An Autonomous GreenOps Architecture for Real-Time Software Energy Measurement and Carbon-Aware Routing

Abstract

The escalating proliferation of software (AI solutions in particular) has triggered a real increase in global energy consumption, and urges nowadays a deep reflection on how to direct digital innovations to meet climate commitments. Within this context, this paper addresses the urgent need for green software by implementing an autonomous carbon-aware framework that treats software as a physical resource. Our proposed architecture is organized through a 3-layer framework: Measurement (Kepler with eBPF), Intelligence (Carbon Aware SDK), and Actuation (time or location shifting). As a proof of concept, we conducted a fine-grained analysis of energy consumption and carbon emissions, by executing multiple times an inference prompt over some widely adopted LLMs across diverse energy grid profiles. The obtained results show that such software-level interventions can reduce carbon emissions up to 40% for our Moroccan grid case, by timely routing tasks in the same country, from coal-gas night grid to solar mid-day grid without degrading service performance; In addition, switching from an energy-intensive LLM to a low energy one for the same task clearly demonstrates a significant contribution to decreasing carbon.