Skip to content

Latest commit

 

History

History
192 lines (150 loc) · 9.33 KB

benchmarks.md

File metadata and controls

192 lines (150 loc) · 9.33 KB

Benchmarks

Environment and method: methodology.md. Raw data: benchmarks/results/. Every number below was measured; where Raze loses, the table says so.

1.1.0

r6 vs systemd — same Raze Desktop, only PID 1 differs (5 interleaved boots each)

benchmarks/desktop-init.py, systems from distro/configurations/vm-desktop.nix (benchmarks/variants/desktop-systemd.nix for systemd). QEMU/KVM, 4 vCPU, 3 GiB, no 3D, warm-up boot discarded. Milestones are process start times read from /proc in the guest; memory 20 s after the greeter started.

Metric (median ± stdev) r6 systemd
Kernel start → greeter (greetd started) 5.08 ± 0.04 s 6.41 ± 0.45 s
Kernel start → NetworkManager started 5.08 ± 0.05 s 6.23 ± 0.45 s
System memory used (MemTotal − MemAvailable) 355 ± 6 MB 379 ± 7 MB
Userspace processes 24 19
Userspace PSS (sum) 93.5 MB 73.0 MB
PID 1 RSS 2.6 MB 15.0 MB

Reading: r6 reaches the login screen 1.3 s sooner and the system uses 24 MB less memory. systemd runs fewer processes and less userspace memory at the greeter because it starts nix-daemon (25.6 MB PSS) through socket activation on first use; r6 has no socket activation yet and starts it at boot. Like systemd, r6 leaves udisks2, upower and rtkit to D-Bus activation (that change took r6 from 105 to 94 MB userspace PSS and from 29 to 24 processes).

r6 vs s6 — Raze Core, 10 interleaved boots each (r6 1.1)

Metric (median ± stdev) r6 s6
PID 1 start → last of 20 services started 32.88 ± 2.17 ms 31.63 ± 1.39 ms
Kernel start → last service started 235.22 ± 4.66 ms 232.11 ± 3.97 ms
Init + supervision processes 2 27
Supervision RSS (sum) 1830 ± 35 KiB 3166 ± 15 KiB
Supervision PSS (sum) 1818 ± 35 KiB 1217 ± 16 KiB
System memory used 50.96 ± 3.45 MB 53.49 ± 0.59 MB

Same picture as 1.0: activation latency equal within noise (s6's median is 1 ms lower), r6 with 25 fewer processes and lower RSS, s6 with lower PSS.

Rift vs niri — same r6 system, virgl GPU (3 interleaved runs each)

benchmarks/compositor.py, systems from benchmarks/variants/desktop.nix. Both render with GLES on the virgl GPU (Rift reports drm-gles). Session = every process of the logged-in user: Rift + Rift Shell + clipboard history, or niri + waybar + mako

  • swaybg; PipeWire, portals and the terminal in both.
Metric (median of 3) Rift + Rift Shell niri + waybar session
Compositor CPU, idle, per 60 s 0.00 s 0.05 s
Session CPU, idle, per 60 s 0.01 s 0.08 s
Compositor RSS 75 MB 104 MB
Compositor PSS 61 MB 83 MB
Session PSS 217 MB 244 MB
System memory used 504 MB 513 MB
Compositor CPU, terminal printing every 50 ms, per 10 s 0.48 s 0.50 s
Session CPU, same workload, per 10 s 0.79 s 0.81 s

Reading: both compositors cost the same under this workload; at idle Rift does no work at all and its session uses 27 MB less memory. The session was not lighter at first; these changes, each measured with this benchmark, got it there:

Rift session change Session PSS System memory used
First measurement (Rift Shell on OpenGL, all overlays loaded, Xwayland at login) 352 MB 600 MB
Overlays created on demand (LazyLoader) 344 MB 596 MB
Rift Shell on Qt's software renderer + Xwayland started on first X client 217 MB 504 MB

(Compositor PSS rose from 42 to 61 MB in the last step only because Rift no longer shares the GL driver pages with the shell; its RSS is unchanged at 75 MB.)

Rift Shell memory

Change Rift Shell PSS, idle (headless Rift, Qt software renderer)
All overlays created at start 130 MB
Launcher, control centre, power menu, clipboard created on demand (LazyLoader) 99 MB

Rift Shell starts Qt's software renderer when Rift composites in software: in the no-3D VM, Quickshell's RSS was 276 MB with OpenGL on llvmpipe and 123 MB with Qt's software renderer.

Idle CPU (no-3D VM, desktop logged in, 30 s)

Process CPU ticks in 30 s
Rift (pixman) 0
Rift Shell (Quickshell) 1

1.0.0

r6 vs s6 — Raze Core, 10 interleaved boots each

QEMU/KVM, 2 vCPU, 256 MiB, Linux 7.3.1 (generic), same initramfs and services.

