Benchmarking Storage with AI Workloads - Performance Analysis and Optimization

Explore a 46-minute technical presentation from the Storage Developer Conference 2023 that delves into the critical relationship between storage systems and AI/ML workloads in modern data centers. Learn how storage performance impacts AI/ML pipelines through detailed analyses of various storage-intensifying approaches, including limited system memory scenarios, simultaneous data ingestion and training, parallel training workloads, and streaming AI inference. Gain insights into system resource utilization patterns, I/O characteristics such as access locality, request sizes, read-write ratios, and file offset patterns. Discover how high-performance storage solutions like PM9A3 can handle challenging workloads, delivering 25x I/O with minimal overhead and 3x higher throughput compared to MLPerf inference implementation. Samsung Semiconductor experts Devasena Inupakutika, Charles Lofton, and Bridget Davis present practical approaches for benchmarking storage systems, understanding performance implications, and optimizing storage configurations to meet the unique demands of AI/ML workloads.