MLCommons MLPerf Storage v2.0 Benchmark Analysis

This technical report analyzes the results of the MLCommons MLPerf Storage v2.0 benchmark, focusing on the performance of Western Digital's OpenFlex Data24 4000 Series. The document outlines the evaluation of storage subsystems using synthetic workloads that reflect real training behavior, specifically through models like ResNet-50 and 3D-UNet. It details the configurations of the OpenFlex Data24, including its performance metrics and efficiency in comparison to other systems. Key findings include the balanced design of the WDC EBOF configuration, which delivers high throughput and efficiency across multiple simulated GPUs. The report also discusses the challenges of data normalization in benchmarking, highlighting the influence of client architecture on I/O behavior and the lack of meaningful power data in current evaluations. It emphasizes the need for a more comprehensive approach to normalization that considers economic and operational factors, aiming to provide clearer insights into the true performance and sustainability of storage solutions.