Desktop performance: the audit and the A/B bench

tools/hypr-bench/ (the bench, run with nix run .#hypr-bench, never installed). What the desktop costs in CPU, GPU, memory and frames, measured with the effects on and off, so an optimization is chosen from a number and not from a guess. Started on 06/11/2026. Nothing here changes the config: every proposal at the end waits for the owner's yes.

The stack that was measuredCopy Anchor Link

MEASURED on 06/11/2026 with hyprctl version, hyprctl monitors -j, hyprctl getoption, hyprctl plugin list and sysfs:

What can be measured here, and what cannotCopy Anchor Link

Why the root-only rows: Hyprland runs with CAP_SYS_NICE (to raise its own priority), which makes the process non-dumpable, so the kernel hides its fdinfo and smaps from its own user. RAPL has been root-only since the Platypus side channel. perf_event_paranoid is 2 here. --privileged reads exactly those files through sudo, and nothing else.

Tools that do NOT work here, checked on 06/11/2026: intel_gpu_top only speaks i915, not xe (bar); pidstat, perf and drm_info are not installed, and the bench needs none of them.

The frame counter is trusted because it matches the physics. In a test compositor at 60 Hz, one animated workspace switch added 21 frames: 350 ms, against the settle spring's 306 ms to rest plus its first and last frame (motion). With HyprGlass disabled at runtime the counter keeps counting (two unanimated switches, two frames), so it measures every scenario, including the ones with the glass off. FPS is a load measure here, never a latency measure: with VRR on, an idle output that composes 0 frames is the IDEAL, not a stall.

What the audit already shows (no bench yet)Copy Anchor Link

MEASURED on 06/11/2026 during normal use, not idle, so these are context, not results:

• Quickshell averaged 13% of one core over 11 h (ps lifetime CPU) and 7% in a 5 s sample, with 227 MiB RSS, about 100 MiB of VRAM and 4% of the render engine's cycles over its lifetime. Hyprland averaged 5.7% of a core over the same 11 h.
• Over those 11 h, DP-2 composed 2 246 720 frames and DP-1 1 818 814 (HyprGlass's counters since the session began): 55.6 and 45 frames per second on average. DP-2 is the LG, whose screen is mostly the glance band. A Twitch stream on DP-1 explains DP-1's share; the LG's 55.6 average is the number the bench has to explain.
• Static reading of the shell (06/11/2026) lists what can keep the LG composing while nobody touches it: the 1 Hz clock with seconds on the band (dash/ClockWeather.qml, fed by bar/Bar.qml's SystemClock), the 1 Hz ping that redraws the bar's net pill (widgets/PingProbe.qml), the page dots that grow 1 px every 682 ms and 455 ms (dash/PageDots.qml, dash/ServiceStrip.qml), the infinite pulses of a degraded service or a tile in alert, and a fork every 2, 3 and 5 s. Each of those repaints a layer that HyprGlass and the blur then re-render. These are HYPOTHESES until the idle workload measures them.

The benchCopy Anchor Link

ScenariosCopy Anchor Link

Each one is applied with hyprctl eval on top of a fresh hyprctl reload config-only, so a block never inherits the one before:

E answers whether HyprGlass costs more than the blur it replaced. --shell-off adds qs ipc call bar hide to every scenario, which stops the bar and the band from drawing while their timers and forks go on, so it splits Quickshell's render cost from its polling cost.

WorkloadsCopy Anchor Link

The bench opens three named workspaces with static terminals of class hypr-bench (two on bench-a and one on bench-b, on the primary; one on bench-c, on the secondary) BEFORE any measurement, so no app start-up lands inside a block, and both screens show the same content in every scenario.

Every block waits 5 s after the scenario changes. Three rounds by default, the order of the scenarios rotated each round so a slow drift (heat, a timer) spreads over all of them. The aggregate is the mean over rounds with its standard deviation. A full run takes about 35 minutes; --quick (one round, short blocks) about 5.

The safety of the deskCopy Anchor Link

• Refuses to start with the session locked, a window of yours fullscreen, a Sunshine stream live (UDP 47998 to 48000), or the GPU already more than 50% busy (a game or a video would swamp the deltas). A block that saw the lock, a fullscreen window, a stream, a change of power profile or a cursor move is kept in rows.csv and marked invalid, never averaged.
• Holds a logind idle inhibitor for the run (systemd-inhibit --what=idle): without it, hypridle locks the screen AND starts power-idle-eco at 300 s, which drops the GPU floor and changes what is being measured. The inhibitor dies with the process, even on SIGKILL.
• Never touches a monitor, a mode, VRR, tearing, direct scanout, a driver or a service.
• Restores in three layers. The finally of the run (also reached by Ctrl+C and SIGTERM) reloads the config without a modeset, closes the panel, unhides the bar, kills only its own terminals and vkcube, and focuses each monitor's workspace back; a systemd-run --user timer armed at the start does the same if the process dies without its finally; and the restore compares getoption for every key it touched against the start, exiting non-zero if one differs. hypr-bench restore runs the same path by hand.

Running itCopy Anchor Link

1. Close games and videos, leave the desk alone (a cursor move invalidates the block).
2. nix run .#hypr-bench -- run --quick to check the mechanism, then nix run .#hypr-bench -- run for the real thing. Add --privileged for Hyprland's own GPU and PSS and the CPU package power (it asks for the sudo password once, at the start).
3. Results land in ~/.local/state/hypr-bench/<date>/: raw.json (every block and the environment), rows.csv, summary.json and report.md. hypr-bench report <dir> rebuilds the last three from raw.json.

hypr-bench probe prints two seconds of every metric and changes nothing: the first thing to run when a number looks wrong.

ResultsCopy Anchor Link

None yet: the bench was written on 06/11/2026 while the desk was in use (a game on DP-1, which the load guard refused, as designed). This section gets the tables and their date from the first full run.

Proposals, waiting for the benchCopy Anchor Link

Hypotheses only, ranked by what they would save if the bench confirms them; none is applied.

RejectedCopy Anchor Link

• A nested Hyprland as the test bed. On 06/11/2026 the API check ran in a Hyprland nested in a headless output of the live one, as the HyprGlass tests did in October. On exit, Hyprland 0.57.1 runs systemctl --user stop graphical-session.target and unsets DISPLAY, HYPRLAND_INSTANCE_SIGNATURE and others in the user manager (both strings are in the binary), so the nested one took the LIVE session's units down: Sunshine, hypridle, the portals, and the apps scoped to the session. hypr-session-ensure brought the target back, but with the dead nested instance's signature, because it picks the newest directory in $XDG_RUNTIME_DIR/hypr. Repaired by hand the same night. The bench needs no nested compositor: hyprctl eval and reload config-only were checked there before it went wrong, and their effect is checked again by getoption at every restore.
• hyprctl keyword for the toggles: dead under the Lua config (keyword can't work with non-legacy parsers).
• Killing Quickshell to isolate it: the bar, the band and the notifications would vanish, and its own counters already isolate its CPU and GPU. --shell-off hides it through its IPC instead.
• intel_gpu_top for the engines: i915 only.