Network SLO-aware container scheduling in Kubernetes

In clouds, various services run on respective containers and have service-level objectives (SLO) that significantly impact service qualities. However, Kubernetes, a widely used container orchestration platform, does not schedule containers with respect to the network SLOs. This paper proposes a new container scheduling technique consisting of a cloud-level and node-level scheduler. The cloud-level scheduler selects a node that is best suited for satisfying the network SLO, and the node-level scheduler adjusts the CPU allocation for the container to satisfy SLOs on the selected node. We implement the cloud-level scheduler in Kubernetes and the node-level scheduler in the Linux kernel module and evaluate them using simulation and actual deployment. The evaluation results show that the cloud-level scheduler reduces the scheduling overhead by 22× compared to DRF, a representative multi-resource scheduling technique. Also, the node-level scheduler increases the number of containers that satisfy SLOs by 2.5× compared to native Kubernetes, which will significantly enhance the service quality of user-facing services.