Benchmarks

这页记录 vue-tui 当前公开 benchmark 口径。它用于 1.0 release quality bar 和后续回归对比，不用于宣称对 @opentui/solid 或其他 TUI renderer 的横向性能优势。

Scope

当前 benchmark 覆盖的是 vue-tui 自身的可复现回归指标：

• dirty rows、scanned nodes、painted nodes
• frame task coalescing、dropped updates、commit count
• DOM renderer cache hit、span reuse、segment reuse
• TVirtualList wheel burst
• TLogView append、retention、wrap、ANSI、OSC8 links、search、exact visual index
• Markdown 图片热绘制、Kitty placement 移动的 stdout bytes、TVirtualMarkdown 单行滚动脏行数

当前 benchmark 不覆盖：

• 真实 terminal emulator input-to-paint latency
• 长时间 streaming 后的 GC heap 曲线
• 常规滚动、输入和 append 场景的 stdout bytes per frame（Markdown Kitty 移动已有专项数据）
• 与 @opentui/solid 的同机器、同 terminal、同场景横向结果

这些缺口必须在 release notes 中明确说明，不能用现有 baseline 推导出 raw throughput 或竞品性能结论。

Reproduce

pnpm run bench:dom-renderer
pnpm run bench:scroll-mailbox
pnpm run bench:phase2
pnpm run bench:markdown-graphics
pnpm run bench:baseline
pnpm run bench:baseline:timing

bench:baseline 是 release gate，默认只检查行为预算。bench:baseline:timing 会额外检查 timing budget，适合本地 release review 或 nightly，不作为默认 CI blocker。

bench:markdown-graphics 是 Markdown 图形专项 benchmark。它测量 1 MiB 图片的重复热绘制、256 KiB Kitty 图片的 100 次移动，以及 TVirtualMarkdown 单行滚动的脏行数。耗时用于同机 before/after 比较；stdout bytes、图片传输次数和脏行数是可重复的行为指标，当前不属于 release gate。

Environment

Non-release Local Sample

This sample is collected outside the release matrix. It is kept only to show result shape and make the sample values below traceable. Do not use it as the 1.0.0 release benchmark record. When cutting 1.0.0, add a new release benchmark record and keep this local sample only as historical context.

Field	Value
Date	2026-05-25
Git SHA	1b36fb4a3b80b8165aa09443f008604440e854f9
Package	@simon_he/vue-tui@1.0.0-rc.0
Host	Darwin arm64, Darwin Kernel Version 23.5.0
Node	v23.11.0
pnpm	10.33.4
Bun	1.3.13
DOM env	happy-dom synthetic benchmark
Command	BENCH_TIMING=1 pnpm exec tsx scripts/check-bench-baselines.ts
Result	[bench:baseline] passed
Raw	JSON, summary
Note	Stable release notes should use Node 20 or another runtime-matrix Node version, and record the exact release machine, tarball digest, workflow URLs, raw artifacts, and run output.

Release Gate Budgets

These budgets come from scripts/bench-baselines.json.

Scenario	Budget	Release gate
DOM cache-hit plain rows	cacheHits = 100	yes
DOM single styled row reuse	spansReused = 100	yes
DOM changed multi-segment row reuse	segmentReuseRows = 100	yes
DOM mixed row-key prepass	rowKeyPrepassChecks = 100	yes
DOM renderer second flush timing	secondFlushDurationMs <= 20/30/50 depending scenario	manual / nightly
1000 render nodes, dirty 1 row	dirtyRows = 1, scannedNodes = 1, paintedNodes = 1	yes
1000 render nodes, dirty 1 row timing	durationMs <= 10	manual / nightly
TVirtualList 100k rows, wheel burst 100	frames = 1, coalescingRatio >= 100, avgScannedNodes <= 1, avgPaintedNodes <= 1	yes
TVirtualList 100k rows timing	maxFrameMs <= 5	manual / nightly
DOM sync flush 20 dirty rows	dirtyRows = 20	yes
DOM sync flush 20 dirty rows timing	domFlushMs <= 10	manual / nightly
TLogView 1000 lines burst	frames <= 2, getLineCalls <= 30, dirtyRows <= 10, domFlushRows <= 20	yes
TLogView 1000 lines burst timing	maxFrameMs <= 10	manual / nightly
TLogView retention 100k, max 1000	retainedLineCount = 1000, getLineCalls <= 30	yes
TLogView retention timing	maxFrameMs <= 10	manual / nightly
TLogView retained search, wrapped lines	matchCount = 2000, getLineCalls <= 25000	yes
TLogView search wrap long lines timing	maxFrameMs <= 10	manual / nightly
TLogView exact index retention append	measuredLineCount = 1000	yes
TList wheel / TLogView append / TVirtualList wheel burst	frameTaskCount = 1, droppedUpdates >= 999, dirtyRows <= 20, commits = 1	yes

Non-release Local Sample Results

Timing values below are examples from the non-release local run above. Treat them as context, not as portable guarantees across machines.

