PolicyCache: Intra-flow Learning in Congestion Control
summary
PolicyCache claims intra‑flow learning that trains and runs a nonparametric Hoeffding Adaptive Tree (HAT) inside each flow, plus backup exploration and micro exploration for safe labeling and continual updates. A Linux prototype and convergence proof back the claim. Evaluations report faster convergence than pure exploration, consistently high throughput and low latency across emulated and intercontinental links, fair sharing across RTTs, balanced coexistence with cubic and BBR, and CPU overhead much lower than NN CC and comparable to heuristic schemes.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:21Limitations of congestion control
- 2:46Intra-flow learning concept
- 4:14Non-parametric models for fast adaptation
- 4:56PolicyCache system design
- 6:21Hoeffding Adaptive Tree
- 7:23Exploration mechanisms
- 8:06Data collection and scoring
- 9:07Implementation details
- 9:52Evaluation results
- 10:57Fairness and overhead analysis
- 12:02Dynamic switching
- 13:04Conclusion
related talks
23mFast Software Cache Design for Network Appliances
16mONCache: A Cache-Based Low-Overhead Container Overlay Network
16mBackpressure Flow Control
25mPCC Vivace: Online-Learning Congestion Control
13mAUTO: Adaptive Congestion Control Based on Multi-Objective Reinforcement Learning...
17mPyrrha: Congestion-Root-Based Flow Control to Eliminate Head-of-Line Blocking in...
19mLearning Production-Optimized Congestion Control Selection for Alibaba Cloud CDN
17mTowards provably performant congestion control
