GPU Dashboard

A web dashboard that summarizes GPU utilization across SSH-accessible nodes by running nvidia-smi on each host.

Setup

1. Requirements

• The machine running the dashboard must be able to SSH directly to each node
• ~/.ssh/id_ed25519_hakone must exist (or edit src/config.py to change SSH_IDENTITY_FILE)
• SSH options (User, IdentityFile) are passed explicitly; ~/.ssh/config is optional

2. Install and run

cd baulab-dashboard  # after: git clone git@github.com:wendlerc/baulab-dashboard.git
uv sync
uv run python -m src.main

Open http://localhost:5000

3. Expose via Cloudflare Tunnel (optional)

To access from outside your network:

# Terminal 1: start the dashboard
uv run python -m src.main

# Terminal 2: start the tunnel
./run-tunnel.sh

The tunnel prints a URL like https://random-name.trycloudflare.com — open it in a browser. No Cloudflare account required.

4. Always-on (systemd, survives reboot)

To run the dashboard automatically on boot with Flask + cloudflared:

1. Configure the env file

mkdir -p ~/.config/baulab-dashboard
nano ~/.config/baulab-dashboard/env

Option A: Cloudflare named tunnel (recommended) — no rate limits, stable URL:

CLOUDFLARE_TUNNEL_TOKEN=eyJ...
CLOUDFLARE_TUNNEL_URL=https://your-tunnel.example.com

Option B: Quick tunnel — no config needed; URL is printed to the log. Use ?api=URL when visiting to point the frontend at the tunnel.

2. Install the systemd user service

cd /path/to/baulab-dashboard

mkdir -p ~/.config/systemd/user
cp baulab-dashboard.service ~/.config/systemd/user/

# Edit the service file: update WorkingDirectory, ExecStart, and EnvironmentFile
# to match your paths (e.g. ~/code/baulab-dashboard)
nano ~/.config/systemd/user/baulab-dashboard.service

systemctl --user daemon-reload
systemctl --user enable baulab-dashboard
systemctl --user start baulab-dashboard

3. Enable user services at boot (without login)

loginctl enable-linger

4. Verify

systemctl --user status baulab-dashboard
journalctl --user -u baulab-dashboard -f

Startup logs are also written to /tmp/baulab-dashboard-startup.log.

How it works

• A background thread runs nvidia-smi on each node every 60 seconds via SSH
• The web UI polls /api/status every 30 seconds
• Nodes are queried in parallel; connection timeout is 15 seconds per node