An Efficient Method for Training Deep Learning Networks Distributed

Abstract

Training deep learning (DL) is a computationally intensive process; as a result, training time can become so long that it impedes the development of DL. High performance computing clusters, especially supercomputers, are equipped with a large amount of computing resources, storage resources, and efficient interconnection ability, which can train DL networks better and faster. In this paper, we propose a method to train DL networks distributed with high efficiency. First, we propose a hierarchical synchronous Stochastic Gradient Descent (SGD) strategy, which can make full use of hardware resources and greatly increase computational efficiency. Second, we present a two-level parameter synchronization scheme which can reduce communication overhead by transmitting parameters of the first layer models in shared memory. Third, we optimize the parallel I/O by making each reader read data as continuously as possible to avoid the high overhead of discontinuous data reading. At last, we integrate the LARS algorithm into our system. The experimental results demonstrate that our approach has tremendous performance advantages relative to unoptimized methods. Compared with the native distributed strategy, our hierarchical synchronous SGD strategy (HSGD) can increase computing efficiency by about 20 times.