OMLSched: Computation-Driven Flow Scheduling for Optical Machine Learning Training Clusters
IEEE Open Journal of the Communications Society · 2026 · 7 · 4100-4111
DOI: 10.1109/ojcoms.2026.3683080 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Open Journal of the Communications Society, 2026. Verfasst von Qiaojun Hu und 6 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Key Research and Development Program of China, Beijing University of Posts and Telecommunications (BUPT) Excellent Ph.D. Students Foundation, BUPT–China Unicom Joint Innovation Center.
Abstract (Originalsprache)
The AI model size and the amount of training data have increased dramatically, causing the computing power of one single GPU to be insufficient. Although distributed machine learning (DML) can enhance computing power by increasing the number of GPUs, the communication overhead between GPUs limits the training speed and utilization. Optical circuit switches (OCSes) are used to construct ML clusters to speedup DML, but there are also problems such as port constraints and reconfiguration delay. While GPU utilization remains a critical metric for ML service providers, existing schedulers lack explicit optimization for this objective. Therefore, refined scheduling strategies are required to overcome hardware limitations. In this paper, we formally model the GPU utilization optimization problem in ML clusters, establish its relationship with job scheduling. We propose both ILP-based optimal solutions and practical heuristic-based online scheduler called OMLSched to schedule the DML jobs while minimizing the inter-job contention for optical ML clusters. Finally, we evaluate the performance of OMLSched and several other existing scheduling methods in simulations. The simulation results show that OMLSched can achieve performance improvements of 48% in training speed and 28% in GPU utilization respectively.
Förderung
- National Key Research and Development Program of China
- Beijing University of Posts and Telecommunications (BUPT) Excellent Ph.D. Students Foundation
- BUPT–China Unicom Joint Innovation Center
Zitieren
APA
Hu, Q., Wang, W., Wang, X., Li, Y., Zhao, Y., Tan, Y., & Zhang, J. (2026). OMLSched: Computation-Driven Flow Scheduling for Optical Machine Learning Training Clusters. IEEE Open Journal of the Communications Society, 7, 4100-4111. https://doi.org/10.1109/ojcoms.2026.3683080
Vancouver
Hu Q, Wang W, Wang X, Li Y, Zhao Y, Tan Y, et al. OMLSched: Computation-Driven Flow Scheduling for Optical Machine Learning Training Clusters. IEEE Open Journal of the Communications Society. 2026;7:4100-4111. doi:10.1109/ojcoms.2026.3683080
BibTeX
@article{Hu2026,
author = {Hu, Qiaojun and Wang, Wei and Wang, Xiaoyu and Li, Yajie and Zhao, Yongli and Tan, Yanxia and Zhang, Jie},
title = {OMLSched: Computation-Driven Flow Scheduling for Optical Machine Learning Training Clusters},
journal = {IEEE Open Journal of the Communications Society},
year = {2026},
volume = {7},
pages = {4100--4111},
doi = {10.1109/ojcoms.2026.3683080}
}
EUVIMED