Loading...

Resource Management for Machine Learning-Based Network Functions in Programmable Data Plane

Babaei, Fatemeh | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58577 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Dolati, Mahdi
  7. Abstract:
  8. This study investigates the direct integration of Binary Neural Networks (BNNs) within network infrastructure. With the increasing volume of data traffic, online services, the Internet of Things, and real-time processing, the execution of advanced functions such as anomaly detection and network security has become essential. Hardware-based solutions are fast but costly and inflexible, while software-based solutions are flexible but suffer from high latency. Programmable infrastructures offer an efficient and adaptable alternative, and BNNs are particularly suitable for deployment within the data path due to their lightweight nature, which utilizes binary weights and activations, and consumes minimal resources. However, the hardware limitations of switches pose a challenge for the direct execution of machine learning models. The proposed approach introduces a two-stage framework. In the first stage, BNN models are trained in a central controller using datasets related to network functions. Once training is completed, the model’s weights and layers are distributed among programmable switches. In the second stage, i.e., the inference phase, these weights are deployed on the switches, and the trained model is executed using the Mininet simulation environment and the P4 programming language, enabling inline and simultaneous packet processing as traffic flows through the network. To achieve this, an optimization model based on integer linear programming is designed to allocate resources efficiently and map neural network layers onto the switches. The implementation demonstrates that even under severe resource constraints on network devices, efficient deployment of machine learning models within the data path is feasible. The results confirm that the proposed framework not only improves resource utilization but also ensures the stable execution of network functions. Furthermore, comparisons with baseline algorithms such as First-Fit and greedy heuristics indicate that, in small-scale scenarios (fewer than ten BNNs), the proposed method achieves an average improvement of 2 to 4 times in resource utilization and inter-layer delay reduction compared to the greedy approach. In larger-scale scenarios (more than ten BNNs), these improvements exceed twelvefold. However, compared to the First-Fit method, the proposed approach performs significantly worse than smoothing and greedy strategies in all scenarios. From this perspective, the present study offers a practical and innovative solution for directly integrating AI capabilities into modern networks and paves the way for designing a new generation of intelligent, scalable, and adaptable networks
  9. Keywords:
  10. Integer Linear Programming ; P4 Programming Language ; Programmable Network ; Binary Neural Network (BNN) ; Mininet Simulation Environment ; Network Functions

 Digital Object List

 Bookmark

...see more