Atlas Bench

atlas.bench is a high-precision, multi-command benchmarking TUI that brings statistical rigor to the terminal on a phosphor-CRT styled console inspired by 1970s engineering workstations.

✨ Features

• 🏎️ Side-by-Side Comparison: Live per-target telemetry with state pills (WARM, RUN, DONE, ERR).
• 🔬 Statistical Rigor: Mean, median, min/max, and standard deviation reported for every target.
• 🔥 Intelligent Warm-up: Unmeasured warm-up runs prime disk caches and OS scheduling before measurement starts.
• 📈 Run-Distribution Sparklines: Each target shows a live sparkline of run durations, with a deeper view on ↵.
• 🏆 Analysis Section: Ratio bar chart vs. the winner (1.00x, 1.23x, 2.40x…), color-graded by how far behind each target is.
• ♻️ Restart in Place: Press r to discard results and re-run the same battery without relaunching.
• 📦 Cross-Platform: Binaries available for Windows, Linux, and macOS (AMD64, ARM64).

🚀 Installation

From Source

git clone https://github.com/fezcode/atlas.bench
cd atlas.bench
go build -o atlas.bench .

⌨️ Usage

Basic Comparison

atlas.bench "node app.js" "bun app.js"

High-Precision Audit

atlas.bench --runs 100 --warmup 10 "python3 process.py" "go run main.go"

Per-Run Timeout

atlas.bench -t 5s "slow-thing" "fast-thing"

Options

Flag	Short	Default	Description
--runs	-r	10	Timed iterations per command
--warmup	-w	3	Unmeasured warm-up runs
--timeout	-t	0	Per-run timeout (e.g. 5s, 100ms; 0 disables)
--shell	-s	auto	Shell to use (auto-detects powershell on Windows, sh elsewhere)

🕹️ Controls

Key	Action
↑/↓ or k/j	Navigate the target list
Enter	Toggle detail (min/max/stddev + per-run bar chart)
r	Restart: wipe results and re-run
q or Ctrl+C	Quit

📐 Statistics

• Mean: Arithmetic average of all measured runs.
• Median: The middle value, resistant to single-spike outliers.
• StdDev: How much the runs varied; a high value indicates jitter from background load.
• Min / Max: Fastest and slowest individual runs.
• Winner: The target with the lowest mean execution time.
• Ratio: Each target's mean divided by the winner's (1.00x = winner).

🏗️ Building for all platforms

The project uses gobake to generate binaries for all supported platforms:

gobake build

Binaries are placed in the build/ directory.

📄 License

MIT License - see LICENSE for details.