Adaptive and Scalable Controller Placement in Software-Defined Networking

Adewale, Adekoya Oladipupo (2023) Adaptive and Scalable Controller Placement in Software-Defined Networking. Doctoral thesis, Birmingham City University.

Adekoya Oladipupo Adewale PhD Thesis published_Final version_Submitted May 2023_Final Award Oct 2023.pdf - Accepted Version

Download (9MB)


Software-defined networking (SDN) revolutionizes network control by externalizing and centralizing the control plane. A critical aspect of SDN is Controller Placement (CP), which involves identifying the ideal number and location of controllers in a network to fulfill diverse objectives such as latency constraints (node-to-controller and controller-controller delay), fault tolerance, and controller load. Existing optimization techniques like Multi-Objective Particle Swarm Optimisation (MOPSO), Adapted Non-Dominating Sorting Genetic Algorithm-III (ANSGA-III), and Non-Dominating Sorting Genetic Algorithm-II (NSGA-II) struggle with scalability (except ANSGA-III), computational complexity, and inability to predict the required number of controllers. This thesis proposes two novel approaches to address these challenges. First, an enhanced version of NSGA-III with a repair operator-based approach (referred to as ANSGA-III) is introduced, enabling efficient CP in SD-WAN by optimizing multiple conflicting objectives simultaneously. Second, a Stochastic Computational Graph Model with Ensemble Learning (SCGMEL) is developed, overcoming scalability and computational inefficiency associated with existing methods. SCGMEL employs stochastic gradient descent with momentum, a learning rate decay, a computational graph model, a weighted sum approach, and the XGBoost algorithm for optimization and machine learning. The XGBoost predicts the number of controllers needed and a supervised classification algorithm called Learning Vector Quantization (LVQ) is used to predict the optimal locations of controllers. Additionally, this research introduces the Improved Switch Migration Decision Algorithm (ISMDA) as part of the holistic contribution. ISMDA is implemented on each controller to ensure even load distribution throughout the controllers. It functions as a plug-and-play module, periodically checking if the load surpasses a certain limit. ISMDA improves controller throughput by approximately 7.4% over CAMD and roughly 1.1% over DALB. ISMDA also outperforms DALB and CAMD with a decrease of 5.7% and 1%, respectively, in terms of controller response time. Additionally, ISMDA outperforms DALB and CAMD with a decrease of 1.7% and 5.6%, respectively, in terms of the average frequency of migrations. The established framework results in fewer switch migrations during controller load imbalance. Finally, ISMDA proves more efficient than DALB and CAMD, with an estimated 1% and 6.4% lower average packet loss, respectively. This efficiency is a result of the proposed migration efficiency strategy, allowing ISMDA to handle higher loads and reject fewer packets.

Real-world experiments were conducted using the Internet Zoo topology dataset to evaluate the proposed solutions. Six objective functions, including worst-case switch-to-controller delay, load balancing, reliability, average controller-to-controller latency, maximum controller-to-controller delay, and average switch-to-controller delay, were utilized for performance evaluation. Results demonstrated that ANSGA-III outperforms existing algorithms in terms of hypervolume indicator, execution time, convergence, diversity, and scalability. SCGMEL exhibited exceptional computational efficiency, surpassing ANSGA-III, NSGA-II, and MOPSO by 99.983%, 99.985%, and 99.446% respectively. The XGBoost regression model performed significantly better in predicting the number of controllers with a mean absolute error of 1.855751 compared to 3.829268, 3.729883, and 1.883536 for KNN, linear regression, and random forest, respectively. The proposed LVQ-based classification method achieved a test accuracy of 84% and accurately predicted six of the seven controller locations.

To culminate, this study presents a refined and intelligent framework designed to optimize Controller Placement (CP) within the context of SD-WAN. The proposed solutions effectively tackle the shortcomings associated with existing algorithms, addressing challenges of scalability, intelligence (including the prediction of optimal controller numbers), and computational efficiency in the pursuit of simultaneous optimization of multiple conflicting objectives. The outcomes underscore the supremacy of the suggested methodologies and underscore their potential transformative influence on SDN deployments. Notably, the findings validate the efficacy of the proposed strategies, ANSGA-III and SCGMEL, in enhancing the optimization of controller placement within SD-WAN setups. The integration of the XGBoost regression model and LVQ-based classification technique yields precise predictions for both optimal controller quantities and their respective positions. Additionally, the ISMDA algorithm emerges as a pivotal enhancement, enhancing controller throughput, mitigating packet losses, and reducing switch migration frequency—collectively contributing to elevated standards in SDN deployments.

Item Type: Thesis (Doctoral)
15 May 2023Submitted
4 October 2023Accepted
Uncontrolled Keywords: Controller Placement, SDWAN, Stochastic Computational Graph, ISMDA, ANSGA-III, and XGBoost
Subjects: CAH11 - computing > CAH11-01 - computing > CAH11-01-01 - computer science
Divisions: Doctoral Research College > Doctoral Theses Collection
Faculty of Computing, Engineering and the Built Environment > School of Computing and Digital Technology
Depositing User: Jaycie Carter
Date Deposited: 14 Nov 2023 10:52
Last Modified: 14 Nov 2023 10:52

Actions (login required)

View Item View Item


In this section...