Ma B, Eitzinger J, Wellein G (2026)
Publication Language: English
Publication Type: Conference contribution, Original article
Publication year: 2026
Publisher: Springer Nature
Series: A Springer Nature Computer Science book series (LNCS)
Conference Proceedings Title: HIGH PERFORMANCE COMPUTING: ISC High Performance 2026 International Workshops
Event location: Hamburg, Berlin
Many HPC centers collect fine-grained telemetry data yet still lack a systematic picture of which applications are running and how efficiently they exploit the hardware. We close this observability gap with a bidirectional pipeline that combines Roofline-grounded telemetry analysis with interpretable tree-based learning. On Fritz, a TOP500-class supercomputer, 2.6 M per-node telemetry records from 0.85 M Slurm jobs are labeled via weak supervision and quality-filtered to 1.6 M records from 0.52 M jobs, then transformed into compact behavioral feature vectors. In the forward direction, a One-vs-Rest XGBoost classifier achieves Micro-F1 0.905 and Macro-F1 0.904 at sub-second inference latency, confirming balanced accuracy across dominant and long-tail workloads. In the backward direction, per-application clustering recovers a stable, physically interpretable dominant execution regime per workload; minority clusters are detected but their attribution reaches a scope boundary that telemetry alone cannot cross without application-level context. The identified scope boundary also opens a natural next step: coupling telemetry-derived regime labels with application-level metadata to enable finer-grained, energy-aware scheduling decisions.
APA:
Ma, B., Eitzinger, J., & Wellein, G. (2026). Automatic Workload Characterization on Production HPC Systems via Roofline Telemetry. In HIGH PERFORMANCE COMPUTING: ISC High Performance 2026 International Workshops. Hamburg, Berlin: Springer Nature.
MLA:
Ma, Bole, Jan Eitzinger, and Gerhard Wellein. "Automatic Workload Characterization on Production HPC Systems via Roofline Telemetry." Proceedings of the 7th ISC HPC International Workshop on “Monitoring, Observability, and Operational Data Analytics”, Hamburg, Berlin Springer Nature, 2026.
BibTeX: Download