Learning Production-Optimized Congestion Control Selection for Alibaba Cloud CDN
summary
AliCCS is a production system that learns to pick among existing congestion controllers using a GAN-based invariant representation to remove /24 hidden-state bias, decision-tree distillation for interpretability, and a prefix-keyed cache to cut inference cost. Deployed on 400+ CDN nodes for over-one-year deployment, it yields 4.9% rebuffer reduction, achieves up to 62% average and 174% peak retransmission reduction, and reports over $10M cost savings while falling back to defaults on low-confidence predictions.
AI-written summary · about
chapters (14)
- 0:00Learning-based congestion control intro
- 1:25Single CC challenges in CDNs
- 3:08Dynamic CC selection need
- 4:34AlicCS framework for CDNs
- 5:17Key challenges in applying machine learning
- 6:22Network type predicts CC
- 8:08Hidden states in metrics
- 9:33IP prefix mapping for generalization
- 10:39Interpretability with decision trees
- 12:04Reducing inference overhead
- 13:08AlicCS design pipeline
- 14:15Evaluation on 400 CDN nodes
- 15:21Comparison with state-of-the-art solutions
- 17:14Lessons from AlicCS deployment
related talks
20mLearning Relaxed Belady for Content Distribution Network Caching
18mCCC: Re-architecting Delay-based Congestion Control in Datacenter Networks
17mBridging the Gap between QoE and QoS in Congestion Control: A Large-scale Mobile...
17mTowards provably performant congestion control
14mPolicyCache: Intra-flow Learning in Congestion Control
25mPCC Vivace: Online-Learning Congestion Control
13mAUTO: Adaptive Congestion Control Based on Multi-Objective Reinforcement Learning...
27mPredictable Network