Enhancing Video Transmission with Machine Learning based Routing in Software-Defined Networks

TL;DR

This paper proposes an AI-based routing algorithm for SDN to optimize video transmission, utilizing machine learning techniques and a newly created dataset to improve network performance in congested environments.

Contribution

Introduces a novel AI-driven routing method in SDN for video transmission, supported by a custom dataset and multiple machine learning models for traffic prediction.

Findings

Machine learning models effectively predict traffic levels.

AI routing improves video quality in congested networks.

New dataset enables better evaluation of network performance.

Abstract

Our study uses the centralized, flexible, dynamic, and programmable structure of Software-Defined networks (SDN) to overcome the problems. Although SDN effectively addresses the challenges present in traditional networks, it still requires further enhancements to achieve a more optimized network architecture. The Floodlight controller utilized in this study employs metrics such as hop count, which provides limited information for routing. In scenarios such as video transmission, this situation is insufficient and the need for optimization arises. For this purpose, an artificial intelligence (AI) based routing algorithm is proposed between the server and the client in the scenario based on NSFNET topology. The topology designed with the Floodlight controller in the Mininet simulation environment includes a client, a server, and 14 switches. A realistic network environment is provided by adding different receivers and creating TCP traffic between these receivers using the iperf3 tool. In three scenarios, video streaming is performed using the FFmpeg tool, and 49 path metrics such as RTT, throughput, and loss are recorded. In these scenarios, PSNR and SSIM calculations are made to observe the differences between the transmitted and the original video in congested and uncongested environments. Due to the lack of a dataset suitable for the proposed network environment in the literature, a new dataset consisting of 876 records is created using continuously transmitted video traffic. Low and high traffic levels are created within the dataset, and different machine learning techniques such as KNN, Random Forest, SVM, AdaBoost, Logistic Regression and XGBoost are applied using the features that affect the traffic levels.