As diverse applications move to high performance computing (HPC), I/O workloads have grown more varied. While parallel file systems have been refined for decades, object stores now offer new models and interfaces. This full-day tutorial explores object storage technologies and modern I/O approaches that leverage both object stores and traditional file systems. Through lectures and hands-on exercises, attendees will learn object storage design, usage, and integration within the evolving I/O stack—gaining the insight to choose the right storage solution and effectively use libraries and tools to simplify access.
Training large language models (LLMs) generates heavy I/O, and vendor guidelines often overestimate bandwidth needs by assuming perfect GPU utilization. Using telemetry from production VAST systems supporting major AI supercomputers, we analyzed over 85,000 checkpoints from 40 LLM training jobs. Even trillion-parameter models required only a few hundred GB/s for efficient checkpointing. From this, we derive a simple demand-side model linking LLM size and checkpoint interval to required bandwidth—helping avoid I/O overprovisioning and freeing more resources for compute.
Deep learning is reshaping the HPC landscape, introducing new I/O challenges for storage systems designed around traditional simulation workloads. Unlike classical models with large, sequential writes, AI workflows require random access for data shuffling and high burst bandwidth for checkpoints—often simultaneously. Coupling simulations with AI models further complicates data movement and storage. This panel explores how I/O systems must evolve to meet these demands and the open challenges ahead.
Session: PDSW'25: The 10th International Parallel Data Systems Workshop
Never before have applications reshaped storage architectures so rapidly. Agentic workloads, KV caching for LLMs, vector databases, relational graphs, and GNNs introduce new challenges—extracting insight from unstructured data, optimizing IOPs/TCO for fine-grained access, managing permissions, and integrating GPU-driven storage with traditional systems. This session brings together experts, vendors, and CSPs to explore emerging technologies and offer fresh perspectives on the future of storage.
This panel explores why traditional HPC systems can’t keep pace with AI factories—and how object interfaces, key-value caching, and global data mobility are reshaping performance. We’ll discuss how legacy file-based I/O and centralized metadata limit modern AI workflows, and how new architectures are emerging to support the adaptive nature of AI data.
