Important Notice: Starting with version 0.10.0, Kepler has undergone a complete ground-up rewrite. This represents a significant architectural improvement while maintaining the core mission of accurate energy consumption monitoring for cloud-native workloads.
📢 Read the full announcement: CNCF Slack Announcement
Authors: Sunyanan Choochotkaew, Marcelo Amaral, Huamin Chen
Originally published on the CNCF TAG ENV blog on 2024-06-06.
⚠️ Note: Kepler 0.10 attributes only active power to workloads and reports idle power per node. Bringing back idle power separation is tracked in Idle power separation for all power interfaces (kepler#2470).
Idle power is the type of power where active compute is not occurring, and includes static power and that consumed by control plane processes. Idle power attribution is challenging in public cloud environment where computation is shared among multiple tenants. It's contribution to the total power for a node is significant and can range from 20% to 60% of the power at maximum utilization. This blog post aims to highlight the importance of transparent, fine-grained energy metrics that distinguish between idle and dynamic power in software energy attribution within public cloud computing. Kepler addresses this need by providing a solution that attributes idle power separately from dynamic power in multi-tenant environments.