High-Performance Input Pipelines for Scalable Deep Learning

Abstract: Learn how to keep your GPUs fed with data as you train the next-generation of deep learning architectures. As GPU technology continues to advance, the demand for faster data continues to grow. In deep learning, input pipelines are responsible for a complex chain of actions that ultimately feed data into GPU memory: defining how files are read from storage, deserializing them into data structures, pre-processing on a CPU, and copying to the GPU. These pipelines bring together complex hardware systems—including cluster networks, peripheral interconnects, modern CPUs, and storage devices—along with sophisticated software systems to drive the data movement and transformation.

In this talk, we present a new benchmark suite for evaluating and tuning input pipelines. We will examine results with TensorFlow’s DataSets API on a DGX-1 with V100 and provide guidance on key tuning parameters and diagnostic techniques for improving performance.

Bio: Ramnath is Senior Manager, Product Marketing for AI and Deep Learning at Pure. Previously, he worked as a Marketing Manager at Mellanox Technologies, leading the market development activities for AI, BigData & Cloud. Before that, he was the RDMA Solutions Evangelist and led Cloud & Big Data strategy at Emulex. Prior to joining Emulex, he worked in two of the most prestigious research labs in Europe - Brain Mind Institute at EPFL, Switzerland and Barcelona Supercomputing Center in Spain. He has 15+ publications in leading conference and journals