Performance evaluation of load balancing algorithms in software-defined networks under dynamic traffic conditions

Bursty arrivals, heterogeneous server capacities, and mixed mice–elephant workloads make software-defined network (SDN) load balancing difficult when dispatch decisions rely on a single load indicator. This study proposes an Adaptive Multi-Metric Load Balancer (AMM-LB), a deterministic, training-free scheme that integrates capacity-normalized connection load, backlog-derived response pressure, congestion-adaptive weighting, and flow-aware adjustment. AMM-LB was evaluated using a Python event-driven simulator with heterogeneous finite-buffer servers and two-state Markov-Modulated Poisson Process traffic, with robustness checked against matched Poisson arrivals. Seven algorithms were compared across offered loads from ρ=0.4 to 1.3 using 12 paired replications and formal statistical testing. At ρ=1.0, AMM-LB achieved the lowest mean latency (176.6 ms) and flow-loss rate (2.71%), representing reductions of 42.4% and 67.1% relative to Random, while maintaining 3547.2 Mbps goodput and 0.9985 Jain fairness. Mean-latency and loss improvements were significant against all baselines after Holm correction. Shortest-Response achieved lower mice-flow p99 latency, revealing an average-delay versus tail-latency trade-off. Ablation, sensitivity, stale-state, flow-classification, traffic-model, and 6–48-server scalability analyses further established performance boundaries. With O(N) decision complexity, AMM-LB provides an interpretable alternative for bursty heterogeneous SDN workloads without offline training.