Agenda

Modern scientific workflows, AI pipelines, and data-intensive applications depend on increasingly complex storage ecosystems that span applications, runtimes, middleware, file systems, operating systems, networks, and heterogeneous storage devices. While the HPC community has invested heavily in monitoring infrastructure, today’s tools largely answer what happened—bandwidth achieved, I/O operations performed, or time spent in a layer—but rarely explain why performance behaved the way it did. As a result, performance variability is often attributed to system noise or complexity, leaving users and administrators to rely on intuition, manual experimentation, and expert knowledge to diagnose bottlenecks.

This keynote argues that future storage systems must be designed not only to store and move data efficiently, but also to continuously explain their behavior. Through a study of modern HPC storage architectures and monitoring ecosystems, we examine the growing gap between observability and explainability. Using a Master Architectural Plan (MAP) of contemporary storage stacks, we highlight the missing links that prevent current monitoring approaches from connecting application behavior, middleware transformations, operating-system activity, and hardware interactions into coherent explanations.

Building explainable storage systems requires a fundamental shift in both monitoring and system design. Monitoring infrastructures must provide cross-layer visibility, causal event association, workload-aware context, low-overhead continuous collection, scalable analysis, and optimization-aware telemetry. At the same time, storage systems themselves must be architected with explainability as a first-class design goal: monitoring must be native rather than bolted on, systems must capture both application and system context, telemetry must span all layers of the software and hardware stack, collected data must support causal reasoning, and monitoring must ultimately target optimization rather than merely reporting metrics.

Finally, the talk presents a vision for self-explaining storage systems—systems capable of automatically detecting anomalies, explaining bottlenecks, recommending optimizations, adapting dynamically to workload behavior, exposing provenance across the storage stack, and integrating AI-driven reasoning into performance analysis. As storage architectures become increasingly software-defined, heterogeneous, and autonomous, explainability will become as critical a design objective as performance, scalability, and reliability. The next generation of storage systems should not only move data—they should explain themselves.