Workshop: 8th Workshop on Programming and Performance Visualization Tools (ProTools)
Authors: Hariharan Devarajan (Lawrence Livermore National Laboratory); Sarah Neuwirth (Johannes Gutenberg University Mainz); Chen Wang (Nanyang Technological University); Ray A. O. Sinurat (University of Chicago); Neeraj Rajesh (Illinois Institute of Technology); Jay Lofstead (Sandia National Laboratories); and Vanessa Sochat, Daniel Milroy, and Thomas Scogland (Lawrence Livermore National Laboratory)
Abstract: Modern scientific workloads move large data volumes through deep, complex I/O stacks in high-performance computing (HPC) systems. State-of-the-art tools such as Darshan capture API-level statistics describing what happened, not why, so performance variation is often dismissed as jitter without direct evidence. Emerging approaches such as eBPF enable fine-grained, low-overhead collection, but naively applied they produce overwhelming data volumes. We present DataCrumbs, an I/O monitoring tool that uses function probing to collect selective, fine-grained performance data across the full I/O stack, from user-space runtimes to kernel-space file-system paths, enabling explainable I/O analysis. DataCrumbs instruments 8,723 functions at only 0.1% profiling and under 3% tracing overhead, while a time-windowed, parallel event-association algorithm scales cross-layer root-cause analysis. Across four workloads, DataCrumbs exposes cross-layer insights invisible to user-space-only tools, including a tuned read-ahead insight that cuts Montage's POSIX I/O time by 1.35x. DataCrumbs thus enables selective, fine-grained, explainable I/O analysis for real HPC workloads.
Back to 8th Workshop on Programming and Performance Visualization Tools (ProTools) Archive Listing
Back to Full Workshop Archive Listing