Netflix's success: Combining collaboration, hardware monitoring & AI
summary
Migration from m5.4xlarge to m5.12xlarge shows only ~20% throughput until PMU sampling with PerfSpect reveals 3x worse CPI, high L1-bound and machine-clears on a subset of nodes; line-level VTune sampling with Intel VTune Profiler maps the hotspots to OpenJDK’s secondary-superclass cache causing false sharing. Padding the data layout (64-byte separation) eliminates the slow band and yields 3.5x throughput at similar CPU and latency. Separately, a learned downsampler (oneDNN on Xeon) raises VMAF by a few points, cuts rebuffering, and improves encoding frames/sec by 15%–2x.
AI-written summary · about
chapters (10)
- 0:00Introduction and collaboration
- 4:09Observability at Netflix
- 7:31Identifying the performance gap
- 10:13Intel's diagnostic methodology
- 13:26Hardware-level investigation
- 17:15Solving the false sharing issue
- 25:30Performance results summary
- 27:14AI in the encoding pipeline
- 36:00Downsampler optimization
- 40:02Hardware acceleration and closing
related talks
45mKeynote Address: Netflix: Streaming Entertainment to 200 Million Members Around the World
38mNetflix's Ray Platform: From Deep Learning to GenAI
22mJustin Basilico, Research/ Engineering Manager, Netflix
43mLinux Container Performance Analysis
37mAsync Inference at Netflix: Batch and Near-Real-Time Serving
17mDelivering better video encodes for legacy devices
40mOptimize Your AI Cloud Infrastructure: A Hardware Perspective
48mFantastic Performance and Where to Find It