Flow conflict eliminations through machine learning for software defoned network

Abstract

Software-Defined Network (SDN) is a modern approach in networking technologies that enables dynamic and programmatically efficient network configuration for improved performance and network monitoring. Similar to the traditional networks, the SDN system is susceptible to conflicts in flows within the network. Flow conflict in SDN occurs in response to adjustment of certain features of flows such as priority, match field, and action. While efforts have been made to address these challenges, the current flow of conflict solutions in SDN has several limitations. First, the control layer does not show nor collect the conflict flows that are affected in the OpenFlow switch. Second, the flow entry detection and classification process are relatively time-consuming. Third, there are no studies on detection methods to avoid flow conflicts using artificial intelligence methods such as Machine Learning (ML) as a solution to flow conflict in SDN. This thesis aims to eliminate flows conflict in SDN by using ML algorithms to detect and classify all flow conflicts in the OpenFlow switch. This thesis aims to develop the flow construction model in the SDN controller, detect the conflict flow using ML algorithm, and classify all the conflict types in the flow table using a classification algorithm. In this work, simulation works were conducted in Mininet software using two types of topologies. Decision trees (DT), support vector machine (SVM), hybrid DT- SVM, and extreme fast decision trees (EFDT) ML algorithms were used to detect the conflicts. The main contribution of this thesis is the development of a flow construction model with conflict rules in the OpenFlow table that enhanced the SDN process. By using accurate and effective ML algorithms designed and implemented in the controller layer, flow conflicts are detected and classified to reduce the adverse effects of conflict in the SDN. The performance of the proposed algorithms was evaluated for their efficiency and effectiveness across a variety of evaluation metrics. The EFDT algorithm produced the best results with a performance accuracy above 90% and 95% in detection and classification respectively for all sizes of flows between 1,000 and 100,000. The proposed algorithms for detection and classification show performance improvements over two different algorithms used as benchmarks.