Metric (median ± stdev) r6 s6
PID 1 start → last of 20 services started 32.31 ± 2.67 ms 32.37 ± 1.19 ms
Kernel start → last service started 234.7 ± 4.33 ms 232.39 ± 4.25 ms
Init + supervision processes 2 27
Supervision RSS (sum) 1778 ± 31.74 KiB 3172 ± 8.04 KiB
Supervision PSS (sum) 1766 ± 31.74 KiB 1223 ± 8.04 KiB
System memory used (MemTotal − MemAvailable) 51176 ± 1831.98 KiB 53270 ± 1144.93 KiB

Reading: activation latency is the same within noise. r6 runs 25 fewer processes and about half the summed RSS. s6 has lower PSS: its many small processes share text pages, while r6-supervise is one larger binary (serde/TOML parsing included). The system-memory difference is within one standard deviation. r6 is not faster than s6; it is as fast with a different process model.

r6 optimisation history (same benchmark)

Change PID 1 → services (median)
Initial (3 boots: 46.3, 43.4, 45.5 ms) ≈ 45 ms
Event-driven rc.init wait (signalfd instead of 10 ms polling) + per-service console messages behind r6.verbose 32.3 ms
Change Poweroff request → power down
Initial (kill(-1, 0) counted kernel threads, so the wait always ran 3 s) ≈ 3.2 s
/proc scan ignoring PF_KTHREAD ≈ 0.2 s

Installed Raze Desktop (QEMU/KVM, 4 vCPU, 4 GiB, OVMF, virgl)

Metric Value
Boot to graphical.target (systemd-analyze) 4.70 s = 0.48 s kernel + 2.15 s initrd + 2.06 s userspace
Memory used after login (niri, waybar, mako, foot open; free -m) 681 MiB
Live ISO boot to graphical.target, zstd squashfs (pre-release build) 6.64 s = 0.50 s kernel + 3.17 s initrd + 2.98 s userspace
Live ISO boot to graphical.target, 1.0.0 ISO (xz), no 3D 12.47 s = 0.50 s kernel + 4.60 s initrd + 7.38 s userspace
Memory used, installed desktop after reboot (niri, waybar, foot) 525 MiB
Full graphical installation (Calamares, erase disk, default desktop) 448 s, host concurrently compiling a kernel
Software-rendered session idle (no 3D; waybar reading) 0.4 GiB

Kernel profiles — Raze Core, r6 as PID 1, 10 boots each (median ± stdev)

QEMU/KVM, 2 vCPU, 256 MiB. "Kernel start → PID 1" is CLOCK_BOOTTIME at r6-init entry.

Metric Generic 7.3.1 Minimal v1 Minimal v2 (1.0.0)
Kernel start → PID 1 201.4 ± 3.2 ms 360.9 ± 4.7 ms 169.6 ± 3.5 ms
Kernel start → all services started 235.9 ± 4.9 ms 377.7 ± 5.5 ms 188.5 ± 4.3 ms
PID 1 → all services started 33.2 ± 3.5 ms 16.9 ± 1.6 ms 18.8 ± 1.6 ms
System memory used 50.2 ± 3.7 MB 35.8 ± 1.2 MB 34.9 ± 1.9 MB
bzImage 14.6 MB 15.8 MB 15.6 MB

v1 (defconfig + kvm_guest + virtio built in) used 30 % less memory but booted 160 ms slower: built-in drivers probe synchronously. Boot-log gap analysis found the RAID6 algorithm benchmark (pulled in by built-in btrfs, ~160 ms), libata probing the QEMU ATA port (~90 ms), netconsole (~48 ms) and the DEBUG_WX page-table scan (~45 ms). v2 disables the benchmark, makes ATA modular, drops netconsole and DEBUG_WX (a boot-time warning, not a protection): 16 % faster kernel boot, 20 % faster to services, 30 % less memory than generic. PID 1 → services is faster on both minimal kernels because no modules need loading in rc.init.

The minimal profile targets QEMU/KVM guests only and has been boot-tested with Raze Core, not as an installed desktop kernel.

Rix search — pinned nixpkgs 1d060086 (113 419 packages)

Host: Ryzen 5 7235HS, idle. hyperfine -N --warmup 3 --runs 20.

Operation Time Peak memory
nix search --no-eval-cache nixpkgs firefox (cold) 26.9 s 4.5 GB
rix update — one-time index build (same evaluation) 27.9 s 4.6 GB
nix search nixpkgs firefox (warm eval cache) 1734 ± 18 ms 44.6 MiB
rix search firefox (index) 63.7 ± 1.2 ms (27× faster) 41.4 MiB

The index is an 11 MB text file per nixpkgs revision. Rix pays the evaluation cost once per revision instead of on every query; results are identical because they come from the same nix search evaluation.

ISO size

Build Size Change
First build (zstd-15, reference closure, NixOS defaults) 3.08 GB —
No speech-dispatcher voices, no default font set, gvfs without samba 2.40 GB −0.68 GB
C GPU probe instead of mesa-demos, wiremix instead of pavucontrol, xz 2.18 GB −0.22 GB
xz + x86 BCJ, no separate reference closure, locale list instead of the all-locales archive, no gparted 2.11 GB (1.961 GiB) −0.07 GB

The remaining size is dominated by linux-firmware (825 MiB, already compressed), Mesa/LLVM (needed for GPU drivers and software rendering) and the Qt stack used by Calamares. Trade-off: xz makes the live session boot slower than zstd did; it does not affect installed systems.