Loading...

Delay-Aware Routing in Software-Defined Networks by Machine Learning Techniques

Siamaki, Mahdi | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57158 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Hemmatyar, Ali Mohammad Afshin; Safaei, Bardia
  7. Abstract:
  8. In the complex and dynamic world of today's communication networks, Software-Defined Networking (SDN) with its flexibility and programmability has opened new horizons in network management. These networks, by decoupling the control plane from the data plane, enable centralized and efficient management. Despite these advantages, optimizing routing in these networks remains a significant challenge, and increased latency in SDNs due to centralized processing in the controller and communications between switches and controllers is a critical issue that can severely impact the quality of service. Traditional routing methods face numerous challenges due to their static nature and lack of adaptability to the dynamics of modern networks. These methods struggle in implementing complex and priority-based routing policies and face scalability challenges in large networks with high computational loads. Additionally, their inability to learn from past experiences and adapt to dynamic network conditions, along with the inability to predict future traffic and prevent congestion, hampers their efficient deployment in SDNs. This thesis aims to address these issues by proposing a novel intelligent routing framework for SDNs. This framework is based on a two-layer hierarchical architecture utilizing Deep Reinforcement Learning (DRL) algorithms and priority-based sampling, supported by a data structure called Sum-tree to optimize training operations, and incorporates traffic prediction capabilities. Initially, using existing SDN measurement mechanisms such as network information collection and processing, comprehensive network data including key parameters like bandwidth, latency, and packet loss rate are extracted and organized into a traffic matrix. Then, by predicting the future traffic matrix and employing the reinforcement learning algorithm, the optimal routing path is determined based on current and future network conditions. Simulation results demonstrate a significant improvement of the proposed method compared to traditional and policy-based approaches. Specifically, compared to the DRL-TP algorithm, the proposed method has resulted in an 8% reduction in latency, a 33% increase in network throughput, and a 29% decrease in packet loss rate. These improvements signify a substantial enhancement in the quality of service and network efficiency. In addition to quantitative improvements, the proposed approach also shows remarkable performance in convergence speed. This high convergence speed indicates the efficiency and responsiveness of the algorithm in finding optimal solutions in dynamic environments, a vital characteristic for modern networks
  9. Keywords:
  10. Software Defined Networks (SDN) ; Reinforcement Learning ; Service Quality ; Network Traffic ; Computer Network Traffic Prediction ; Routing Optimization

 Digital Object List

 Bookmark

...see more