Accelerate Distributed Deep Learning with a Fast Reconfigurable Optical Network
- Resource Type
- Conference
- Authors
- Li, Wenzhe; Yuan, Guojun; Wang, Zhan; Tan, Guangming; Zhang, Peiheng; Rouskas, George N.
- Source
- 2024 Optical Fiber Communications Conference and Exhibition (OFC) Optical Fiber Communications Conference and Exhibition (OFC), 2024. :1-3 Mar, 2024
- Subject
- Communication, Networking and Broadcast Technologies
Photonics and Electrooptics
Training
Deep learning
All-optical networks
Electric potential
Optical switches
Decentralized control
Prototypes
- Language
We propose a fast-reconfigurable and scalable optical network architecture, which employs a flow-based transmit scheduling scheme to accelerate data parallelism in distributed deep learning. Experimental results demonstrate that the 4-node prototype achieves training times comparable to those of ideal electrical switching.