Self-Programming Networks: Network Clock Synchronization, Event Playback and Real-time Action
summary
Claims that a NIC-based probe mesh (each server probing 10–15 random peers) plus accurate NIC timestamps and LASSO plus probe mesh reconstruction recovers millisecond-averaged per-switch queue lengths in a 128-server testbed, preserves ~97% signal power (one millisecond avg for 10 Gb links), and keeps clock sync under ~40 ns at load. A shallow neural net learns LASSO outputs and accelerates reconstruction by ~1e3–1e4×, enabling near-real-time (~1 ms) telemetry and a live dashboard demo showing per-flow attribution.
AI-written summary · about
chapters (12)
- 0:00Introduction to research
- 1:36Data center infrastructure
- 3:40Network sensing and control
- 6:47NIC-based sensing approach
- 9:28Reconstruction pipeline
- 10:37Lasso algorithm for inference
- 15:27Algorithm validation
- 19:01High-fidelity reconstruction
- 26:51Machine learning acceleration
- 30:32Clock synchronization
- 36:07Loop-based correction
- 39:36System dashboard demo
related talks
26mExploiting a Natural Network Effect for Scalable, Fine-grained Clock Synchronization
27mSIMON: A Simple and Scalable Method for Sensing, Inference and Measurement in...
28mRock around the clock (synchronization
21mTime Synchronization using ML Techniques in Electronic Trading
17mHow to diagnose nanosecond network latencies in rich end-host stacks
26mSoNIC: Precise Realtime Software Access and Control of Wired Networks
19mAI/ML Networking Challenges: The Fast and the Finicky!
32mActive Fault Finding in Networks