OPTIMIZATION OF CONGESTION CONTROL IN SOFTWARE-DEFINED NETWORKS (SDN) BASED ON REINFORCEMENT LEARNING

• Main Author: Hendarto Fajar Nugroho, Tri
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/87638
• Institution: Institut Teknologi Bandung

Software-Defined Networking (SDN) technology has emerged as a solution for network management, as it enables centralized network programmability and control. Congestion is one of the key factors contributing to the degradation of network performance (Quality of Service). Several techniques have been employed to address congestion, including load balancing, Johnson’s algorithm, Dijkstra’s algorithm, routing algorithms, and Reinforcement Learning (RL). RSCAT is one of the congestion control approaches that utilizes Reinforcement Learning to detect and take action when congestion occurs. However, challenges remain in handling dynamic traffic patterns and optimizing network performance. The RL+ algorithm presents a solution to these challenges. This algorithm was compared to Dijkstra’s algorithm in simulation testing using Mininet to evaluate delay, jitter, and throughput for both TCP and UDP protocols. The evaluation was conducted using D-ITG and Iperf3 tools under different network load conditions. The results indicate that the RL+ algorithm is more effective in mitigating congestion than Dijkstra’s algorithm. In mesh topology, RL+ achieved a delay reduction of up to 34.4% compared to Dijkstra’s algorithm, particularly under high payload scenarios. In the scenario without background traffic, RL+ achieved a TCP delay of 4.973 ms, a jitter of 0.0091 ms, and a throughput of 93.9 Mbps at 75% payload in the presence of background traffic. In terms of throughput, RL+ demonstrated a higher throughput of up to 94.2 Mbps, whereas Dijkstra’s algorithm experienced a significant throughput drop, reaching only 57.7 Mbps under high payload conditions, with a higher packet loss rate compared to RL+.