A Methodology for System-Scale I/O Pattern Taxonomy for HPC Workloads
Théo Jolivel, François Tessier, Jakob Luettgau, Gabriel Antoniu, Philippe Deniel · INRIA a CCSD electronic archive server · 2025
As High-Performance Computing (HPC) races toward Exascale, I/O bottlenecks threaten to throttle the performance of data-intensive applications. Unlocking optimized I/O strategies such as I/O-aware job scheduling or data prefetching on intermediate storage tiers, to name a few, hinges on deeply understanding data access patterns by applications, a critical yet under-explored frontier. This paper unveils a novel methodology to classify I/O patterns through three lenses: temporality, periodicity, and metadata load. By harnessing advanced techniques to analyze execution traces, we instantiate our new methodology through MOSAIC, a robust Python library that transforms raw I/O data into valuable insights. Validated on two datasets from top-tier supercomputers, our approach reveals relevant correlations, such as near-permanent accesses with larger data writes, and exposes how application diversity shapes distinct I/O landscapes.