Framework optimizes LLM training over multi-datacenter optical networks

Researchers propose a new communication-bound-aware cross-domain resource assignment framework to optimize pipeline-parallel distributed LLM training over multi-datacenter optical networks.

• The framework targets pipeline-parallel distributed training of large language models across multiple data centers connected by optical networks, reducing iteration time by 31.25% and blocking requests by 13.20% compared to baseline approaches.
• The work is categorized under Networking and Internet Architecture, Artificial Intelligence, and Optics, and was submitted to arXiv on 23 Dec 2025 with an arXiv-issued DOI 10.48550/arXiv.2512.20080, licensed under CC BY 4.0.