Provisioning with Fine-grained Affinity for Container-enabled Cloud-edge System

Abstract

Containerization effectively decouples the applications and facilitates those user-defined dependencies for extreme elasticity via existing container management frameworks. However, the variance occurs among container-enabled hosts at edges, since their copies of images and libraries, and available resources are quite different. Then, even for the same container, the costs of retrieving the dependency also vary over hosts, since existing mechanism forces each host to retrieve the missing parts from the central repository only upon its own status. Although previous researches have optimized the management of dependencies, the deployment mainly focuses on the entire image. Instead, upon the fine-grained difference on dependencies, we propose to minimize the overall latency of both retrieving the missing parts and the executions under constrained resources. We propose an integer linear program and then design a randomized algorithm to select suitable hosts to deploy the containers. The probability calculated in algorithm shows the preference and the affinity of hosts, and is derived from the status of fine-grained dependencies. Via rigorous proof, the result of our algorithm is concentrated on its optimum with high probability. Extensive simulations confirm that our algorithm outperforms the alternatives up to 44% regarding the average completion time of containers.