Using application informed pacing to be a friendly internet neighbor
summary
Netflix implements application-informed pacing and a joint ABR-plus-rate-control called Sammy that sets per-chunk pace rates (6× top bitrate when buffer empty, 3× when full) and paces TCP emissions. Experiments show a 76% throughput reduction per session with no QoE loss, retransmissions down 51%, median RTTs improved 38%, client artwork renders 5–6% faster, neighbor TCP throughput +28% and response time −18% (UDP neighbor delay −51%), demonstrating a built, measured, and effective smoothing system.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:29Streaming traffic latency
- 4:42Burstiness in HTTP/TCP
- 8:18Reducing burstiness with pacing
- 12:37Pace rate calculation and benefits
- 14:51Real-world results and demo
- 17:22Summary
related talks
03mImproving ABR at Netflix
22mProviding better video experiences for the next billion users.
17mDelivering better video encodes for legacy devices
42mPredictive Load-Balancing: Unfair but Faster & more Robust
16mImproving Streaming Experience with Bayesian Optimization, from AB to...
10mGoing Beyond Bitrate: Adaptation Based on Video Context
45mNetflix's success: Combining collaboration, hardware monitoring & AI
13mSENSEI: Aligning Video Streaming Quality with Dynamic User Sensitivity