Tri-Ring: Asynchronous Service Provisioning with Online Learning in Edge Cloud Networks

Container-based microservices are ideal for service provisioning due to their lightweight deployment and fast startup, making them well-suited for resource-limited edge cloud networks. Furthermore, the inherent layered structure enables efficient layer scheduling and caching to address computing node resource constraints. However, dynamic and time-varying user service requests, coupled with the diversity in microservice container layers, present significant challenges for layer-aware service provisioning in edge cloud scenarios. These challenges include time-exceeded offline service provisioning, tangled microservice orchestration, and layer cache redundancy. In this paper, we introduce Tri-Ring, an asynchronous service provisioning framework with online learning in edge cloud networks. This framework addresses a joint optimization problem across three time scales, aiming to maximize the utility of edge cloud nodes. We formulate service provisioning as a mixed-integer nonlinear programming (MINLP) problem, which can be transformed into a linear programming (LP) subproblem and submodular optimization subproblem, solved using the estimator-assessor algorithm. The proposed framework is thoroughly evaluated using real-world datasets. The results demonstrate that Tri-Ring outperforms other baselines in service provisioning, increasing utility by 22.31%. Additionally, it reduces microservice startup time by 63.35% and optimizes storage resources by 30.07%.