Scalability of Memory Subsystem in AI Storage Systems - IO Profiling and Performance Analysis
Learn about critical storage system performance factors for AI workloads in this 17-minute technical talk from the Open Compute Project. Explore real-world IO profiling data from AI customers while examining how overlapping pipeline stages and the IO Blender effect impact system performance. Gain insights into metadata overhead costs and understand how filesystem latency acts as a bottleneck for pipeline progress. Discover why Generative AI's micro-burst performance needs challenge traditional metrics, analyze training epoch delays caused by media silos, and explore WEKA's solutions to these storage challenges. Presented by Isaiah Weiner from Weka IO, dive deep into memory subsystem scalability considerations for both traditional and generative AI applications.