EUVIMEDEUVIMED entdecken
Crossref

Task placement and traffic interleaving for cross-datacenter LLM training over optical networks

Qiaojun Hu, Wei Wang, Chongzhu Huang, Xiaoyu Wang, Yajie Li, Yongli Zhao, Yanlei Zheng, Yanxia Tan, Jie Zhang

Journal of Optical Communications and Networking · 01.02.2026 · 18 (2) · 137

DOI: 10.1364/jocn.579324 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Optical Communications and Networking, online erschienen am 28.01.2026. Verfasst von Qiaojun Hu und 8 weiteren Autor:innen. Die Arbeit wurde bisher 4-mal zitiert (Crossref) und nennt 26 Literaturangaben. Gefördert von: National Key Research and Development Program of China, BUPT Excellent Ph.D. Students Foundation, BUPT-China Unicom Joint Innovation Center.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

With the continuous growth of large language model (LLM) sizes, individual datacenters are increasingly insufficient to support LLM training. Cross-datacenter training has become a feasible solution, where optical networks play a crucial role in data transmission. However, current optical networks suffer from severe over-provisioning issues, primarily due to the conflict between the bursty nature of LLM traffic and the fixed bandwidth of optical network connections. Therefore, this paper presents CrossOptic, a novel, to our knowledge, framework for optimizing cross-datacenter LLM training over optical networks. The proposed solution addresses the critical challenges of bandwidth under-utilization in optical networks and excessive communication overhead in distributed LLM training. CrossOptic integrates two synergistic components: adaptive GPU resource orchestration (AGRO) for communication-minimizing task placement and traffic-interleaved connectivity provisioning (TICP) for bandwidth-efficient flow aggregation. Through comprehensive simulations using realistic GPT model workloads, CrossOptic outperforms conventional dedicated connection approaches by 38.1% in terms of optical resource utilization. For a GPT model with 529.6B parameters, compared with the training methods using data parallelism and pipeline parallelism between datacenters, CrossOptic reduces the iteration time by 55% and 14%, respectively. The framework achieves significant cost efficiency in the optical infrastructure while supporting the computational demands of LLM training.

Förderung

  • National Key Research and Development Program of China
  • BUPT Excellent Ph.D. Students Foundation
  • BUPT-China Unicom Joint Innovation Center
  • State Key Laboratory of Information Photonics and Optical Communications

Zitieren

APA

Hu, Q., Wang, W., Huang, C., Wang, X., Li, Y., Zhao, Y., Zheng, Y., Tan, Y., & Zhang, J. (2026). Task placement and traffic interleaving for cross-datacenter LLM training over optical networks. Journal of Optical Communications and Networking, 18(2), 137. https://doi.org/10.1364/jocn.579324

Vancouver

Hu Q, Wang W, Huang C, Wang X, Li Y, Zhao Y, et al. Task placement and traffic interleaving for cross-datacenter LLM training over optical networks. Journal of Optical Communications and Networking. 2026;18(2):137. doi:10.1364/jocn.579324

BibTeX
@article{Hu2026,
  author = {Hu, Qiaojun and Wang, Wei and Huang, Chongzhu and Wang, Xiaoyu and Li, Yajie and Zhao, Yongli and Zheng, Yanlei and Tan, Yanxia and Zhang, Jie},
  title = {Task placement and traffic interleaving for cross-datacenter LLM training over optical networks},
  journal = {Journal of Optical Communications and Networking},
  year = {2026},
  volume = {18},
  number = {2},
  pages = {137},
  doi = {10.1364/jocn.579324}
}
Kostenlos starten →