Config-Snob: Tuning for the Best Configurations of Networking Protocol Stack
summary
Config-Snob selects configurations from an eight-knob protocol space using historical data, two-level client grouping, and a multi-arm bandit that bootstraps a Bayesian blackbox optimizer with controlled, small-step exploration to avoid large user regressions. A causal inference reward estimator filters harmful changes and refines search. Deployed in a QUIC-based server across Turkey and the Philippines (20% baseline kept on default), it yields 15–36% mean and 62–70% median completion-time improvements in production and 49% mean / 55% median gains in simulated QUIC experiments.
AI-written summary · about
chapters (13)
- 0:00Config-Snob overview
- 1:47Static tuning challenges
- 3:14Black-box optimization
- 5:26Two-level client grouping
- 6:29Multi-armed bandit
- 7:54Conservative exploration
- 9:22Calibration inference
- 11:08System design overview
- 12:16Production results
- 13:40Performance variations
- 14:44Comparison with approaches
- 15:26Conclusion and future
- 16:07Q&A
related talks
14mInternet Connection Splitting: What’s Old is New Again
18mNetKeeper: Enhancing Network Resilience with Autonomous Network Configuration...
21mConfig2Spec: Mining Network Specifications from Network Configurations
23mACK to the Future: TCP in 2025
15mMortise: Auto-tuning Congestion Control to Optimize QoE via Network-Aware Parameter
18mImproving Transfer Times in the Backbone Network Using QUIC Jump Start
24mNetBlocks: Staging Layouts for High-Performance Custom Host Network Stacks
32mSnabb Switch simple and fast packet networking