Distributed Spatial-Temporal Demand and Topology Aware Resource Provisioning for Edge Cloud Services

Abstract

Current edge cloud providers offer a wide range of on-demand private and public cloud services for customers. Predictive demand monitoring and supply optimisation are necessary to deliver truly elastic distributed edge cloud services with resizable resource and compute capacity to adapt to dynamically changing customer requirements. However, current state-of-the-art monitoring and provisioning systems remain reactive which often results in over or under service provisioning, incurring unnecessary costs for customers or deterioration in the quality of service for the end-user. This paper proposes an adaptive protocol, ARPP, that enables distributed real-time demand monitoring and automatic resource provision based on the dynamically changing spatial-temporal workload patterns. ARPP leverages distributed predictive analytics and deep reinforcement learning at the edges to predict the dynamically changing spatial-temporal demand and allocate the appropriate amount of resources at the right times and right locations. We show that ARPP outperforms benchmark and state of the art algorithms across a range of criteria in the face of dynamically changing mobile real-world topologies and user interest patterns.