Carbon-efficient AI Training Workloads Placement in Cloud Datacenters over Optical Networks

The latest breakthroughs in generative AI have sparked a surge in demand for AI training, posing significant environmental challenges due to its large energy consumption. In this paper, we propose a carbon-aware workload placement (CAWP) algorithm to optimize the distribution of AI training workloads in geographically distributed datacenters and to reduce the overall carbon emissions of AI training workloads, while meeting the communication and time requirements of the workloads. We conducted extensive experiments to compare our proposed algorithm with local placement and random placement algorithm. Compared with other algorithms, our proposed CAWP algorithm can optimize the AI workload distribution based on data center hardware resources, carbon emission coefficients, and data transmission bandwidth between data centers, resulting in a significant reduction in carbon emissions (~20%).