Re-architecting Traffic Analysis with Neural Network Interface Cards
TalkinItalianat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
The paper argues for replacing traditional supervised ML models for online network traffic analysis with binary neural networks, and introduces N3IC, a system that compiles binary neural nets into implementations that run directly in SmartNIC data planes. N3IC targets multiple hardware platforms by generating micro-C and P4 data-plane code, and evaluations on traffic identification and anomaly detection show up to 100× lower classification latency and 1.5–7× higher throughput while forwarding at 40 Gbps and freeing the host CPU; the authors also present an FPGA-based NIC primitive implementing binary neural network support using under 1–2% of a Virtex-7’s logic and memory.
AI-written summary · about
chapters (7)
- 0:00Neural network NICs intro
- 1:44Traditional traffic analysis fails
- 4:14Traffic classification with neural nets
- 6:44Smart NIC trade-offs
- 9:16Hardware acceleration for real-time traffic
- 12:11Case studies and applications
- 13:57Summary
related talks
14mBrain-on-Switch: Towards Advanced Intelligent Network Data Plane via NN-Driven Traffic...
01h 00mCommSec D1 - An Isolated Data Center Security Policy Model Using SmartNICs
18mNICA: An Infrastructure for Inline Acceleration of Network Applications
17mSRNIC: A Scalable Architecture for RDMA NICs
22mTransparent MultiNIC Routing for Large AI Models
18mLeo: Online ML-based Traffic Classification at Multi-Terabit Line Rate
01h 57mIntroduction to Neuromorphic Computing
30mAccelerating Neural Networks for Vision Systems via FPGAs