부산대학교 도서관

Containers technology has become very popular in recent years, since it allows users to focus on designing their applications in a modular way and abstracting away the environments in which they actually run. Cloud providers such as AWS (Amazon Web Services) and GCP (Google Cloud Platform) offer their users managed containers platforms that orchestrate, schedule and execute multiple containers over a multi-tenant cloud infrastructure. As these services gain popularity, it is becoming more and more challenging to manage them in a way that effectively utilized the existing resources. The latter has a significant economical impact on cloud providers when it comes to their compute infrastructure investment costs and the price they can offer to their customers. In this paper, we approach this challenge by developing multidimensional container resource allocation algorithms designed to be deployed in dynamic cloud environments with different types of applications under varying loads scenarios. Our algorithms allocate for each container an available engine to execute it, in a way that maximizes the overall revenue. We design our algorithms and provide a constant worst-case approximation bound using the Local Ratio technique. Our evaluation, based on real-world scenarios, indicates that the performance of our algorithms is up to a factor of two better than the performance of existing scheduling algorithms, when the available resources are scarce.