HPC Bottleneck Detection Dataset: Labelled Performance Counter Traces

This dataset contains labelled hardware performance-counter time series used to train and evaluate machine-learning and rule-based HPC bottleneck detectors, collected via the xbat continuous benchmarking tool. It has three parts: training_corpus/, 20 real HPC applications profiled on an Intel Xeon 8360Y node, weak-supervision labelled via a heuristic detection engine hsuper_dataset/, 7 held-out applications profiled on the HSUper cluster on matching hardware, used as an out-of-sample generalization test hpas_fault_injection/, synthetic anomaly ground truth from HPAS Performance Anomaly Suite covering the 6 bottleneck anomalies, captured on both the Intel node and an AMD EPYC 9654 (Genoa) node. See MANIFEST.md for the full job-to-application mapping, labelling parameters, and output format. Licensed under CC BY 4.0.