Machine learning inside the network
ML inference running directly inside a programmable network switch, bringing market prediction closer to the data.
Paper
In-Network Market Prediction Using Machine Learning and Limit Order Books
Authors
Xinpeng Hong, Changgang Zheng, Joshua Lilley, Stefan Zohren, Noa Zilberman
In-Network Market Prediction Using Machine Learning and Limit Order Books presents LOBIN, a system that builds a limit order book and runs ML inference directly on a programmable network switch.
Implemented in P4 (a programming language for defining how network switches process and forward packets) on an Intel Tofino switch, LOBIN processes market-by-order data in the network data plane rather than sending it to a server. The system supports tree-based models operating on three levels of the order book.
The switch achieves microsecond-level inference latency and processes up to 3.2 Tbps. A hybrid design routes high-confidence predictions through the switch and sends uncertain cases to a larger server-side model, processing around 45% of market-data traffic on the switch in the authors' experiments.
