Shuffle-Exchange Brings Faster: Reduce the Idle Time During Communication for Decentralized Neural Network Training

As a crucial scheme to accelerate the deep neural network (DNN) training, distributed stochastic gradient descent (DSGD) is widely adopted in many real-world applications. In most distributed deep learning (DL) frameworks, DSGD is implemented with Ring-AllReduce architecture (Ring-SGD) and uses a computation-communication overlap strategy to address the overhead of the massive communications required by DSGD. However, we observe that although $O(1)$ gradients are needed to be communicated per worker in Ring-SGD, the $O(n)$ handshakes required by Ring-SGD limits its usage when training with many workers or in high latency network. In this paper, we propose Shuffle-Exchange SGD (SESGD) to solve the dilemma of Ring-SGD. In the cluster of 16 workers with 0.1ms Ethernet latency, SESGD can accelerate the DNN training to $1.7 \times$ without losing model accuracy. Moreover, the process can be accelerated up to $5\times$ in high latency networks (5ms).