Scenario	Result
1000 render nodes, dirty 1 row	durationMs = 0.361, dirtyRows = 1, scannedNodes = 1, paintedNodes = 1
TVirtualList 100k rows, wheel burst 100	frames = 1, maxFrameMs = 0.323, coalescingRatio = 100, avgScannedNodes = 1
DOM sync flush 20 dirty rows	dirtyRows = 20, domFlushMs = 0.218
TLogView 1000 lines burst	frames = 2, maxFrameMs = 0.486, getLineCalls = 20, dirtyRows = 10
TLogView long-line burst	frames = 2, maxFrameMs = 1.611, getLineCalls = 20, dirtyRows = 10
TLogView wrapped long-line burst	frames = 2, maxFrameMs = 0.439, getLineCalls = 1, dirtyRows = 10
TLogView retention 100k, max 1000	frames = 1, maxFrameMs = 0.444, retainedLineCount = 1000, getLineCalls = 20
TLogView ANSI retention, max 1000	frames = 1, maxFrameMs = 2.622, retainedLineCount = 1000, getLineCalls = 20
TLogView OSC8 links retained	frames = 1, maxFrameMs = 8.919, retainedLineCount = 1000, getLineCalls = 20
TLogView retained search wrapped lines	matchCount = 2000, getLineCalls = 20000, maxFrameMs = 0.432
TLogView exact index retention append	measuredLineCount = 1000, visualRowCount = 4000, maxFrameMs = 0.513
TLogView Lab smoke	lineCount = 600, visibleLinks = 14, matchCount = 67, maxFrameMs = 16.196

Markdown Graphics Optimization Sample

这是 2026-08-13 在同一台机器上，用 bench:markdown-graphics 分别运行 origin/main (bd678c6d2f905ca67de6b9a8e4d72ab95cd6399d) 和本 PR 工作树得到的非 release 样本。耗时是 5 次样本的中位数；不能跨机器解释为性能保证。

Scenario	origin/main	This PR	Change
1 MiB image hot paint, 100 iterations	732.373 ms	1.500 ms	488.2x faster
256 KiB Kitty image move, 100 moves	810.515 ms	126.773 ms	6.4x faster
Kitty move stdout bytes	13,362,925	465,884	96.5% fewer
Kitty image transmissions	50	1	98.0% fewer
Kitty placement operations	50	99	placement-only after initial transmit
Kitty delete operations	99	0	no clear/retransmit cycle
TVirtualMarkdown dirty rows for one-row scroll	6	1	83.3% fewer

Environment: macOS 14.5 / Darwin 23.5.0 arm64, Node v24.16.0, pnpm 10.34.5. Raw values and comparison metadata are recorded in the JSON artifact and the result summary.

DOM Renderer Cache Sample

Scenario	Result
Cache-hit plain rows	cacheHits = 100, secondFlushDurationMs = 0.599
Changed plain rows	plainTextRows = 100, secondFlushDurationMs = 0.942
Single styled row reuse	spansReused = 100, secondFlushDurationMs = 2.86
Changed multi-segment row reuse	segmentReuseRows = 100, spansReused = 200, secondFlushDurationMs = 8.304
Mixed row-key prepass	rowKeyPrepassChecks = 100, cacheHits = 55, secondFlushDurationMs = 1.443

Mailbox Sample

Component	Scenario	Result
TList	wheel burst	frameTaskCount = 1, droppedUpdates = 999, dirtyRows = 20, commits = 1
TLogView	wheel burst	frameTaskCount = 1, droppedUpdates = 999, dirtyRows = 20, commits = 1
TLogView	append burst	frameTaskCount = 1, droppedUpdates = 999, dirtyRows = 20, commits = 1
TVirtualList	wheel burst	frameTaskCount = 1, droppedUpdates = 999, dirtyRows = 20, commits = 1

User-Perceived Metrics Still Needed

The next benchmark report should add real terminal/browser runs for:

Metric	Target
input-to-paint p95	Small chrome/input updates at 80x24 and 120x40, with p95 and p99 latency
long streaming heap	100k line append with retention, GC-after-run heap, and no sustained growth across repeated run
stdout bytes/frame	ANSI bytes written per frame for scroll, input, and append scenarios

These should be measured before using performance claims in external comparison copy.

@opentui/solid Comparison Status

There is no professional same-scenario comparison yet.

OpenTUI's public docs describe OpenTUI as a native Zig terminal UI core with TypeScript bindings and currently Bun-first installation. The @opentui/solid npm page describes Solid.js support for OpenTUI, and the OpenTUI renderer docs say createCliRenderer() loads the native Zig rendering library.

That means the credible comparison today is product fit, not raw speed:

Dimension	vue-tui	@opentui/solid
Framework	Vue 3 components and Vue reactivity	Solid JSX binding
Runtime target	Node runtime package target plus browser DOM, CLI stdout, and headless	Bun-first OpenTUI terminal app path
Renderer route	JS terminal buffer with DOM renderer, stdout renderer, and headless use	OpenTUI native Zig renderer through TypeScript binding
Strong story	Vue, browser-hosted terminal, Node/Vite install, logs/markdown/agent UI	Solid JSX plus native OpenTUI terminal rendering
Claim to avoid	Do not claim faster raw renderer throughput without same-scenario data	Not applicable

If vue-tui compares against @opentui/solid, use the same machine, same terminal, same viewport, same runtime constraints, and application-level scenarios. The full comparison protocol lives in Compare With OpenTUI Solid.

Scenario	Metrics
100k virtual list scroll	cold start, first paint, 100 scroll events total, p95 input-to-paint, frames, stdout bytes, heap delta, CPU
streaming log / transcript	append throughput, p95/p99 frame time, render calls per visible frame, heap after retention, detached scroll
agent console	markdown update p95, input typing p95, overlay open/close p95, stdout bytes/frame, dropped frames over 60s
browser-hosted terminal	vue-tui-only DOM renderer data: first render, dirty row flush, cache-hit flush, Vite import, SSR import