SLA-aware cooperative service caching in MEC via graph-attention demand prediction and migration-constrained reinforcement learning

Abstract The MEC infrastructure supports latency-sensitive IoT applications by placing computation and storage resources close to end users. However, limited cache size, unpredictable service demands, mobile users, and cache replacement make adaptive service management difficult. Improper replacement may lead to increased cold-start latencies, migration costs, cloud failover, energy utilization, and SLA breaches. In this paper, we propose ACGSR: Adaptive Cooperative Graph-Attention Service Replacement, which represents a learning-enabled MEC architecture for adaptive service caching and offloading. ACGSR incorporates service demand forecasting using graph attention networks, cache replacement using reinforcement learning, service eviction based on utility maximization, and cooperative edge service identification. The service caching and offloading problem is defined as a normalized multi-objective optimization function, considering service latency, cache hit rate, energy consumption, migration cost, queue stability, and SLA compliance. The graph attention network learns spatial correlations between demands from nearby edge nodes. The reinforcement learning agent chooses suitable caching, replacement, and offloading operations under cache and computation resource constraints. Moreover, cooperative edge identification improves cloud fallback by assisting cache-missing services from adjacent edge nodes. Simulations using an EUA-derived edge-server topology and synthetic service workloads show that ACGSR outperforms CSC-noLM by 30–40%, CDOSC by 24–32%, DRL by 18–24%, and TOSC by 12–18% regarding service latencies. Furthermore, ACGSR boosts cache hit rates by 12–22% and lowers energy utilization by 20–35%.