gpumanager

Monitor for a shared GPU workstation (Windows + WSL2 Ubuntu, multiple SSH users). Everyone gets a live dashboard showing who is logged in, each user's CPU / RAM / VRAM usage, and limit breaches — so people stop stepping on each other's runs.

• Backend: Python (FastAPI + psutil + nvidia-smi), runs inside WSL as a systemd service, samples every 3 s.
• Frontend: React (Vite), served by the backend at http://<host>:8595.
• CLI: gpuwho prints the same info in any SSH session.
• Alerts: global thresholds (combined VRAM/RAM/CPU) and per-user limits. On breach: wall broadcast to every SSH session, direct message to the offender's terminal, optional Slack/Discord webhook for the admin, and the dashboard shows it. Repeats are suppressed for cooldown_minutes.

Install (inside WSL Ubuntu)

1. Enable systemd in WSL (once), in /etc/wsl.conf:

   [boot]
   systemd=true

   Then from Windows: wsl --shutdown and reopen WSL.

2. Build the frontend (needs Node; can be done on Windows or in WSL):

   cd frontend && npm install && npm run build

3. Install the service:

   sudo bash deploy/install.sh

4. Set your limits in /etc/gpumanager/config.yaml (see config.example.yaml for every option), then sudo systemctl restart gpumanager.

Dashboard: http://<workstation>:8595 · API: /api/status · CLI: gpuwho

Reaching the dashboard from other machines

WSL2 is NAT'd by default. Two options (you likely already solved this for SSH):

• Mirrored networking (Windows 11, recommended): in %UserProfile%\.wslconfig set [wsl2] networkingMode=mirrored, then wsl --shutdown. WSL ports are then reachable on the Windows host's LAN IP directly.

• Port proxy: in an admin PowerShell on Windows:

  netsh interface portproxy add v4tov4 listenport=8595 listenaddress=0.0.0.0 connectport=8595 connectaddress=(wsl hostname -I).Trim()
  New-NetFirewallRule -DisplayName gpumanager -Direction Inbound -LocalPort 8595 -Protocol TCP -Action Allow

How measurement works (and its limits)

• Sessions: from utmp (same source as who), with idle time from tty activity — users see who is on and whether they're actually active.
• CPU / RAM per user: aggregated over each user's processes. cpu_percent is percent of the whole machine (all cores = 100). RAM is summed RSS, which double-counts shared pages slightly. Totals are for the WSL VM, i.e. the RAM/CPU you allotted WSL in .wslconfig — not the whole Windows box.
• VRAM per user: from nvidia-smi --query-compute-apps, mapping each GPU process to its owner. Caveat: some WSL driver versions don't report per-process GPU memory. gpumanager then falls back to scanning /dev/dxg handles: it still shows which users have GPU work running (and card-level totals stay exact), but per-user GB shows as "GPU active" instead of a number. Update the NVIDIA Windows driver to get per-process numbers.
• Alerts notify, they don't enforce — nothing is killed or throttled. Hard enforcement (per-user cgroups for RAM/CPU, CUDA_VISIBLE_DEVICES / MPS for GPU) is a natural next step.

Development

# backend (WSL):  python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
.venv/bin/python -m gpumanager --config config.example.yaml
# frontend (hot reload, proxies /api to :8595):
cd frontend && npm run dev

Note: the dashboard has no authentication — anyone on the LAN can view it. Keep it on the trusted network, or put a reverse proxy with auth in front.