ML-QTLB: A Hybrid Machine Learning and Queueing-theory Model for Server-side Load Balancing in SDN

Abstract

Software-Defined Networking (SDN) enables centralized and programmable network management, making it a key enabler for intelligent load balancing in modern network infrastructures.However, existing load balancing approaches rely on static algorithms or instantaneous resource metrics that fail to capture the dynamic queuing behaviour of server traffic, leading to suboptimal response times, uneven load distribution, and starvation of lowpriority flows.This paper proposes ML-QTLB, a hybrid Machine Learning and Queueing-Theory based Load Balancing model for server-side dynamic load balancing in SDN environments.ML-QTLB constructs a dataset of 25,000 records from Python-generated SDN traffic (1 MB, 10 MB, 25 MB, 50 MB, and 75 MB) captured using Wireshark on Linux within a Mininet-based testbed.A three-stage preprocessing pipeline (data cleaning; feature selection, reducing 15 to 8 features; and min-max normalization) is applied, followed by Fuzzy C-Means (FCM) clustering with K=2 determined by Silhouette analysis and Random Forest (RF) classification, achieving 99.88% accuracy.An M/M/1-based QT scheduler selects the server with the minimum expected system time at each routing epoch, and an aging-based threshold mechanism prevents starvation of low-priority flows.Evaluated against four classical algorithms (RR, Random, WRR, and LC), ML-QTLB reduces average response time by an overall mean of 33-40% and improves the degree of load balancing by an overall mean of 21-30% across all tested traffic loads.Evaluated against five server-metric-based algorithms (LRAM, LCPU, LCPURAM, LC, and LCLCPURAM), ML-QTLB reduces waiting time by an overall mean in the range of 87-94% and service time by an overall mean in the range of 85-94%.The results confirm that integrating data-driven ML classification with mathematically grounded queueing theory yields substantial and consistent performance improvements over both classical and resource-metricbased load balancing approaches in SDN environments.