Endura

Measured, reversible battery-life tuning for Linux laptops — with a GUI.

中文说明

Endura is a small root daemon, two user agents and a PyQt6 app. It applies a set of power tweaks that were each A/B-measured on real hardware, can switch every one of them on and off individually, and records per-app CPU time and wakeups so you can see what is actually draining your battery.

It started as an investigation on one laptop that had been "fixed" by clamping the CPU to 1.2 GHz and still only lasted ~6 hours. After the work below, the same machine idles at 4.4 W with the screen on at 30 % (≈ 22 h on its 99 Wh battery) while staying more responsive than before. See the case study.

What it does

Tested and not adopted (no measurable gain on the reference laptop): AMD ABM levels 1/3, 120 Hz + VRR vs 60 Hz, SMT off, Bluetooth USB autosuspend, cgroup CPU quotas for chatty apps (quotas do not stop wakeups — only freezing does).

Smoothness

Saving power by making the machine sluggish was explicitly not allowed. Endura's criterion, checked by endura-bench:

From idle, a UI-like task that takes 10 ms at full performance must finish with p90 ≤ 16.7 ms (one 60 Hz frame); a 2 ms task p90 ≤ 8.3 ms (one 120 Hz frame).

Telemetry and tools

• Overview: live battery and SoC power, 1 h graph, every tweak with status (active / battery only / needs reboot / not available on this hardware).
• Apps: per-app CPU time and wakeups per second over 1 h … 7 days (battery time only), plus a freeze probe that measures an app's real power by freezing and thawing its cgroup a few times. Wakeups often matter more than CPU %: a backend at 3 % CPU but 3000 wakeups/s cost 4 W.
• Smoothness: runs the benchmark above.
• lab/ has the A/B scripts used for every number on this page.

Requirements

• systemd, cgroup v2, Python 3.10+, PyQt6, pulsectl, dbus-python, PyGObject, PipeWire (pulse)
• Laptop with a battery exposed in /sys/class/power_supply
• Best support: AMD Ryzen APUs (SoC power telemetry, amdgpu PSR, TSC fix) and KDE Plasma 6 on Wayland (focus tracking for App Nap, panel monitor hiding). Generic tweaks work on Intel and other desktops; tweaks whose hardware is missing show as not available.

Tested distributions

endura-install resolves dependencies per distribution, installs what is packaged, vendors the rest into /usr/lib/endura/vendor (pulsectl; PyQt6 from PyPI where the distro has none), and disables a feature instead of failing if something is still missing. Verified in clean containers (dependency installation and imports for /usr/bin/python3):

The GUI connects to the daemon as the desktop session's user or any wheel/sudo/admin member (checked with SO_PEERCRED), so no group setup is needed. KDE-specific parts (App Nap focus tracking, hiding panel monitors) are skipped elsewhere; on other desktops App Nap naps apps only when there is no keyboard/mouse input.

Install

From PyPI:

pipx install endura                # or: pip install --user endura
sudo "$(which endura-install)"     # installs deps (dnf / apt / pacman), daemon, GUI
endura-gui                         # or "Endura" in the application menu

From source:

git clone https://github.com/qvictl/endura.git
cd endura
sudo ./install.sh

The system services run from a self-contained copy in /usr/lib/endura with the distribution's /usr/bin/python3, so they do not depend on where pip put the package.

Most tweaks are on by default and apply only on battery. Reboot-scoped ones (TSC resync, PSR, stutter) are off until you enable them in the GUI. For the TSC fix, install tscsync first — Endura only toggles it.

sudo "$(which endura-uninstall)"   # or ./uninstall.sh: reverts every tweak first, then removes Endura

Every file a tweak modifies is backed up in /var/lib/endura/backup; module options go to a separate override (/etc/modprobe.d/zz-endura-amdgpu.conf) instead of editing yours.

Configuration

/etc/endura/config.json:

{
  "enabled": {"app_nap": true, "amdgpu_psr": true},
  "options": {
    "battery_epp": "balance_performance",
    "slack_deep": ["sunloginclient", "ToDesk_Service"],
    "nap_apps": {
      "com.slack.Slack": {"globs": ["app-flatpak-com.slack.Slack-*.scope"], "cls": ["slack"], "tier": 1}
    }
  }
}

Defaults and all keys: endura/defaults.py. nap_apps matches the app's systemd cgroup (flatpak scopes, KDE/GNOME app-*.service units); tier 1 naps when unfocused, tier 2 only after 5 minutes without input.

