Explainable Anomaly Detection for HPC Storage during Computational Modelling: A Reproducibility-Centred Evaluation

Abstract

Scientific-computing storage experiences bursts of parallel reads and checkpoint writes that can resemble genuine infrastructure anomalies. An evaluation that overlooks these constraints may reach conclusions that cannot be reproduced. The research question is whether it is possible to separate workload-driven congestion from storage faults while maintaining interpretable alerts. Previous work provides context for sampling, input validation and study boundaries [3–7]. The planned analysis uses a trace-replay experiment using computational-model I/O patterns and injected failures under controlled concurrency. Prior work on storage operations provides technical context for validation [1], while complementary systems evidence motivates service-continuity checks [2]. Under matched evaluation conditions, the protocol compares static queue thresholds, robust multivariate detectors and explanation-producing anomaly models. Methodological sources are used to specify transparent comparison and reporting practices [8–12]. The reproducibility extension records dataset versions, model settings, exclusion rules and provenance before any held-out evaluation. Performance will be judged using tail latency, false alerts per workload, fault lead time and root-cause agreement with diagnostic logs. Interpretation will account for cache warm-up, burst timing, shared-file interference and operator interpretation of alarms. An independent replication exercise would repeat the main analysis from the documented artefacts and quantify disagreements. The supplementary technical literature is used to identify plausible failure mechanisms [13–17]. No study result is assumed in advance; the analysis is designed to show both when the method helps and when it fails, with reporting context from the remaining sources [18–23].