TensorLightning: A Traffic-Efficient Distributed Deep Learning on Commodity Spark Clusters

Abstract

With the recent success of deep learning, the amount of data and computation continues to grow daily. Hence a distributed deep learning system that shares the training workload has been researched extensively. Although a scale-out distributed environment using commodity servers is widely used, not only is there a limit due to synchronous operation and communication traffic but also combining deep neural network (DNN) training with existing clusters often demands additional hardware and migration between different cluster frameworks or libraries, which is highly inefficient. Therefore, we propose TensorLightning which integrates the widely used data pipeline of Apache Spark with powerful deep learning libraries, Caffe and TensorFlow. TensorLightning embraces a brand-new parameter aggregation algorithm and parallel asynchronous parameter managing schemes to relieve communication discrepancies and overhead. We redesign the elastic averaging stochastic gradient descent algorithm with pruned and sparse form parameters. Our approach provides the fast and flexible DNN training with high accessibility. We evaluated our proposed framework with convolutional neural network and recurrent neural network models; the framework reduces network traffic by 67% with faster convergence.