Towards Evaluation of Tensorflow Performance in a Distributed Compute Environment

摘要

Tensorflow (TF) is a highly popular Deep Learning (DL) software framework. Neural network training, a critical part of DL workflow, is a computationally intensive process that can take days or even weeks. Therefore, achieving faster training times is an active area of research and practise. TF supports multiple GPU parallelization, both within a single machine and between multiple physical servers. However, the distributed case is hard to use and consequently, almost all published performance data comes from the single machine use case. To fill this gap, here we benchmark Tensorflow in a GPU-equipped distributed environment. Our work evaluates performance of various hardware and software combinations. In particular, we examine several types of interconnect technologies to determine their impact on performance. Our results show that with the right choice of input parameters and appropriate hardware, GPU-equipped general-purpose compute clusters can provide comparable deep learning training performance to specialized machines designed for AI workloads.