Design and evaluation of an Energy-Aware scheduling plugin for Kubernetes clusters

Abstract

The rapid proliferation of cloud-native orchestration via Kubernetes has led to significant escalations in data center energy consumption and carbon footprints. While traditional scheduling frameworks prioritize resource utilization, they often neglect the environmental impact of workload distribution. This dissertation proposes CARMEN, a novel energy-aware scheduling plugin for Kubernetes (K3s) that integrates carbon-aware computing principles into the orchestration layer. CARMEN implements a simple, deterministic ARM-priority rule that systematically places pods on ARM nodes (8W) over x86 nodes (35W), using real-time metrics from the Kubernetes Metrics Server. The CARMEN framework consists of five components: an external scheduler for ARM-first placement, a proactive rescheduler that migrates pods when energy gain exceeds 15%, a failure descheduler that automatically detects node failures and recovers affected pods within 5 seconds, a real-time Streamlit dashboard, and the Kubernetes Metrics Server. Evaluated through 30 consecutive tests on a simulated 3-node heterogeneous cluster (one ARM worker, one x86 worker, one x86 master), CARMEN achieves a mean energy gain of 38.86% compared to random placement, reducing power consumption from 35.96W to 22.00W, with gains ranging from 32.93% to 49.54% and perfect reproducibility (zero variance). Carbon emissions are reduced by the same percentage using the IPCC global average factor of 475 gCO2/kWh. Our findings demonstrate that CARMEN effectively balances operational performance with ecological sustainability, providing a scalable framework for green cloud-native computing