Accelerate Data Loading in Distributed Training with Fast Reconfigurable Optical Network
We enhance our prior all-optical reconfigurable network to accelerate the data loading with minimized commu-nication distance between compute and storage nodes. A novel corresponding data loading scheme is proposed to further reduce the distance. Simulation results show up to 3.3x improvement compared to fat-tree.
Authors
- Wenzhe Li
- Guojun Yuan (ORCID: https://orcid.org/0000-0002-4864-7677)
- Peiheng Zhang
- X. Zhang (ORCID: https://orcid.org/0009-0006-2475-7536)
- Wenxuan Jia (ORCID: https://orcid.org/0000-0001-7596-2704)
- Zhuohan Chen (ORCID: https://orcid.org/0000-0001-8820-3459)
- Zhan Wang (ORCID: https://orcid.org/0000-0002-5791-1097)
Institutions
- Chinese Academy of Sciences (CN)
- Institute of Computing Technology (CN)
Publication Details
- Published
- 2024-11-02
- DOI
- https://doi.org/10.1109/acp/ipoc63121.2024.10809388
- Primary Topic
- Interconnection Networks and Systems
- Type
- article
- Field-Weighted Citation Impact
- 0.00