부산대학교 도서관

Multi-access edge computing (MEC) aims to execute mobile cloud services in edge servers located near end-users to provide higher Quality of Experience (QoE). Centralised methods for dynamic service placement in MEC require central access to control every server, which is challenging where servers belong to different administrative domains. Other approaches use distributed decision-making, but most of them do not consider cooperation between servers. Dynamic, distributed service placement can potentially provide workload orchestration and higher QoE. However, this is challenging where service demand patterns are non-stationary. This paper proposes a multi-armed bandit-based method where servers cooperatively decide to host service replicas by applying reinforcement learning that uses service characteristics and context information to minimise response time and backhaul traffic. By sharing cached services, the servers achieve better workload distribution while autonomously making placement decisions. The simulations demonstrate improvements in response time and reduction in backhaul traffic compared to close baselines.