Efficient All-Reduce for Distributed DNN Training in Optical Interconnect Systems
Fei Dai, Yawen Chen, Zhiyi Huang, Haibo Zhang, Fangfang Zhang
Abstract
All-reduce is the crucial communication primitive to reduce model parameters in distributed Deep Neural Networks (DNN) training. Most existing all-reduce algorithms are designed for traditional electrical interconnect systems, which cannot meet the communication requirements for distributed training of large DNNs due to the low data bandwidth of the electrical interconnect systems. One of the promising alternatives for electrical interconnect is optical interconnect, which can provide high bandwidth, low transmission delay, and low power cost. We propose an efficient scheme called WRHT (Wavelength Reused Hierarchical Tree) for implementing all-reduce operation in optical interconnect systems. WRHT can take advantage of WDM (Wavelength Division Multiplexing) to reduce the communication time of distributed data-parallel DNN training. Simulations using real DNN models show that, compared to all-reduce algorithms in the electrical and optical network systems, our approach reduces communication time by 75.76% and 91.86%, respectively.