AI models of video and email traffic for anticipatory networking

Om, Khandu (2022) AI models of video and email traffic for anticipatory networking. PhD thesis, Murdoch University.

Abstract

As the number of applications available over the Internet rapidly grows, providing services with high Quality of Service (QoS) and Quality of Experience (QoE) becomes an increasingly critical issue for wired and wireless networks. Email is one of the most successful applications and has remained a principal communication channel, both in enterprise and social settings. In addition, the demand for multimedia services, especially streaming content such as video and high-definition television, video conferencing, user-generated video and video-based social networking, which are characterized by high burstiness and large bandwidth requirements, has rapidly increased to the point of video becoming the dominant type of network traffic.

The work in this thesis falls into the research area of anticipatory networking. It addresses, via three studies, the need to improve the modelling of email and video traffic towards accurately predicting future traffic loads and hence achieving high bandwidth utilization with improved QoS.

The first focuses on the understanding of email traffic workload properties and patterns. It provides a comprehensive comparison between the performance of recurrent neural networks (RNN) and long short-term memory (LSTM) models and demonstrates that both approaches can achieve high modelling accuracy, outperforming existing work, over four large datasets acquired from different universities’ servers.

To handle the burstiness of video traffic in ways that will improve users’ satisfaction while also achieving high utilization of network resources, prediction of video frame sizes can play a significant role. The second study focuses on accurate traffic prediction for videos encoded with H.264 and H.265, which are major state-of-the-art standards. Our work uses single-step and multi-step approaches to capture the long-range dependence and short-range dependence properties of variable bit rate video traces and evaluates the accuracy of LSTM, Convolutional Neural Networks (CNN) and Sequence-to-Sequence (seq2seq) models and compares them with existing approaches, outperforming them in accuracy for a variety of videos.

The third study uses the work of the second in order to focus on the problem of policing video traffic from H.264 and H.265 sources. Building on work that has shown that classicAI Model of Video and Email Traffic for Anticipatory Networking traffic policing schemes can lead to unnecessarily strict policing for conforming video sources, we propose the use of Artificial Intelligence – based traffic policing schemes for video traffic. We propose mechanisms that are shown to clearly outperform both the widely used token bucket mechanism and other mechanisms from the literature, when used on conforming and non-conforming video users.

Item Type:	Thesis (PhD)
Murdoch Affiliation(s):	IT, Media and Communications
Supervisor(s):	Koutsakis, Polychronis, Wong, Kevin, McGill, Tanya and Dixon, Michael
URI:	http://researchrepository.murdoch.edu.au/id/eprint/65109

Item Control Page

Downloads