Bin Wang, Ph.D. (Advisor); Phu Phung, Ph.D. (Committee Member); Krishnaprasad Thirunarayan, Ph.D. (Committee Member); Yong Pei, Ph.D. (Committee Member)
Doctor of Philosophy (PhD)
Network traffic has been growing exponentially due to the rapid development of applications and communications technologies. Conventional routing protocols, such as Open-Shortest Path First (OSPF), do not provide optimal routing and result in weak network resources. Optimal traffic engineering (TE) is not applicable in practice due to operational constraints such as limited memory on the forwarding devices and routes oscillation. Recently, a new way of centralized management of networks enabled by Software-Defined Networking (SDN) made it easy to apply most traffic engineering ideas in practice. \par Toward creating an applicable traffic engineering system, we created a TE simulator for experimenting with TE and evaluating TE systems efficiently as this tool employs parallel processing to achieve high efficiency. The purpose of the simulator is two aspects: (1) We use it to understand traffic engineering, (2) we use it to formulate a new traffic engineering algorithm that is near-optimal and applicable in practice. We study the design of some important aspects of any TE system. In particular, the consequences of achieving optimal TE by solving the multi-commodity flow problem (MCF) and the consequences of choosing single-path routing over multi-path routing. With the help of the TE simulator, we compare many TE systems constructed by combining different paths selection techniques with two objective functions for rate adaptations: load balancing (LB) and average delay (AD). The results confirm that paths selected based on the theoretical approach known as Oblivious Routing combined with AD objective function can significantly increase the performance in terms of throughput, congestion, and delay.\par However, the new proposed system comes with a cost. The AD function has a higher complexity than the LB function. We show that this problem can be tackled by training deep learning models. We trained two models with two different neural network architectures: Multilayer Perceptron (MLP) and Long-Short Term Memory (LSTM), to get a responsive traffic engineering system. The input training data is based on synthetic data obtained from the simulator. The output of the two models is the split ratios that the SDN controller uses to instruct the switching devices about how to forward traffic in the network. The result confirms that both models are effective and can be used to forward traffic in an optimal or near-optimal way. The LSTM model has shown a slightly better result than MLP due to its ability to predict a longer output sequence.
Department or Program
Department of Computer Science and Engineering
Year Degree Awarded
Copyright 2022, some rights reserved. My ETD may be copied and distributed only for non-commercial purposes and may not be modified. All use must give me credit as the original author.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 3.0 License.