Predicting the Power Impact of Scheduling Decisions in Kubernetes Using Fine-Grained Monitoring and XGBoost

Abstract

The rapid growth of cloud data centres has increased their energy consumption and environmental footprint, highlighting the need for more energy-efficient resource management. Kubernetes has become a widely adopted container orchestration platform for automating the deployment, scaling, and management of containerized workloads. This study investigates the impact of Kubernetes scheduling decisions on cluster power consumption. A fine-grained monitoring system was implemented to characterize application behaviour and cluster state by collecting metrics at the node, shared-resource, and container levels. Controlled pod-placement scenarios were designed to evaluate how workload distribution topology, microservice affinity, and resource contention affect power consumption. Using the collected dataset, an XGBoost model was developed to predict cluster power consumption associated with pod placement based on pre-scheduling system-state metrics. The model achieved an (R2) score of 93.2%, demonstrating high predictive accuracy. Building on these results, future work will focus on developing a customized Kubernetes scheduler based on reinforcement learning and integrating the power-prediction model to enable energy-aware pod placement. The proposed approach aims to support more sustainable and energy-efficient cloud data centre operations.