Enabling Seamless Integration of ML-Based Network Functions into the Network Dataplane

Abstract

The current network device architecture lacks a comprehensive framework for deploying machine learning (ML) network functions in the data plane. This letter presents a primitive-based ML framework for deploying network functions directly onto the programmable dataplane. The ML primitives are introduced as modular building blocks that enable software-to-hardware model translation. We demonstrate the framework’s functionality by defining primitives for an Artificial Neural Network model. Additionally, a two-stage approximation–model pruning and hardware-aware primitive tuning–reduces the implementation complexity of the ML model. The resulting implementation maintains inference accuracy and resource efficiency, making it suitable for resource-constrained data plane environments. © 2019 IEEE.