Skip to content

Dashboard

The NØMAD dashboard provides real-time monitoring of your HPC cluster(s).

Launching

nomad dashboard

Open http://localhost:8050 in your browser.

Tabs

Cluster Tabs

Per-cluster views showing: - Node status (idle, allocated, down) - Partition utilization - CPU and memory pressure indicators - Active jobs count

Network

3D job similarity network visualization: - Jobs as nodes, colored by health - Edges connect similar jobs - "Safe zone" and "danger zone" regions emerge from data - Failure clustering analysis

Resources

Resource usage views for administrators: - CPU hours by cluster/group - Filter by cluster, group, and time period - Resource consumption patterns - Quietest hours for job scheduling

Activity

Job activity and history: - Recent job completions - Failed job categories - Interactive session monitoring (RStudio/Jupyter)

Planned Features

  • Education Tab: Visual proficiency trajectories and per-student breakdowns (CLI available now via nomad edu)

Remote Access

By default, the dashboard only listens on localhost. For remote access:

  1. SSH Tunnel (simple): 

       ssh -L 8050:localhost:8050 user@hpc-head

  2. Reverse Proxy (production): See System Install

GPU Panel

The node detail panel shows GPU metrics for any node with NVIDIA GPUs.

nvidia-smi utilization

The primary GPU utilization bar reflects the kernel-launch duty cycle reported by nvidia-smi. This is the standard metric available on all NVIDIA GPU nodes.

Real Utilization (DCGM)

When DCGM is active on a node, a second bar labeled Real Util appears below the nvidia-smi bar. Real Utilization is a weighted composite of pipeline activity counters (SM, Tensor, DRAM, GR engine) that reflects what the GPU is actually computing, not just whether any kernel is scheduled.

A significant gap between nvidia-smi utilization and Real Util is a signal worth investigating — see GPU Monitoring.

Workload badge

A color-coded badge below the utilization bars shows the dominant workload category for the node's GPUs: tensor-heavy compute, FP64 / HPC compute, memory-bound, compute-active, idle, and others. Colors follow the Okabe-Ito colorblind-safe palette. See Workload Classification for the full category definitions.

Health indicator

A WARN, HOT, or CRIT badge appears next to the GPU name when hardware health issues are detected via DCGM (PCIe replay errors, high temperature, ECC errors, or row remap failures). No badge means OK.

DCGM pill

A small DCGM pill next to the GPU name indicates that enhanced metrics are active for that node. Nodes without DCGM show the nvidia-smi bar only and no badge or Real Util row.