Containerized Service Placement and Resource Allocation at Edge: A Hybrid Reinforcement Learning Approach

Abstract

Container has already become a default and prevalent solution due to its efficient and easy-to-deploy in edge computing. However, constrained resources in edge nodes may introduce significant deployment costs and increase service response latency in containerized services. Existing studies mainly focus on optimizing container placement strategies, while largely overlooking computational resources configuration. To tackle this problem, we introduce a joint optimization approach for containerized service placement and computational resources configuration from the perspective of image layer sharing. Specifically, we define a profit-driven mixed integer nonlinear programming (MINLP) problem and propose a graph-aware hybrid reinforcement learning (GAHRL) algorithm. By capturing inter-layer sharing dependencies and edge resource distribution, our algorithm optimizes containerized service placement while ensuring efficient computational resources configuration. Extensive experimental results show that the proposed algorithm outperforms other baseline algorithms in maximizing profits as well as reducing service delays and deployment costs.

Keywords: edge computing, containerized service placement, resources allocation, reinforcement learning.