In-Network ML project: Planter
Planter: Rapid Prototyping of In-Network Machine Learning Inference
Using programmable network devices to aid in-network machine learning provides high throughput and low latency. Still scaling to multiple models, targets and use-cases is hard,
Planter is a modular framework for fast prototyping of trained machine learning models to programmable devices.
Provided with a configuration file and a dataset, Planter automatically offloads machine learning classification task into a programmable data plane.
Planter supports 11 machine learning models (with 52 mapping variations), 5 targets, 4 use cases, 14 datasets, and can be easily extended.
Planter achieves better machine learning performance and significantly lower resource consumption than previous model-tailored works, while co-existing with network functionality and maintaining line rate, providing billions of classification decisions per second.
Contributions by Changgang Zheng, Xinpeng Hong, Riyad Bensoussane, Liam Perreault, Noa Zilberman (Oxford), Mingyuan Zang (DTU), Shay Vargaftik and Yaniv Ben-Itzhak (VMWare).