Architecture

The daemon itself was tuned too: a 5 s loop, few ACPI EC reads (each power_supply read is an EC transaction on many laptops), batched SQLite commits, no PAM sessions per user command. It costs ≈ 0.5 % of one core.

Case study

Reference machine: Lenovo ThinkBook 14 G8+ AHP — AMD Ryzen 7 H 255 (Zen 4, Radeon 780M), 32 GB LPDDR5X, 3072×1920 120 Hz IPS panel, Intel AX210, 99.9 Wh battery, Fedora 43, KDE Plasma 6 Wayland, kernel 7.0. Daily apps that had to stay open: Feishu, Chrome, Sunlogin, Discord, v2rayN, WeChat. Constraints: Wi-Fi and Bluetooth on, 30 % brightness, no loss of smoothness.

Starting point. A previous "fix" ran a feedback loop that lowered scaling_max_freq until battery draw fell under 12 W, plus cgroup CPU quotas on the chat apps. It sat at the 1.2 GHz floor all day; the machine drew 14–16 W and lasted ≈ 6 h. A 10 ms UI task took 40 ms.

Investigation, in the order things were found (whole-system battery power, A/B, screen on, apps open unless noted):

1. The clamp was counter-productive. Removing it lowered draw by ~1.5 W: slow cores stay in C0 longer and keep the fabric awake. Power-saver's EPP power was the next culprit: balance_performance with boost off was −2.07 W and 2.6× faster.
2. A browser tab played digital silence for hours, keeping the codec and speaker amp on: −1.2 W when the stream was parked on a null sink. This became the audio gate.
3. The kernel was on HPET. dmesg: Measured 3168609404 cycles TSC warp between CPUs, turning off TSC clock. The BIOS boots CPU0 ~0.8 s behind the other 15 threads; AMD has no TSC_ADJUST. Every clock_gettime() became a syscall plus a slow MMIO read (one chat backend did 3000 per second). Fixed with tscsync; offsets after boot ±16 cycles.
4. amdgpu.dcdebugmask=0x12 (an old flicker workaround) disabled both PSR and memory stutter: FCLK was pinned at 1875 MHz. Re-enabled; FCLK idles at 400 MHz, no flicker so far.
5. One chat app's backend was stuck in a busy-wake loop: its Rust task scheduler woke 15 workers every 5 ms (~3000 wakeups/s at 3 % CPU). Freezing just that process dropped the whole system from 14.8 W to 10.1 W — ~4 W from one process. An update + restart fixed it; cgroup CPU quotas (what the old setup used) did nothing because they do not stop wakeups.
6. A remote-desktop daemon was in a start-timeout loop (Type=forking, never forks; 113 restarts) and, once fixed, still polled every 10 ms on 6 threads. Timer slack took it from 586 to 16 wakeups/s without affecting remote sessions.
7. Freeze probes per app with a static screen: chat app #1 0.51 W, chat app #2 0.30 W, browser 0.08 W, Discord ≈ 0, all four together 1.35 W (more than the sum — the package only reaches deep idle when everything is quiet). This became App Nap: −0.40 W with the chat app in the background.
8. The screen was never static. Tracing drm_vblank_event showed ~55 flips/s on an idle desktop. Freezing cgroups one by one pinned it on the Plasma panel's seven system-monitor widgets. Hidden on battery → 0 flips/s, PSR residency ~90 %.
9. The SSD never slept: 57 writes/s from writeback and 30 s btrfs commits. commit=120,noatime on battery halved the write count.
10. Housekeeping: an updater crash-looping every 10 s (1286 restarts) → crash-loop guard; the daemon's own cost cut from 2.5 % to 0.5 % of a core.

Static screen, 30 % brightness, all apps open:

Remaining floor (all apps frozen, PSR active, CPU < 2 %): SoC ≈ 2.9 W (display engine, memory controller, fabric), panel ≈ 1.2 W of which the 30 % backlight is ≈ 0.6 W.

Evaluated and rejected: rebuilding the kernel for CONFIG_RCU_LAZY / the TEO idle governor (Fedora ships neither; RCU woke only ~15×/s here — not worth a custom kernel on every update).

License

MIT. third_party/tscsync/set-mode.sh is from tscsync (MIT).