Storage I/O Patterns in Deep Learning Workloads: Characterization, Optimization Strategies, and Performance Implications

Abstract

Deep learning workloads exhibit heterogeneous storage I/O patterns that differ considerably fromconventional enterprise workloads. In this article, we present a thorough characterization of storage I/Oacross the entire ML lifecycle — encompassing training data loading, model checkpointing, and inference serving — and identify optimization techniques