Enhancing Network Reliability in Software Defined Networks: Controller Placement and Fast Reroute Mechanisms

Abstract

Modern network infrastructures are undergoing a major transformation driven by Software Defined Networking (SDN). However, migration from legacy hardware to fully programmable architectures is typically incremental, resulting in hybrid environments where legacy routing protocols and centralized SDN controllers coexist. Managing these heterogeneous networks requires coordinated optimization across the physical infrastructure, control plane, and data plane. This thesis presents a multi-layer optimization framework for planning, deploying, and operating homogeneous and hybrid SDN environments. At the infrastructure layer, the Controller Placement Problem (CPP) is formulated as a multi-objective Integer Linear Programming (ILP) model and solved using exact ILP solvers and a localized Tabu Search approach. The framework determines controller placement and quantity to maximize network centrality and throughput while minimizing deployment cost and propagation delay. The ILP model reduces propagation delay by 16.4\% and 24.1\%, while the localized Tabu Search further improves transmitted data by 15.6\% and 26.2\% for the selected topology. At the control-plane layer, the thesis addresses protocol heterogeneity and route redistribution across administrative boundaries. Five routing protocols---BGP, EIGRP, IS-IS, OSPF, and RIP---are evaluated in terms of round-trip time (RTT), convergence delay, and a redistribution feasibility index(capturing topology compatibility, load sensitivity, and link stability). The optimization results show that while EIGRP provides strong proprietary performance, IS-IS emerges as the most resilient open-standard protocol for hybrid control planes. At the data-plane layer, the thesis develops two port-state-aware Fast Reroute (FRR) mechanisms for unpredictable link failures: PSA-FRR, a proactive rule-based approach for homogeneous networks, and PSAR-FRR, an automated deep neural network approach for hybrid environments. The neural model maps real-time interface status (port status) directly to backup egress paths using a formulated traffic engineering dataset. Experiments on the Abilene topology using Mininet, Ryu, OpenDaylight, and GNS3 show that both approaches restore traffic within 30--100~ms. PSAR-FRR achieves a data-plane switching latency of 0.123~ms, more than 70\% lower than PSA-FRR lookup latency and faster than the other evaluated machine learning methods. Overall, this thesis provides an end-to-end mathematical and machine learning framework for designing dependable, low-latency, and scalable SDN infrastructures.