Networking · Cluster
Congestion Control
Signals, fairness, and queue management.
- 01 Ack Clocks and the Congestion Window: How a Sender Infers Available Capacity Why a TCP sender measures the path with returning acknowledgements instead of a timer, how the congestion window combines with the receive window, and what a stalled clock does to the estimate.
- 02 Additive Increase, Multiplicative Decrease, and the Fairness of the Sawtooth Why the increase is additive and the decrease is multiplicative, how the resulting sawtooth converts loss rate into throughput, and where the convergence argument stops holding.
- 03 Slow Start: Exponential Probing and the Initial Window Problem A new connection has no estimate of path capacity, so it doubles its window each round trip until a threshold or a loss. The cost is a burst that a small buffer cannot absorb.
- 04 Duplicate Acknowledgements and How a Sender Distinguishes Loss from Reordering Three duplicate ACKs are a heuristic for one lost segment, not a proof. Fast retransmit and fast recovery act on it, and the ambiguity with packet reordering shapes the whole design.
- 05 Queueing Delay as a Congestion Signal and the Base RTT Problem A queue that is filling raises round-trip time before it drops anything, so delay can warn earlier than loss. Turning that warning into a control signal requires a stable estimate of the delay without queues.
- 06 Tail Drop, Buffer Bloat, and How Much Queue a Link Actually Needs A drop-tail queue signals congestion only by overflowing, and the delay it adds while full is the latency every other flow experiences. The bandwidth-delay product gives a floor, not an optimum.
- 07 Active Queue Management and Explicit Congestion Notification Random early detection marks or drops before the buffer fills, and ECN lets the receiver echo the mark instead of losing a packet. What the queue can signal and what the sender must do with it.
- 08 Model-Based Congestion Control: Delivery Rate, Minimum RTT, and Pacing Estimating the bottleneck rate and the propagation delay separately lets a sender set its window from the product of the two instead of probing for loss. The estimators are the whole design.
- 09 RTT Fairness: Why Shorter Paths Win and How Algorithms Compensate Two flows on the same bottleneck with different round trips do not get equal shares, and the bias is a direct consequence of the control law. The fixes trade that bias against stability.
- 10 DCTCP and the Proportional Response to Congestion Marks A single ECN mark carries one bit, but the fraction of marked packets carries the severity of congestion. DCTCP converts that fraction into a window reduction sized to match.