Geo-distributed efficient deployment of containers with Kubernetes

Software containers are changing the way applications are designed and executed. Moreover, in the last few years, we see the increasing adoption of container orchestration tools, such as Kubernetes, to simplify the management of multi-container applications. Kubernetes includes simple deployment policies that spread containers on computing resources located in the cluster and automatically scale them out or in based on some cluster-level metrics. As such, Kubernetes is not well-suited for deploying containers in a geo-distributed computing environment and dealing with the dynamism of application workload and computing resources. To tackle the problem, in this paper we present ge-kube (Geo-distributed and Elastic deployment of containers in Kubernetes), an orchestration tool that relies on Kubernetes and extends it with self-adaptation and network-aware placement capabilities. Ge-kube introduces flexible and decentralized control loops that can be easily equipped with different deployment policies. Specifically, we propose a two-step control loop, in which a model-based reinforcement learning approach dynamically controls the number of replicas of individual containers on the basis of the application response time, and a network-aware placement policy allocates containers on geo-distributed computing resources. To address the placement issue, we propose an optimization problem formulation and a network-aware heuristic, which explicitly take into account the non-negligible network delays among computing resources so to satisfy Quality of Service requirements of latency-sensitive applications. Using a surrogate CPU-intensive application and a real application (i.e., Redis), we conducted an extensive set of experiments, which show the benefits arising from the combination of elasticity and placement policies, as well as the advantages of using network-aware placement solutions.