LEO SC26 Artifact Evaluation — Profiling Data Tarball

Pre-collected HPCToolkit PC-sampling measurement databases for the 21 workloads evaluated in "LEO: Tracing GPU Stall Root Causes via Cross-Vendor Backward Slicing" (SC26). Contains measurement databases for three GPU platforms — NVIDIA GH200 (sm_90, Hopper), AMD MI300A (gfx942, CDNA3 APU), and Intel GPU Max 1100 (Ponte Vecchio, Xe-HPC) — across 15 RAJAPerf kernels plus miniBUDE, XSBench, LULESH, Kripke, llama.cpp, QuickSilver, and HipKittens RMSNorm. Companion to the source artifact at 10.5281/zenodo.19704349. The measurement databases enable LEO analysis on any x86-64 CPU-only machine without GPU hardware, which is required for reproducing Figure 5 (SDC coverage) in the paper. Size: ~1 GB compressed, ~5.6 GB extracted. Expected directory layout after extraction: per-workload directories named [vendor]-[machine]-[workload]-[timestamp]/ under results/, plus results/per-kernel/. Collected with the HPCToolkit feature/hidden-stall-metrics branch. See the companion source artifact (10.5281/zenodo.19704349) for the exact profiling commands and Dockerfiles.