Copa: Practical Delay-Based Congestion Control for the Internet
summary
Algorithm sets target rate = 1/(δ · dq) and estimates dq by taking a minimum over a half-RTT window to remove positive noise, periodically emptying queues so new flows learn true RTT. Control loop makes additive-increase/additive-decrease toward that target with a velocity term, and a detector switches to AIMD on 1/δ when buffer-fillers appear. Claims: empties queues every 5 RTTs, yields Cubic-like throughput with much lower delay, fairer across RTTs, robust to non-congestive loss and large buffers.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:21Why are we still talking about it
- 4:05Basic Goals
- 4:25Target Rate
- 8:24Copa Algorithm
- 9:14One Sender Case
- 10:56Two Consequences
- 13:12Experimental Results
- 14:02Limitations
- 15:24Performance
- 17:44Summary
- 18:55Questions
related talks
17mTowards provably performant congestion control
18mCCC: Re-architecting Delay-based Congestion Control in Datacenter Networks
16mManaging Congestion Control Heterogeneity on the Internet with Approximate Performance...
19mOSCAR: O(1)-Step Convergence and Readily-deployable Congestion Control
17mControlling Arbitrary Internet Queues with Titrate
43mHow internet congestion control actually works in the bufferbloat age
16mR-TCP: A Framework to Optimize TCP Performance Over Rate-Limiting Networks
18mRevisiting Congestion Control for Lossless Ethernet
