Dynamic data partitioning strategy for distributed learning on heterogeneous edge system

Distributed machine learning on edge systems has attracted attention due to the development of artificial intelligence and edge computing. One challenge is straggler problem for synchronous updates during training, in which some edge nodes that complete training first have to wait for the nodes that complete training later. This results in long waiting time and downgrades the performance of distributed learning. In this paper, we investigate dynamic data partition for load balance among heterogeneous edge nodes. We propose experience-driven algorithms based on actor–critic deep reinforcement learning to optimize model training in distributed edge systems. It can learn the network environment and the computing capabilities of edge nodes, and thus strategically allocate training data to edge nodes. We conduct experiments on two commonly used datasets, i.e., MNIST and CIFAR-10, to evaluate the performance of the proposed method. The results show that the proposed DDPS can significantly reduce training latency, compared to random partition strategy, even partition strategy, greedy partition strategy and A2C strategy.