A Dynamic I/O Sensing Scheduling Scheme in Kubernetes
Dong Li, Yi Wei, Bing Zeng
Abstract
With the rapid development of the Container-as-a-Service (CaaS), Kubernetes has become the de facto standard for deploying containerized applications on cloud environments. However, the Kubernetes scheduler does not take the disk I/O load of nodes into account, which leads two problems: (1) Multiple I/O-intensive applications may be dispatched to the same node, which cause I/O bottlenecks. (2) Pods are less likely to be scheduled on node with idle I/O and insufficient CPU, resulting in the waste of the node's I/O resource. To address these problems, we first propose a dynamic scheduling algorithm named by Balanced-Disk-IO-Priority (BDI) to improve the disk I/O balance between the nodes. Moreover, we also propose a dynamic scheduling algorithm called Balanced-CPU-Disk-IO-Priority (BCDI) to solve the issue of load imbalance of CPU and disk I/O on single node. The experimental results show that the BDI algorithm and BCDI algorithm are more effective than the Kubernetes default scheduling algorithms.