Linux Resource Monitor

A GTK3 system-tray monitor for CPU, RAM, every GPU and disk/NAS I/O, plus a live "Mission Control" HTML dashboard that the tray writes to ~/Desktop/DASHBOARD.html.

Features

• Auto-detects every GPU on the machine, any vendor, any count — no card names or indexes are hardcoded, so plugging a card in or pulling one out is picked up on the next tick without touching the code.

• Per-vendor load sources, chosen from what the kernel/driver actually exposes:

  Board	Load source
  NVIDIA	NVML (libnvidia-ml.so.1) via ctypes; falls back to nvidia-smi
  AMD (amdgpu)	/sys/class/drm/cardN/device/gpu_busy_percent
  Intel (xe/i915)	DRM fdinfo engine counters (busy cycles vs total cycles, or busy ns)
  anything else	fdinfo if present, otherwise the card is still listed with the stats that do exist

• Expandable dashboard cards (<details>/<summary>): collapsed each card shows the load bar, temperature and wattage; expanded it carries every temperature sensor the driver publishes, power (and cap), fan, clocks, voltages, per-engine load and device identity. Open/closed state is kept in localStorage, so it survives a page refresh.

• Consistent cards: every GPU card carries the same sections and row labels, and a metric a driver does not publish says n/a rather than vanishing — the 780M has no fan sensor, so it shows Fan: n/a next to cards that show a real reading. Only genuine hardware differences remain: sensor count, vendor clock names, engine counters. Vendor extras (voltage rails, Intel GT idle state, the NVIDIA thermal limits) are appended to the shared rows.

• Toolbar controls: text size (A- / A+, 85–150 %, plus Reset) and a light/dark switch. Every size in the stylesheet is in rem, so the text, the load bars and the padding scale together. The stylesheet is driven by custom properties, so a theme switch is a variable swap on <html>.

• Honest telemetry: a figure the driver does not publish is shown as n/a and explained, never invented. Intel VRAM size is read from the Vulkan device-local heap (validated against nvidia-smi's memory.total); Intel VRAM usage is summed over DRM clients and labelled as such, because xe exposes no device-wide memory counter.

• Storage cards for local disks plus /mnt and /media mounts, with per-device read/write rates.

Requirements

• python3-gi, gir1.2-ayatanaappindicator3-0.1, python3-psutil
• Optional: vulkaninfo (VRAM size for boards whose driver publishes none, e.g. Intel xe), lspci (friendly card names)
• A graphical session: the script needs DISPLAY/WAYLAND_DISPLAY and DBUS_SESSION_BUS_ADDRESS.

Run it with the system python3. It needs the distro gi module, which a project virtualenv does not have.

Install

install -m 0755 resource_monitor.py ~/scripts/resource_monitor.py
install -m 0755 packaging/nate-systray ~/.local/bin/nate-systray
install -m 0644 packaging/nate-systray.desktop ~/.local/share/applications/
install -m 0644 packaging/nate-systray.autostart.desktop ~/.config/autostart/nate-systray.desktop
update-desktop-database ~/.local/share/applications

Then either launch it from the app grid, run nate-systray, or let the autostart entry bring it up at login. The tray menu has "Open Mission Control (Web)" and "Exit Monitor".

packaging/nate-systray is a wrapper that refuses to start a second icon, prefers the live script and falls back to the bundled binary, logs to ~/.local/state/nate-systray.log, and needs a graphical session.

Verify

python3 test_gpu_detect.py

Self-contained offline suite (no GUI interaction): card enumeration, per-vendor load sources, the cycle-ratio and nanosecond load maths against synthetic fixtures, Vulkan parsing (including a cross-check that the number for an NVIDIA card equals nvidia-smi), n/a-vs-zero semantics, dashboard structure and idempotency (no duplicated regions, stable file size), the expanded-card panel contents, the theme/text-size controls, and per-card row/section parity.

Dashboard notes

The dashboard is a static file rewritten every 2 seconds; the browser shows whatever it last loaded, so refresh to re-read values. Card expand state, text size and theme all persist across refreshes where the browser allows localStorage on file:// URLs (Firefox does), and the theme and text size are applied by a <head> script before first paint so there is no flash of the wrong theme.

The monitor owns the whole <style> block plus the two marked <script> blocks and the toolbar; those are regenerated on startup, so hand edits there will be overwritten. The card/table markup outside them is preserved.

Layout

resource_monitor.py          tray app + dashboard writer
test_gpu_detect.py           offline verification suite
packaging/                   launcher wrapper + desktop entries
bin/resource_monitor         prebuilt standalone binary (see note below)

bin/resource_monitor is a PyInstaller build and is stale — it predates the multi-GPU work. Rebuild with pyinstaller resource_monitor.spec when a fresh standalone binary is wanted.