Latency‐ and Load‐Aware MEC Host Selection in Simu5G: A Lightweight Orchestration Policy for Delay‐Sensitive 5G Applications

Multiaccess edge computing (MEC) reduces end‐to‐end (E2E) latency by offloading computation from user equipment (UE) to nearby edge hosts, but host‐selection policies in simulators such as Simu5G typically focus on service availability or CPU headroom. They do not explicitly account for dynamic UE–host latency and can therefore steer delay‐sensitive flows to topologically distant or temporarily congested hosts, inflating tail delay and limiting effective throughput. We extend Simu5G's ETSI‐compliant MEC orchestrator with LatencyAwareSelectionBased, a lightweight host‐selection policy that jointly accounts for (i) a dynamic UE–host delay estimate derived from measurable run‐time delay statistics and smoothed via an exponential moving average and (ii) the host's normalized CPU load. These two signals are combined in a simple score, S(h)=L˜(h)1+αC˜(h)$$ S(h)=\tilde{L}(h)\kern0.3em \left(1+\alpha \kern0.3em \tilde{C}(h)\right) $$, controlled by a single trade‐off parameter α$$ \alpha $$. The policy is implemented as a plug‐in selection module activated through Simu5G's native selectionPolicy hook, requiring only minimal changes to existing scenarios and preserving the standard orchestration pipeline. We evaluate the proposed policy in Simu5G's MultiMecHost scenario under three operating points (Best, Moderate, and Worst) and an additional multitenant burst configuration with up to 20 UEs generating bursty arrivals. Using Simu5G scalar outputs, we take downlink MAC delay and per‐UE throughput as primary metrics and also examine CPU‐only, latency‐only, and mixed‐score variants for different values of α$$ \alpha $$. Across all configurations, the latency‐aware policy consistently maintains low tail latency (95th‐percentile downlink MAC delay below approximately 7ms$$ 7\kern0.3em \mathrm{ms} $$) while improving per‐UE throughput relative to Simu5G's default CPU‐ and service‐based strategies and avoiding persistent host hot spots. The resulting design provides a simple, reproducible baseline for latency‐sensitive MEC orchestration and a practical integration point for future mobility‐aware and learning‐based host‐selection schemes.