Modeling the Impact of Fiber Latency on Compute-Communication Overlap in Geo-Distributed Multi-Datacenter AI Training
This study investigates the impact of optical fiber latency on the efficiency of computation-communication overlap in geographically distributed, multi-datacenter AI training. By developing a discrete-event simulation model that integrates data-parallel training architectures with the physical characteristics of fiber-optic transmission, the work systematically evaluates system performance across varying inter-site distances. The research identifies, for the first time, an optimal deployment range of 10–100 kilometers for AI clusters to maximize overlap efficiency. Furthermore, it demonstrates that employing hollow-core fiber can enhance computation-communication overlap efficiency by 25%, offering critical theoretical insights for designing low-latency, wide-area AI infrastructure.