SNAP: A Communication Efficient Distributed Machine Learning Framework for Edge Computing

More and more applications learn from the data collected by the edge devices. Conventional learning methods, such as gathering all the raw data to train an ultimate model in a centralized way, or training a target model in a distributed manner under the parameter server framework, suffer a high communication cost. In this paper, we design Select Neighbors and Parameters (SNAP), a communication efficient distributed machine learning framework, to mitigate the communication cost. A distinct feature of SNAP is that the edge servers act as peers to each other. Specifically, in SNAP, every edge server hosts a copy of the global model, trains it with the local data, and periodically updates the local parameters based on the weighted sum of the parameters from its neighbors (i.e., peers) only (i.e., without pulling the parameters from all other edge servers). Different from most of the previous works on consensus optimization in which the weight matrix to update parameter values is predefined, we propose a scheme to optimize the weight matrix based on the network topology, and hence the convergence rate can be improved. Another key idea in SNAP is that only the parameters which have been changed significantly since the last iteration will be sent to the neighbors. Both theoretical analysis and simulations show that SNAP can achieve the same accuracy performance as the centralized training method. Compared to the state-of-the-art communication-aware distributed learning scheme TernGrad, SNAP incurs a significantly lower (99.6% lower) communication cost.