Storage Benchmarking and I/O Characterization for Next Generation AI Workloads

Learn how to benchmark and characterize storage systems for next-generation AI workloads in this 10-minute conference talk from the Open Compute Project. Explore the evolution of AI from compute-focused training to storage-intensive applications including vectorDBs, large language model inference, and generative AI systems that challenge traditional storage architectures with demands for both large sequential datasets and low-latency small read patterns. Discover comprehensive workload analysis methodologies using eBPF-based profiling, MLPerf storage benchmarking extensions, and real-workload analysis to understand VectorDB operations with small random reads and LLM KV cache system requirements. Examine vector similarity search benchmarking across major platforms and KV cache management for large parameter models through detailed workload characterization. Gain actionable insights for optimizing storage systems as an AI infrastructure architect, storage system engineer, or practitioner working with next-generation AI storage solutions.