Two major trends are reshaping HPC storage needs: (1) the shift from traditional bulk-synchronous models to complex workflows integrating in-situ analysis, AI, and data analytics, which challenge existing storage systems, and (2) the emergence of new technologies like GPUDirect, CXL, and computational storage which, despite their potential, are still in early stages of adoption. As storage architectures move beyond two-tier models to advanced hierarchies with fast, temporary layers near compute resources, they often fail to fully accommodate emerging workload demands and leverage new storage innovations. This Dagstuhl Seminar aims to bridge these gaps to refine the research agenda for HPC storage and I/O, focusing on understanding workload requirements, leveraging new technologies, addressing limitations of current optimization tools, and shaping future storage system design.

The goal of this seminar is to bring together experts in I/O analysis, optimization, file systems, and storage systems from the HPC, data center, and cloud communities. Together, we will evaluate how the community captures and analyzes I/O workloads in HPC systems, supports the evolving I/O needs of modern workloads, and integrates recent storage advancements. We will also identify gaps in current methodologies and explore ways to develop a deeper understanding of future HPC workload and system requirements. We expect our discussions to result in (1) a set of common terminologies across the community to describe the emerging I/O workloads and their I/O requirements (2) concrete recommendations on how the latest and predicted storage advances can be utilized to support the modern applications; and (3) a special issue of a journal documenting our findings and providing the needed outreach to the wider community.

In this seminar, we will explore key topics and questions aimed towards our goal of improving understanding of parallel I/O behavior, emerging workload requirements, and opportunities in new storage technologies. We expect the following topics and questions to generate lively discussions:

• Emerging Workloads and their Requirements: How are HPC workloads evolving and how do their I/O requirements differ from the past? What impact do machine learning and in-situ analysis have on I/O behavior? How will these trends shape future workloads?
• Emerging Storage Technologies and their Potential: Which advanced storage system designs can enhance future workflow performance? How can we leverage emerging technologies like DPUs, CXL, and computational storage to optimize I/O?
• Explainable I/O for HPC Systems: Current I/O tools identify bottlenecks but lack explainability, making it difficult to pinpoint root causes of performance issues. As workloads and storage architectures grow more complex, existing tools fail to capture critical data flows. What are the limitations of current monitoring methods, and what new techniques can provide a more comprehensive view of I/O behavior?
• Performance Analysis and Optimization: How effective are current performance analysis and optimization techniques for emerging workloads and storage? What gaps exist in optimization strategies? Given the increasing focus on energy efficiency, how to achieve multiple objectives of performance, energy efficiency, and programmer productivity? How can AI-driven methods automate optimization without burdening application developers?

The results of this seminar will have broad applicability for those interested in improving I/O performance of HPC applications, which is a typically overlooked bottleneck that results in inefficient system utilization. We also anticipate our meeting to spark long-term, international collaboration across HPC I/O researchers that share the goal of better preparing the HPC system for future applications.

Creative Commons BY 4.0

Suren Byna, Shadi Ibrahim, Sarah Neuwirth, and Chen Wang