Despite its growing importance, scalable deep learning (DL) remains a difficult challenge. Scalability of large-scale DL is constrained by many factors, including those deriving from data movement and data processing. DL frameworks rely on large volumes of data to be fed to the computation engines for processing. However, current hardware trends showcase that data movement is already one of the slowest components in modern high performance computing systems, and this gap is only going to increase in the future. This includes data movement needed from the filesystem, within the network subsystem, and even within the node itself, all of which limit the scalability of DL frameworks on large systems. Even after data is moved to the computational units, managing this data is not easy. Modern DL frameworks use multiple components---such as graph scheduling, neural network training, gradient synchronization, and input pipeline processing---to process this data in an asynchronous uncoordinated manner, which results in straggler processes and consequently computational imbalance, further limiting scalability. This thesis studies a subset of the large body of data movement and data processing challenges that exist in modern DL frameworks.

For the first study, we investigate file I/O constraints that limit the scalability of large-scale DL. We first analyze the Caffe DL framework with Lightning Memory-Mapped Database (LMDB), one of the most widely used file I/O subsystems in DL frameworks, to understand the causes of file I/O inefficiencies. Based on our analysis, we propose LMDBIO---an optimized I/O plugin for scalable DL that addresses the various shortcomings in existing file I/O for DL. Our experimental results show that LMDBIO significantly outperforms LMDB in all cases and improves overall application performance by up to 65-fold on 9,216 CPUs of the Blues and Bebop supercomputers at Argonne National Laboratory.

Our second study deals with the computational imbalance problem in data processing. For most DL systems, the simultaneous and asynchronous execution of multiple data-processing components on shared hardware resources causes these components to contend with one another, leading to severe computational imbalance and degraded scalability. We propose various novel optimizations that minimize resource contention and improve performance by up to 35% for training various neural networks on 24,576 GPUs of the Summit supercomputer at Oak Ridge National Laboratory---the world's largest supercomputer at the time of writing of this thesis.