Framework for Analysing a Policy-driven Multi-Tenant Kubernetes Environment

Abstract

Kubernetes (K8s) is gaining adoption in cloud computing for container management, deployment automation, and resource scheduling. As K8s matures, with increased stability and scalability, it is important to study how it can be effectively customized for use in different application scenarios. The focus of our work is on studying one of its main core components, kube-scheduler, which is in charge of scheduling pods on worker nodes. The K8s default scheduler implements the First Come First Serve (FCFS) algorithm as the pods are ordered and sequenced for execution based on the timestamp of when tasks arrive, when no priority is set to the pods. In this paper, we present a Policy-driven Multi-Tenant K8s (PMK) framework to study how policies of multiple tenants on resource requests and job arrivals affect fairness for the tenants individually in terms of makespan, average waiting time, and average turnaround time. PMK allows re-sequencing of tasks, submitted by multiple tenants, via well-known or customized scheduling algorithms before they enter the K8s scheduling queue. Our evaluation uses well-known algorithms such as Round Robin (RR), FCFS and Dominant Resource Fairness (DRF). In addition, we introduce a Cluster-Based Fairness (CBF) scheduling algorithm, which is a variation of DRF. CBF considers overall cluster utilization and resource availability to determine which task to choose from new requests. Our analysis shows that PMK can provide insights to cluster and cloud infrastructure managers on the factors affecting fairness and accordingly in some cases obtain 61.0% improvement in average waiting time for tenants with homogeneous individual demands. In addition, our customized CBF scheduling policy, when used with with PMK on K8s, can reduce overall cluster average waiting time by up to 4%.