Sebastian Werner

The rapid digitization and the increasing use of emerging technologies such as AI models have significantly contributed to the emissions of computing infrastructure. Efforts to mitigate this impact typically focus on the infrastructure level such as powering data centers with renewable energy, or through the specific design of energy-efficient software. However, both strategies rely on stakeholder intervention, making their adoption in legacy and already-deployed systems unlikely. As a result, past architectural and implementation decisions continue to incur additional energy usage - a phenomenon we refer to as energy debt. Hence, in this paper, we investigate the potential of serverless computing platforms to automatically reduce energy debt by leveraging the unique access to function source code. Specifically, we explore whether large language models (LLMs) can translate serverless functions into more energy-efficient programming languages while preserving functional correctness. To this end, we design and implement ReFaaS and integrate it into the Fission serverless framework. We evaluate multiple LLMs on their ability to perform such code translations and analyze their impact on energy consumption. Our preliminary results indicate that translated functions can reduce invocation energy by up to 70%, achieving net energy savings after approximately 3,000 to 5,000 invocations, depending on the LLM used. Nonetheless, the approach faces several challenges: not all functions are suitable for translation, and for some, the amortization threshold is significantly higher or unreachable. Despite these limitations, we identify four key research challenges whose resolution could unlock long-term, automated mitigation of energy debt in serverless computing.

Current approaches to designing energy-efficient applications typically rely on measuring individual components using readily available local metrics, like CPU utilization. However, these metrics fall short when applied to cloud-native applications, which operate within the multi-tenant, shared environments of distributed cloud providers. Assessing and optimizing the energy efficiency of cloud-native applications requires consideration of the complex, layered nature of modern cloud stacks. To address this need, we present a comprehensive, automated, and extensible experimentation framework that enables developers to measure energy efficiency across all relevant layers of a cloud-based application and evaluate associated quality trade-offs. Our framework integrates a suite of service quality and sustainability metrics, providing compatibility with any Kubernetes-based application. We demonstrate the feasibility and effectiveness of this approach through initial experimental results, comparing architectural design alternatives for a widely used open-source cloud-native application.