A machine learning framework to implement network function on programmable data plane

Abstract

A novel machine learning framework is proposed in this thesis to enable network function deployment on the programmable data plane. Using a modular and unified approach, the framework finds a minimum set of reusable primitives able to support a wide class of ML models. This work is motivated by the growing demand for intelligent and programmable networks by fusing the machine learning capabilities directly into the network data plane. However, the present network infrastructure lacks hardware architecture and a framework to allow effective development and deployment of ML-based network functions on the hardware. To address this gap, we proposed a modular hardware framework that uses basic building blocks to construct a developed ML model on the hardware. These basic building blocks are known as primitives. In this work, we have illustrated the working of our framework for an ANN and Tree-based model, which are used to build and deploy a network function over the data plane. Furthermore, to reduce computational overhead, the framework incorporates different model approximation strategies. These approximations help to maintain a balance between accuracy and performance, ensuring feasibility in high-speed, resource-constrained network environments. Additionally, the framework is built to be scalable and programmable, such that a myriad of ML architectures can be supported as the framework evolves. This approach not only enhances real-time packet processing capabilities but also opens various new directions for intelligent in-network decision making and autonomous network management. Overall, this work lays a foundational step toward enabling scalable, ML-driven, real-time processing in next-generation programmable network devices.