Brain-on-Switch: Towards Advanced Intelligent Network Data Plane via NN-Driven Traffic...
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
Implements a prototype on P4 Barefoot Tofino and claims line‑speed neural inference by running a binarized RNN with sliding‑window recurrence entirely on the data plane while offloading only ambiguous flows to an off‑switch Transformer via DPDK+GPU. System escalates under 5% of flows, classifies over 95% of flows on‑switch, and reports average F1 gains of 133% and 31% versus NetBeacon and NN‑on‑NIC (up to 177%/39% on harder tasks). Uses ~20% SRAM and ~1% TCAM, and end‑to‑end IMS latency stays under two seconds for typical loads.
AI-written summary · about
chapters (9)
- 0:00Brain-on-Switch project
- 1:19Traditional traffic analysis limits
- 3:06RNNs on switch motivation
- 5:14Brain-on-Switch design goals
- 6:15On-switch RNN architecture
- 8:47Ambiguous flow escalation
- 10:15Off-switch Transformer inference
- 11:00Experimental evaluation results
- 13:10Conclusion and future
related talks
14mRe-architecting Traffic Analysis with Neural Network Interface Cards
13mOptimizing Dynamic Neural Networks with Brainstorm
13mFENIX: Enabling In-Network DNN Inference with FPGA-Enhanced Programmable Switches
16mAccelerating Distributed Graph Learning by Using Collaborative In-Network...
15mEnabling In-situ Programmability in Network Data Plane: From Architecture to Language
01h 11mMachine Learning Acceleration in an Embedded World
01h 54mDNNs on FPGAs
16mA High-Speed Stateful Packet Processing Approach for Tbps Programmable Switches
