Performance Benchmarks

Real-world search benchmarks across multiple tools, corpora, and pattern types. ripgrep consistently leads — often by a wide margin.

Benchmarks adapted from Andrew Gallant's original benchmark blog post. Results vary by hardware, pattern, and corpus.

Linux Kernel Source Tree

Pattern: [A-Z]+_SUSPEND — searches the full Linux kernel source (~25 million lines of code, ~900 MB). All tools search recursively with line numbers enabled.

536 matches found by all tools.

Tool	Command	Time (s)	vs ripgrep
ripgrep fastest	rg -n -w "[A-Z]+_SUSPEND"	0.082	—
hypergrep	hgrep -n -w "[A-Z]+_SUSPEND"	0.167	2.0×
git grep	git grep -P -n -w "[A-Z]+_SUSPEND"	0.273	3.3×
ugrep	ugrep -r -n -w "[A-Z]+_SUSPEND"	0.301	3.7×
ag	ag -w "[A-Z]+_SUSPEND"	0.443	5.4×
grep (GNU)	grep -rn -P -w "[A-Z]+_SUSPEND"	0.671	8.2×
ack	ack -w "[A-Z]+_SUSPEND"	3.231	39.4×

Lower is better. Times are medians over multiple runs.

Single Large File (13.5 GB)

Pattern: Sherlock Holmes — a simple literal search on a single 13.5 GB text file (concatenated Project Gutenberg corpus). Tests raw throughput on a single large file.

All tools read the full file without any ignore-file filtering.

Tool	Command	Time (s)	vs ripgrep
ripgrep fastest	rg "Sherlock Holmes"	6.73	—
ugrep	ugrep "Sherlock Holmes"	7.02	—
grep (GNU)	grep "Sherlock Holmes"	9.20	1.4×
ag (memory-maps entire file)	ag "Sherlock Holmes"	34.60	5.1×

Lower is better.

Methodology

Test Environment

CPU: Intel i9-12900K @ 5.2 GHz
RAM: 64 GB DDR5
Storage: NVMe SSD (warm cache)
OS: Linux (Ubuntu 22.04)
Measurement: hyperfine (median of 10 runs, warm cache)

Corpora

Linux kernel: v6.6 source tree, ~75,000 files, ~900 MB
Large file: 13.5 GB concatenated Project Gutenberg texts
Cache: Files pre-read into OS page cache before timing

Run your own benchmarks: Install hyperfine and compare tools on your own codebase. Results vary significantly based on pattern type, corpus size, file system, and hardware. Benchmarks adapted from Andrew Gallant's original benchmark post.

Why is ripgrep so fast?

⚡

SIMD acceleration

Literal patterns use CPU vector instructions to scan 16–32 bytes per cycle.

🧵

Parallel search

All CPU cores search different files simultaneously using a work-stealing thread pool.

🗺️

Memory-mapped I/O

Large files are mapped directly into memory — no read() syscall overhead.

🙈

Smart skipping

Binary files, hidden dirs, and git-ignored files are excluded before they are even opened.

See it yourself

Download ripgrep and run it on your own codebase.

Download ripgrep Compare with grep →