Optical Networks for Distributed AI Training: Flow Completion Time and Training Time Analysis

We present a comprehensive framework for evaluating distributed AI training performance across single and multi-datacenter environments. Our approach combines 2 existing simulators, NeuronaBox's high-fidelity traffic generation and m3's efficient flow-level simulation to analyze various parallelization strategies and synchronization methods much faster than with existing packet-level simulators. We propose and evaluate two datacenter architectures: a single datacenter network (S-DCN) supporting 128k servers across 32 AI pods with non-blocking connectivity, and a distributed datacenter network (D-DCN). Through extensive evaluation, we demonstrate that 800G DCI links can reduce training time by 52.8% compared to 200G links at 50km distance and that introducing an optical circuit switch within a DC achieves superior failure recovery, maintaining flow completion time within 21% of normal operation during pod failures compared to 26% degradation in traditional electronic packet-based architectures, translating to a 2.5 day advantage in recovery scenarios.