Communication-Efficient Distributed Learning With MST-Based Scheduling Over Cross-Datacenter Optical Networks
IEEE Journal on Selected Areas in Communications · 2026 · 44 · 6033-6049
DOI: 10.1109/jsac.2026.3742502 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Journal on Selected Areas in Communications, 2026. Verfasst von Ruikun Wang und 9 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Natural Science Foundation of Shandong Province, Qingdao Natural Science Foundation.
Abstract (Originalsprache)
Many emerging artificial intelligence (AI) applications continue to grow in model size and computational complexity. To meet these increasing requirements, distributed learning (DL) has become a key paradigm, enabling model training to be partitioned and executed across multiple geographically distributed datacenters (DC). However, DL inherently requires frequent synchronization of model parameters. As the number of participating local models and the size of AI models grow, DL generates substantial communication overhead, resulting in heavy demands on cross-DC optical networks. To solve this problem, as our main and original contribution, we propose a minimum spanning tree (MST)-based strategy that transforms the DL scheduling problem into a procedure based on the construction of multiple spanning trees. This approach enables flexible routing and model aggregation, effectively reducing communication costs in terms of both resource consumption and latency reduction. We further formulate an integer linear programming (ILP) model to obtain the optimal solutions for benchmarking. We evaluate the proposed methods on 9-node, 30-node and 68-node substrate networks, where the number of local models and DL requests are varied to provide a comprehensive assessment. Numerical results show that the MST-based strategy achieves near-optimal performance, with a gap of 1.1% to 3.99% compared to the optimal ILP solution. Meanwhile, the proposed MST strategy shows strong scalability and outperforms existing heuristic benchmarks in terms of resource efficiency and latency reduction.
Förderung
- National Natural Science Foundation of China
- Natural Science Foundation of Shandong Province
- Qingdao Natural Science Foundation
- Science & Technology Cooperation Program of Shandong
- Open Fund of IPOC
- Shandong Key Laboratory of Deep Sea Equipment Intelligent Networking
Zitieren
APA
Wang, R., Zhang, Q., Zhang, J., Wang, X., Gu, Z., Yang, X., Musumeci, F., Ji, Y., Wang, J., & Tornatore, M. (2026). Communication-Efficient Distributed Learning With MST-Based Scheduling Over Cross-Datacenter Optical Networks. IEEE Journal on Selected Areas in Communications, 44, 6033-6049. https://doi.org/10.1109/jsac.2026.3742502
Vancouver
Wang R, Zhang Q, Zhang J, Wang X, Gu Z, Yang X, et al. Communication-Efficient Distributed Learning With MST-Based Scheduling Over Cross-Datacenter Optical Networks. IEEE Journal on Selected Areas in Communications. 2026;44:6033-6049. doi:10.1109/jsac.2026.3742502
BibTeX
@article{Wang2026,
author = {Wang, Ruikun and Zhang, Qiaolun and Zhang, Jiawei and Wang, Xin and Gu, Zhiqun and Yang, Xinghai and Musumeci, Francesco and Ji, Yuefeng and Wang, Jingjing and Tornatore, Massimo},
title = {Communication-Efficient Distributed Learning With MST-Based Scheduling Over Cross-Datacenter Optical Networks},
journal = {IEEE Journal on Selected Areas in Communications},
year = {2026},
volume = {44},
pages = {6033--6049},
doi = {10.1109/jsac.2026.3742502}
}
EUVIMED