Analyzing the impact of the MPI allreduce in distributed training of convolutional neural networks

For many distributed applications, data communication poses an important bottleneck from the points of view of performance and energy consumption. As more cores are integrated per node, in general the global performance of the system increases yet eventually becomes limited by the interconnection network. This is the case for distributed data-parallel training of convolutional neural networks (CNNs), which usually proceeds on a cluster with a small to moderate number of nodes. In this paper, we analyze the performance of the Allreduce collective communication primitive, a key to the efficient data-parallel distributed training of CNNs. Our study targets the distinct realizations of this primitive in three high performance instances of Message Passing Interface (MPI), namely MPICH, OpenMPI, and IntelMPI, and employs a cluster equipped with state-of-the-art processor and network technologies. In addition, we apply the insights gained from the experimental analysis to the optimization of the TensorFlow framework when running on top of Horovod. Our study reveals that a careful selection of the most convenient MPI library and Allreduce (ARD) realization accelerates the training throughput by a factor of 1.2×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2\times $$\end{document} compared with the default algorithm in the same MPI library, and up to 2.8×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.8\times $$\end{document} when comparing distinct MPI libraries in a number of relevant combinations of CNN model+dataset.