As machine learning (ML) and artificial intelligence (AI) technologies become increasingly prevalent in society, concerns about their environmental sustainability have grown. Developing and deploying ML-enabled systems, especially during training and inference, are resource-intensive, raising sustainability issues. Green AI has emerged as a response, advocating for reducing the computational demands of AI while maintaining accuracy. While recent research has identified various green tactics for developing sustainable ML-enabled systems, there is a gap in understanding the extent to which these tactics are adopted in real-world projects and whether additional, undocumented practices exist. This paper addresses these gaps by presenting a mining software repository study that evaluates the adoption of green tactics in 168 open-source ML projects on GitHub. In doing so, we introduce a novel mining mechanism based on large language models to identify and analyze green tactics within software repositories. Our results provide insights into the adoption of green tactics found in the literature and expand previous catalogs by providing 12 new tactics, with code examples to support wider implementation. This study contributes to the development of more sustainable ML systems by identifying adopted green tactics that offer substantial environmental benefits with minimal implementation effort. It provides practical insights for developers to green their systems and offers a path for future research to automate the integration of these tactics.
翻译:暂无翻译