Sustainable, Situation-Aware Multi Objective Spatial Load Scheduling for Data Centers

Data centers are already among the largest consumers of electricity worldwide and recent advancements in Large Language Models (LLM) will further increase this demand.Through their operation, they are responsible for large amounts of indirect carbon emissions as well as direct and indirect water consumption.This work proposes a novel multi-objective load shifting optimization focusing on low costs and reduced resource depletion.For the resource depletion we combine the long term effects of carbon emissions with the short term effects of water consumption, which in a novel approach is weighted based on a newly developed spatial and temporal drought risk indicator.We investigate a use case where we optimally shift the inference load of a LLM spatially when hosted in ten different European countries.Additionally, we extend previous research on the water intensity of the electric grid by providing the first automated framework called IntensityLib for calculating water intensity factors for the European electricity grid as well as integrating a flow-tracing algorithm which models the effect of imports and exports.We find that adjusting for water scarcity leads to entirely different solutions compared to simply minimizing water consumption and can prevent further stress on regions with reoccurring drought risks.