Investigating the Potential of Kepler Toward Power Observability for Sustainable Cloud Computing

Power monitoring is a cornerstone of sustainability efforts, especially as the energy demands of Artificial Intelligence (AI) workloads continue to rise. However, accurately measuring power consumption in cloud-native environments remains challenging due to technical constraints such as fine-grained monitoring requirements and the high abstraction introduced by virtualization and containerization. Kepler (Kubernetes-based Efficient Power Level Exporter), an open-source tool for container-level power monitoring, has emerged as a promising solution for cloud power observability. This paper outlines the key requirements for accurate power tracking in containerized environments and assesses Kepler’s alignment with these criteria. However, its precision remains unvalidated, mainly due to the lack of a systematic evaluation methodology. To fill this gap, we introduce a novel accuracy validation framework tailored to assess container-level power monitoring tools under dynamic controlled multi-tenancy environments, including CPU frequency scaling, C-state transitions, and varying co-runner workloads (i.e., co-hosted containers executing concurrently on other cores of the same processor socket). Using this framework, we perform the first in-depth evaluation of Kepler’s accuracy in real-world cloud scenarios (e.g., dynamic power configuration settings, dynamic workloads). Our results show that Kepler’s container-level power estimation exhibits a root mean squared error (RMSE) of 11.9 Watts against the RAPL ground truth, corresponding to an overestimation of approximately 15×. Its accuracy is highly sensitive to runtime factors such as CPU configuration and C-state transitions, which reveals critical limitations in Kepler’s current power model and highlights the need for refinement. This work establishes the foundation for more precise and effective power observability in cloud computing and paves the way for sustainable cloud computing.