Abstract
NØMAD (NOde Monitoring And Diagnostics) is a lightweight monitoring and predictive analytics tool designed for computing infrastructure that requires minimal deployment overhead. At its core, metric collectors scheduled via systemd timers gather system metrics—disk, CPU, memory, I/O, and GPU utilization—from standard Linux tools, storing everything in a single SQLite database. When SLURM is available, these collectors extend to capture job-level analytics from scheduler commands and per-process I/O statistics, enabling deeper insight into workload behavior. The system employs machine learning (ML) models to predict job failures before they occur, while a data readiness estimator helps administrators determine when sufficient historical data has been collected for reliable predictions. Beyond prediction, diagnostic tools provide targeted analysis of network performance, storage health, and node-level bottlenecks. A key innovation is modeling jobs as nodes in a similarity network, where edges connect jobs with comparable resource usage patterns; because jobs with similar characteristics tend to share similar outcomes, the network structure reveals failure-prone regions in the feature space. The tool includes a real-time web dashboard for visualization and supports alerts via email, Slack, or webhooks. By requiring no external databases or complex infrastructure, NØMAD is particularly well-suited for small-to-medium high-performance computing (HPC) centers and research groups seeking sophisticated monitoring without enterprise-scale overhead.