Conference:  KubeCon + CloudNativeCon Europe 2023

Authors: Yuan Chen

2023-04-18

The Vanilla Kubernetes scheduler does not take power into account when placing pods on nodes and racks. As a result, the power usage of servers and racks in a data center can exceed the allocated power envelope and lead to service outages and costly downtime. In this short talk, Yuan Chen from Apple will give an overview of a new scheduler feature to support power-aware scheduling in Kubernetes. The proposed power-aware scheduling can protect power supply infrastructures and improve workload stability in large scale Kubernetes clusters by (1) enforcing power cap at both the server and rack levels, and (2) by optimizing pod placement to more evenly distribute workloads and power demand across servers and racks. The enhanced scheduling strategy is implemented by extending the default Kubernetes scheduler via scheduler plugins using the standard Kubernetes scheduling framework API. Specifically, power capping is implemented using the scheduler Filter plugin, and workload and power distribution optimization is achieved via the scheduler Scoring plugin. The improved scheduler can help safely increase server hardware and data center infrastructure size, and improve resource utilization and workload reliability for Kubernetes clusters.