SPB Git

Repository Public

BENCHMARKS.md 6.1 KB

# SP Term — Benchmarks

Measured on 2026-11-17 on one machine: MacBook Pro, Apple M5 Max (18 cores), 48 GB RAM, macOS 27.0, built-in display 1312×848 pt @2x, on AC power, with other applications (and other automation) running. Single runs unless a spread is given: treat the numbers as orders of magnitude, not as guarantees.

Renderer under test: the engine's default CoreGraphics renderer. The Metal renderer exists behind an "experimental" switch and was not measured; SP Term makes no claim of GPU acceleration.

# Application (signed universal Release build, arm64 slice)

Command: Scripts/bench.sh "dist/SP Term.app"

Measure Result Method
Launch → first window on screen 250–259 ms (5 launches: 256, 251, 255, 259, 250; median 255) from open until the window server lists the window, polled every 10 ms, after one warm-up launch
Memory, 1 pane, idle 57 MB physical footprint (100 MB resident) footprint, ps 4 s after launch
Memory, 8 panes (8 shells), idle 64 MB footprint (107 MB resident) preset grid8, 6 s after launch
CPU, 8 idle panes 1.2 % ps (includes the 1 Hz process sampler)
Bundle size / DMG size 23 MB / 8.3 MB universal binary

# Terminal pipeline (in-process, Release build with testability)

Command: xcodebuild … -configuration Release ENABLE_TESTABILITY=YES ARCHS=arm64 test -only-testing:SPTermTests/PerformanceTests

These run the real PTY, shell and emulation engine inside the app process with the window not on screen: they measure transport + parsing + buffer maintenance and exclude drawing.

Measure Release Debug Method
Large stdout: cat of 51 MB / 400 000 lines 0.51 s — ≈ 100 MB/s (≈ 780 000 lines/s) 13.5 s — 3.8 MB/s time until cat returned and the last line was parsed
Echo latency: key → PTY → echo → parsed into the screen buffer median 0.26 ms, p95 0.36 ms, max 0.70 ms (200 keys) median 0.69 ms, p95 0.94 ms program cat in the pane; busy-wait on the buffer
300 view resizes while a loop prints continuously 0.32 s, kernel TTY size equals the view afterwards 5.2 s setFrameSize × 300, then stty size
Scrollback bound 200 044 lines kept of 400 000 fed (setting: 200 000) same line count of the buffer
Memory with a full 200 000-line scrollback (~125 columns) ≈ 365 MB more (142 → 508 MB resident; includes the 51 MB test file held by the test) same task_info before/after
Backward search for one string in 200 000 lines 2.5 s 6.9 s findPrevious + match summary

# Reading these numbers

  • Typing latency to glass was not measured. The 0.26 ms figure stops at the screen buffer; the time to the next displayed frame (bounded by the display refresh) is not included, and no camera or frame-timing measurement was made.
  • Throughput of ~100 MB/s is without drawing; with a visible window the engine coalesces frames, which was not quantified here.
  • The first throughput attempt reported 0.4 MB/s. That was a flaw of the test (it re-read the whole 200 000-line buffer every 30 ms, blocking the parser); it is mentioned because reading the full buffer of a very long scrollback is genuinely expensive: SP Term only does it on explicit export, on "attach last 60 lines" and once when an SSH session ends.
  • A long scrollback is memory-hungry (roughly 1.8 KB per line at this width). The default is 20 000 lines; the setting goes up to 1 000 000 and the cost grows linearly.
  • Searching a 200 000-line scrollback takes seconds and runs on the main thread: the window is unresponsive meanwhile. With the default 20 000 lines this is proportionally shorter (not measured separately).
  • Debug builds are 10–25 times slower on the parsing paths; always measure Release.
  • Not measured: energy use, Intel performance, many simultaneous heavy streams, sustained multi-hour sessions.

# SP Term Next (2.0.0) against 1.0.0 — 2026-11-19

Command: Scripts/bench-instance.sh <app> (one isolated instance measured by process id; the binary is started directly, so launch times are not comparable with the open-based figures above). Same machine, other applications running, single session. 1.0.0 is the notarized dist build; 2.0.0 is a local universal Release build.

Measure 1.0.0 2.0.0
Launch → first window on screen, 7 launches, two series each 152–190 ms (medians 166 and 157) 159–182 ms (medians 179 and 171)
Memory, 1 idle pane 57 MB footprint (96 MB resident) 58 MB (101 MB)
Memory, 8 idle panes 63 MB (105 MB) 67 MB (111 MB)
CPU, idle, one 2 s top sample 0.2 % (1 pane), 0.6 % (8 panes) 0.9 % (1 pane), 0.1 % (8 panes)

Reading: the new chrome costs about 10 ms at launch and 1 to 4 MB. The idle CPU samples are single readings within noise of each other; they do not show a difference in either direction. Animation smoothness was not measured.

Terminal pipeline on 2.0.0 (in-process, Release with testability, single run, 2026-11-19). The command of the section above now needs CODE_SIGN_INJECT_BASE_ENTITLEMENTS=YES: without it the hardened Release host refuses the test bundle and the runner hangs ("hung before establishing connection"); that setting was turned off for distribution after the 1.0.0 figures were taken.

Measure 1.0.0 (2026-11-17) 2.0.0 (2026-11-19)
cat of 51 MB / 400 000 lines 0.51 s, ≈ 100 MB/s 0.54 s, 94 MB/s
Echo latency to the screen buffer median 0.26 ms, p95 0.36 median 0.16 ms, p95 0.27
300 view resizes under output 0.32 s 0.72 s, kernel TTY size still equals the view
Backward search in 200 000 lines 2.5 s 1.4 s

The two columns are not the same experiment: in 1.0.0 the window controller (and with it the off-screen test window) was released right after opening, so these tests ran on terminal views that were no longer in a window; in 2.0.0 the window lives. The slower resize figure may come from that, from the new card layout, or both; it was not separated. Throughput and echo latency, which do not depend on layout, are unchanged within run-to-run noise.

Files main
Menu

Theme