A delay-sensitive resource allocation algorithm for container cluster in edge computing environment

Abstract

With the development of lightweight container virtualization technology, it becomes possible to quickly and efficiently deploy software in edge computing environments. When allocating resources to containers in edge nodes, end-to-end latency is an important performance indicator for evaluating strategies. However, it is often difficult to develop a delay model of the container cluster to obtain the real-time end-to-end delay of the service request flow. The state information is always extremely complicated in edge computing environment. Different container service combinations reach edge computing nodes at different time periods. The arrival rate of each container service flow also changes over time. In this paper, we first develop a dynamic M/D/1 queuing model to analyze the end-to-end delay of the data packets of the container service flow and use the average packet delay as the optimization goal of the Edge container resource allocation problem. Then a delay-sensitive resource allocation algorithm based on A3C (Asynchronous Advantage Actor–Critic) is proposed to solve this problem. Finally, we utilize the ESN (Echo state network) to improve the traditional A3C algorithm. Simulation shows that the ESN-based critic A3C(EC-A3C) algorithm has better performance by at least 10.9% than other algorithms in terms of latency and throughput and greatly improves the convergence speed of the network.