Scheduling Parallel Optical Circuit Switches for AI Training

The paper introduces Spectra, an algorithm for scheduling a single AI traffic demand matrix across multiple parallel optical circuit switches while accounting for reconfiguration delays. Spectra decomposes the demand into weighted permutations, assigns them load‑aware across switches, and then equalizes load via controlled permutation splitting. Experiments on GPT and Qwen MoE workloads show Spectra reduces schedule makespan by 1.4×, 2.7×, and 2.4× respectively compared to state‑of‑the‑art baselines, approaching newly derived lower bounds.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

• llms
• benchmarks