Enhancing AI Backend Network with Dynamic Load Balancing Using Open Source Network Operation System SONiC

This paper evaluates dynamic load balancing (DLB) with the open-source network operating system SONiC to enhance AI backend networks. Our experiments demonstrate that DLB significantly improves network utilization, reducing link imbalance by as much as 26% in undersubscribed topologies and increasing bandwidth by up to 25% in all2all operations. By optimizing load distribution, DLB provides a critical advantage for AI-driven environments. The findings suggest that further exploration of DLB in diverse network configurations could maximize its benefits, contributing to a more efficient and reliable AI infrastructure.