Scholarlink

To diagnose issues in HPC data centers, we compare what jobs are doing with facility signals (temperature, power) across many nodes, and we link nodes, jobs, and users. We present two interactive views for incident analysis: PowerMap, a node–time view for one metric at a time with thresholding and user/job provenance, and Time-Indexed PCA Trajectories, a multi-metric projection (PC1–PC2) with optional time paths and drill-down (radar and line charts). Applied to 110 CPU-only nodes over 48 h (29–31 Jul 2025, UTC−05) that include a high room-temperature event (30 Jul, \(\sim \) 15:00–18:00), PowerMap shows cluster-wide temperature alerts, higher system power, and full CPU usage. A PCA snapshot (17:00) shows a user-colored cluster with loadings aligned to higher cpu_usage, cpu_power, and temperature; provenance links many concurrent jobs to user100049 (momaruf). Overall, heavy workload set the stage, and a facility temperature rise pushed many nodes past limits. The views support quick triage, clear attribution, and multi-metric monitoring.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

PowerMap and Time-Indexed PCA Trajectories: Interactive Visual Forensics for HPC Incidents

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Phornsawan Roemsri,
  • Amauri Henrique Ribeiro,
  • Tommy Dang

摘要

To diagnose issues in HPC data centers, we compare what jobs are doing with facility signals (temperature, power) across many nodes, and we link nodes, jobs, and users. We present two interactive views for incident analysis: PowerMap, a node–time view for one metric at a time with thresholding and user/job provenance, and Time-Indexed PCA Trajectories, a multi-metric projection (PC1–PC2) with optional time paths and drill-down (radar and line charts). Applied to 110 CPU-only nodes over 48 h (29–31 Jul 2025, UTC−05) that include a high room-temperature event (30 Jul, \(\sim \) 15:00–18:00), PowerMap shows cluster-wide temperature alerts, higher system power, and full CPU usage. A PCA snapshot (17:00) shows a user-colored cluster with loadings aligned to higher cpu_usage, cpu_power, and temperature; provenance links many concurrent jobs to user100049 (momaruf). Overall, heavy workload set the stage, and a facility temperature rise pushed many nodes past limits. The views support quick triage, clear attribution, and multi-metric monitoring.