Mutant: Learning Congestion Control from Existing Protocols via Online Reinforcement...
summary
Mutant is an implementation that runs a kernel protocol manager plus a user-space learner to switch among up to six in-kernel TCP schemes with a 10 ms switching granularity, using a contextual bandit with linear UCB and a GRU-based encoder. Evaluation on Pantheon, 5G cellular traces and an intercontinental fabric shows lower delay and higher throughput than each pooled protocol and often outperforms pre-trained ML schemes, while keeping fairness impact negligible.
AI-written summary · about
chapters (6)
- 0:00Mutant introduction
- 2:53Traditional CC design cycle
- 5:43Mutant architecture
- 8:12Reinforcement learning module
- 12:26Mutant evaluation
- 16:21Conclusion
related talks
13mAUTO: Adaptive Congestion Control Based on Multi-Objective Reinforcement Learning...
13mOne Protocol to Rule Them All: Wireless Network-on-Chip using Deep Reinforcement Learning
25mPCC Vivace: Online-Learning Congestion Control
17mSupercharging Packet-level Network Simulation of Large Model Training via Memoization...
11mScaling Distributed Machine Learning with In-Network Aggregation
16mUNUM: A New Framework for Network Control
16mCredence: Augmenting Datacenter Switch Buffer Sharing with ML Predictions
19mTowards Domain-Specific Network Transport for Distributed DNN Training