Leo: Online ML-based Traffic Classification at Multi-Terabit Line Rate
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
System implements decision trees in the switch data plane and claims packet-by-packet classification at multi‑terabit line rate on Intel Tofino by combining a multiplexed ALU abstraction with subtree flattening. Compiler provisions a representative tree for a user‑specified class, allows on‑the‑fly tree swaps with no downtime, and trades table SRAM/TCAM for far fewer pipeline stages. Evaluation shows up to twice the depth of prior systems, F1 ≈ 94% on CICIDS2017, and complete trees to depth 13 (22 with 124 leaves).
AI-written summary · about
chapters (7)
- 0:00Introduction and motivation
- 2:02Programmable switch potential
- 3:12Leo system design
- 6:39Multiplexed ALU abstraction
- 9:36Subtree flattening
- 13:03Evaluation and results
- 17:38Conclusion
related talks
36mOffloading Network Traffic Classification to Hardware
17mSPLIDT: Partitioned Decision Trees for Scalable Stateful Inference at Line Rate
18mLearning-Enhanced High-Throughput Pattern Matching Based on Programmable Data Plane
14mRe-architecting Traffic Analysis with Neural Network Interface Cards
26mEvolution of Application Visibility for Telcos: From Programmable Switches to Cloud Native
18mTiara: A Scalable and Efficient Hardware Acceleration Architecture for Stateful Layer-4
16mA High-Speed Stateful Packet Processing Approach for Tbps Programmable Switches
35mLive Technology Panel