TurboTest: Learning When Less is Enough through Early Termination of Internet Speed Tests
summary
A deployable two-stage system decouples Stage 1 regressor (predicts final throughput from partial TCP info: RTT, retransmissions, congestion window, throughput samples) from a Stage 2 classifier that decides optimal stopping; a transformer-based termination model exposes ε tunable accuracy tolerance. Evaluated on 1M M-Lab NDT tests (2024–2025), TURBOTEST cuts data transfer by 1.8–4.4× vs. BBR heuristics and, at ε=5%, shrinks a 12 PB/month workload to ~1 PB while reducing median error.
AI-written summary · about
chapters (12)
- 0:00TurboTest introduction
- 1:20Speed test scaling costs
- 2:22Speed test data costs
- 3:26Heuristic early termination limits
- 4:51Two-stage ML early termination
- 6:16Predicting throughput with time series
- 7:20Optimal stopping classifier
- 8:42TurboTest framework
- 9:44TurboTest vs BBR
- 11:30Real-world cost impact
- 11:51Gigabit Ethernet performance
- 13:14Key takeaways
related talks
08mProbabilistic Analysis of WCETs on RT Systems: When Your Worst Case Isn't Worst
28mThe Hitchhiker's Guide to Advanced A/B Testing Techniques
12mCATO: End-to-End Optimization of ML-Based Traffic Analysis Pipelines
01h 02mThe Key Ingredients of Optimizing Test-Time Compute and What's Still Missing
31mTesting Machine Learning Models
49mHarnessing the Twister: Zephyr Data-Driven Testing and Insights
38mEnough statistics so that Zed won't yell at you
21mThe Need for Speed: The Story of How We Achieved Industry Leading TTP