Problem
Research questions and friction points this paper is trying to address.
fiber latency
geo-distributed AI training
compute-communication overlap
data parallelism
multi-datacenter
Modeling the Impact of Fiber Latency on Compute-Communication Overlap in Geo-Distributed Multi-Datacenter AI Training
🤖 AI Summary
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.
📝 Abstract
We use discrete-event simulation to quantify the impact of fiber latency on the efficacy of geo-distributed AI model training with data parallelism. We conclude that the optimum distances between two AI clusters is 10-100km, over which hollow-core fiber enables 25% higher compute-communication overlap.
Innovation
Methods, ideas, or system contributions that make the work stand out.
fiber latency
compute-communication overlap
geo-distributed AI training
hollow-core fiber
discrete-event simulation
🔎 Similar Papers
No similar papers found.
