simple-system-monitor: fix GPU and CPU sensors, CPU usage and network rates - #1889
Conversation
… rates The GPU temperature was read from thermal_zone0, the same motherboard sensor used for the CPU, so the GPU row showed the CPU value on every install. It is now read from the GPU itself: nvidia-smi for NVIDIA, hwmon for AMD and Intel. The CPU sensor is likewise detected rather than defaulting to thermal_zone0, preferring the on-die package sensor. CPU usage counted iowait as busy time and left out nice, irq and softirq, so a slow disk showed as CPU load while real work went uncounted. It is now measured as the share of time that is neither idle nor waiting on I/O. Network traffic was summed over every activated interface including loopback and any VPN, so traffic through a tunnel was counted twice, once on the tunnel and again on the hardware carrying it. Only physical interfaces are counted now. The rate itself divided bytes by milliseconds and labelled the result KB, and the label was missing the per-second part. The refresh interval becomes a setting, keeping 1 second as the default. GPU temperature is polled no more than once every 2 seconds since it costs a subprocess. The settings object was rebuilt on every settings change while the previous bindings stayed live, so each change cost more than the one before it.
Best-practices scannerThis is a regex-based check for API usage that can pose security, performance or This check is not perfect and will not replace a normal review.Found 6 potential issue(s):
|
1 participant
@arielandrade
Thanks for this desklet, it has been on my desktop for a while. While looking at
why the GPU row never seemed to move I found a few things worth fixing, all in
the 5.8 version.
The GPU temperature never read the GPU.
customGPUPathandcustomCPUPathboth default to
/sys/class/thermal/thermal_zone0, and that default is writteninto the saved settings on first run, so the GPU row has always shown the same
motherboard sensor as the CPU row. On my machine the two rows sat on an identical
number permanently, which is what made me look.
The sensors are now detected instead. NVIDIA is read through
nvidia-smi, sincethe proprietary driver publishes no hwmon entry, and that call is asynchronous
and rate limited to once every 2 seconds so it cannot stall the shell. AMD, Intel
and nouveau are read from hwmon. The CPU prefers the on-die package sensor
(
coretemp/k10temp) and falls back tox86_pkg_temp, thenthermal_zone0.A path the user has actually chosen still wins. Because the old default is sitting
in existing users' settings files,
/sys/class/thermal/thermal_zone0is treatedas "detect automatically" rather than as a deliberate choice, otherwise the fix
would not reach anyone who has already run the desklet.
CPU usage was measured wrongly in both directions. It used
(user + sys + iowait) / total, which counts iowait as busy, so a slow disk readas CPU load, while
nice,irqandsoftirqwere not counted at all. It is now(total - idle - iowait) / total.Network traffic was counted more than once.
update_connectionssums everyactivated NetworkManager device, which on a machine with a VPN includes both the
tunnel and the physical card carrying it, plus loopback. My figures were roughly
doubled whenever the VPN was up. Only interfaces with a
deviceentry in sysfsare counted now, with a fallback to the old list if that leaves nothing.
Network rate units. The rate divided bytes by milliseconds and labelled the
answer KB, which is about 2.4 % high, and the label had no per-second part, so a
speed read like a total. Now computed as bytes per second and shown as
B/s,KB/sorMB/s.Refresh interval is now a setting, defaulting to the existing 1 second, range
0.5 to 30. This one is a feature rather than a fix, so I am happy to drop it from
the PR if you would rather keep the interval fixed.
Settings object was rebuilt on every change.
setupUIcreated a newDeskletSettingsand rebound every property each time it ran, while the previousbindings stayed live and also called
setupUI, so each settings change cost morethan the one before. It is now built once, with the UI rebuild split into its own
function. The
changed::desklet-decorationshandler had the same problem.Tested on Cinnamon 6.6.7, Fedora 44, on a laptop with an Intel UHD iGPU and an
NVIDIA RTX 2060, with Proton VPN and Tailscale both up. CPU and GPU temperatures
were cross checked against
nvidia-smiand thecoretemphwmon entry, and thenetwork figures against the raw interface counters. Versions 3.0 and 4.0 are
untouched.
🤖 Generated with Claude Code