Temperature-Aware Scheduling of LLM Inference in Large-Scale Geo-Distributed Edge Data Centers with Distributed Optimization

摘要

The environmental impact of Large Language Models (LLMs) on data centers hosting these models is becoming a significant concern. While many efforts have focused on reducing the substantial training overhead of LLMs, carbon and water consumption during the inference phase can often surpass the costs associated with their training. The cooling systems of data centers are crucial in this context, but they are frequently modeled with a location-independent efficiency term. However, their energy efficiency is highly influenced by ambient temperature, which can vary significantly across different geographical locations. Leveraging this temperature diversity can help reduce total cooling energy costs and improve the performance of edge data centers. To address these critical sustainability issues related to LLMs, this study proposes a temperature-aware approach that co-optimizes LLM energy costs, carbon emissions, time-to-first token, and water consumption. The approach employs a distributed optimization algorithm based on an alternating direction method of multipliers, aimed at enhancing the sustainability of LLM hosting across geo-distributed edge data centers in Australia. Our method demonstrates reductions in cooling energy consumption and improves overall cost efficiency for geo-distributed cloud environments.