate reduction. The script introduces additive increase multiplicative decrease (AIMD), the cornerstone of TCP's congestion control, and慢启动, an exponential rate increase at connection start-up. Variants like TCP Cubic, which adjusts the rate increase strategy, and delay-based approaches that use RTT measurements to manage congestion, are also discussed. The video touches on Explicit Congestion Notification (ECN), a network-assisted method, and concludes with a reflection on TCP fairness, emphasizing the importance of TCP in internet reliability and performance.

Takeaways

• 🌐 **TCP's Role in Internet Success**: TCP's congestion control algorithms are crucial to the internet's success, reflecting its central role in the TCP/IP protocol stack.
• 📈 **Congestion Collapse**: Before TCP's congestion control, the internet experienced congestion collapse, where increasing aggregate TCP sender rates led to decreased throughput.
• 🔄 **TCP's Congestion Control Mechanism**: TCP uses loss events to detect congestion, rather than signals from congested routers, employing an end-to-end approach.
• 📊 **AIMD Algorithm**: TCP uses the Additive Increase Multiplicative Decrease (AIMD) algorithm to control congestion, increasing the sending rate gradually and decreasing it significantly upon loss.
• 🚀 **TCP's Slow Start**: TCP starts with a slow start phase, exponentially increasing the sending rate until it switches to the AIMD phase at a certain threshold.
• 🔄 **TCP Cubic**: TCP Cubic is a modification to the AIMD algorithm that increases the window size based on the cube of time since the last congestion event, improving performance.
• 📏 **Delay-Based Approach**: An alternative to loss-based congestion control, delay-based approaches measure RTTs to infer congestion and adjust the sending rate accordingly.
• 🔔 **Explicit Congestion Notification (ECN)**: ECN is a network-assisted approach where routers signal congestion to TCP senders, which then react without packet loss.
• ⚖️ **Fairness in TCP**: TCP's AIMD algorithm ensures fairness among TCP senders by allowing equal sharing of bandwidth when flows are congested.
• 🏗️ **Building on UDP**: While it's possible to build applications on top of UDP with reliability and flow control, they wouldn't have built-in congestion control and could unfairly compete with TCP flows.

Q & A

What is TCP congestion control?

TCP congestion control is a mechanism used by TCP (Transmission Control Protocol) to prevent network congestion collapse by adjusting the rate at which data is sent over a network.

How does classic TCP congestion control work?

Classic TCP congestion control works by gradually increasing the sending rate until packet loss occurs, then backing off. It follows an 'additive increase, multiplicative decrease' (AIMD) approach.

What is the significance of Van Jacobson's paper on congestion avoidance and control?

Van Jacobson's paper laid the foundation for modern TCP congestion control mechanisms. It introduced the AIMD algorithm that has been in use for over 30 years.

What is the 'sawtooth' behavior in TCP congestion control?

The 'sawtooth' behavior refers to the pattern where the TCP sending rate increases until a loss event occurs, then decreases, and then starts increasing again.

How does the AIMD algorithm contribute to TCP's congestion control?

AIMD contributes to TCP's congestion control by increasing the sender's rate by one segment per RTT when no loss is detected, and decreasing the rate by half when a loss is detected.

What is the purpose of the slow start phase in TCP?

The slow start phase is an initial phase where TCP exponentially increases its sending rate by doubling it every RTT until it reaches a certain threshold or experiences packet loss.

What is TCP Cubic and how does it differ from AIMD?

TCP Cubic is a congestion control algorithm that operates similarly to AIMD but with a different method of increasing the congestion window size. It ramps up more quickly to near the rate of the last experienced loss and then increases more cautiously.

What is the concept behind delay-based congestion control?

Delay-based congestion control aims to keep the network path just full enough to keep the congested router busy without causing congestion loss, by using RTT measurements.

What is Explicit Congestion Notification (ECN) and how does it help in congestion control?

ECN is a network-assisted approach where routers signal congestion by setting a bit in the IP datagram header. TCP senders react to this signal by reducing their sending rate, similar to how they react to packet loss.

What is fairness in the context of TCP congestion control?

Fairness in TCP congestion control refers to the equal sharing of bandwidth among competing TCP flows, ensuring that no single flow dominates the bottleneck link.

How does TCP ensure fairness among multiple flows sharing a bottleneck link?

TCP ensures fairness by having flows that exceed the available bandwidth reduce their sending rates when packet loss occurs, allowing other flows to increase their rates until the bandwidth is fairly shared.