Explainable Learning-Based Timeout Optimization for Accurate and Efficient Elephant Flow Prediction in SDNs.

Area of Science:

• Computer Science
• Network Engineering

Background:

• Software-defined networks (SDNs) require efficient traffic prediction for optimal performance.
• Current methods using complete traffic data incur significant bandwidth and delay costs.

Purpose of the Study:

• To develop a predictive strategy for elephant flows using incomplete traffic data.
• To reduce control channel overhead and network latency in SDNs.

Main Methods:

• Implementing a flow entry timeout strategy with an initial hard timeout (Tinitial) and a rate of increase (r).
• Utilizing logistic regression for elephant flow modeling and Bayesian optimization for tuning Tinitial and r.
• Employing feature selection, model learning, and optimization on sampled traffic data.

Main Results:

• Achieved over 90% generalization accuracy across diverse datasets (campus, backbone, IoT).
• Successfully predicted elephant flows for approximately 50% of their lifetime.
• Demonstrated significant reduction in controller-switch interactions for campus and IoT networks.

Conclusions:

• The proposed incomplete traffic-based prediction strategy is effective for optimizing SDNs.
• This method offers a viable solution for reducing network overhead while maintaining high prediction accuracy.
• Further considerations for packet completion may be needed in networks with very short packet inter-arrival times.