pForest: In-Network Inference with Random Forests.

摘要: The concept of "self-driving networks" has recently emerged as a possible solution to manage the ever-growing complexity modern network infrastructures. In self-driving network, devices adapt their decisions in real-time by observing traffic and performing in-line inference according machine learning models. recent advent programmable data planes gives us unique opportunity implement this vision. One open question though is whether these are powerful enough run such complex tasks? We answer positively presenting pForest, system for in-network supervised models on top planes. key challenge design classification that fit constraints (e.g., no floating points, loops, limited memory) while providing high accuracy. pForest addresses three phases: (i) it optimizes features selection capabilities devices; (ii) trains random forest tailored different phases flow; (iii) applies real time, per-packet basis. fully implemented Python (training), P4_16 (inference). Our evaluation shows can classify at line rate hundreds thousands flows, with an accuracy on-par software-based solutions. We further show practicality deploying existing hardware (Barefoot Tofino).