Speaker
Description
As High-Performance Computing (HPC) scales toward Exascale, optimizing energy efficiency and resource utilization requires a unified, data-driven approach. The EuroHPC JU project SEANERGYS [1] addresses this by establishing a telemetry-to-intelligence pipeline built around three pillars: Monitoring (CMI), Data Analytics (AIDAS), and Resource Management (DSRM). The AIDAS transforms the CMI monitoring data into insights, predictions, and actionable feedback that can be used by DSRM. The main AIDAS component is the Model Zoo, a centralized framework containing a curated set of AI/ML models providing system-, node-, scheduler-, and job-centric predictions. To enable model training prior to full CMI deployment, anonymized “Day-0 datasets” were collected from eight supercomputing centers, standardized into the unified SEANERGYS format, and stored as compressed Apache Parquet [2] files in a centralized data repository. Day-0 datasets establish a baseline for pre-training and benchmarking the Model Zoo’s models, enabling their preparation prior to integration with live monitoring data streams. Day-0 datasets may also be made publicly available in the future through a data repository to support research and reproducibility.