zbench

Compare the performance of one or more commands using Linux hardware performance counters, wall-clock time, and peak memory — in a single run.

zbench repeatedly executes each command, measures it with perf_event_open plus wait4, and prints a colored table of statistics. When you pass more than one command, every command after the first is compared against the first with a two-sample t-test, so you can see whether a difference is actually significant.

Linux-only.

Install

From crates.io:

cargo install zbench

Or build from source:

cargo build --release
# binary at target/release/zbench

The release profile uses thin LTO + panic = "abort" + symbol stripping.

Usage

Usage: zbench [options] <command1> ... <commandN>

Compares the performance of the provided commands.

Options:
 -d, --duration <ms>    (default: 5000) how long to repeatedly sample each command
 -r, --runs <n>         run each command exactly <n> times (overrides --duration)
 -w, --warmup <n>       (default: 0) unmeasured warmup runs before sampling each command
     --max-runs <n>     (default: 10000) cap on samples per command in duration mode
     --color <when>     (default: auto) color output mode
                            available options: 'auto', 'never', 'ansi'
 -f, --allow-failures   (default: false) compare performance even on a non-zero exit code

Commands are split on spaces — there is no shell, so no quoting, globbing, pipes, or variable expansion. This keeps shell startup out of the measured numbers; it also means a command like 'sleep 0.1' is sleep with argument 0.1, not a shell string.

Examples

# Sample one command for ~2s
zbench -d 2000 ./bench_c

# Compare two builds, exact run count, with warmup
zbench -w 5 -r 200 ./old ./new

# Three-way comparison (first command is the reference)
zbench ./bench_c ./bench_rs ./bench_go

Each run reports: wall_time, peak_rss, cpu_cycles, instructions, cache_references, cache_misses, and branch_misses — as mean ± σ, min … max, an outlier count (Tukey's fences), and (for comparisons) a delta vs. the reference. A 💩 marks a significant regression, ⚡ a significant improvement.

Measurement model

• No fixed warmup by default. zbench runs each command at least 3 times, then keeps sampling until --duration elapses (capped by --max-runs). --runs <n> switches to an exact count instead. --warmup <n> adds unmeasured priming runs to warm caches / the page cache / the dynamic linker.
• Hardware counters are user-space only (exclude_kernel), opened on the benchmarked process with inherit + enable_on_exec, so counting starts when the child execves and includes its threads/children.
• Peak RSS comes from wait4's ru_maxrss. zbench forces a fork+exec spawn (not posix_spawn) so the reported RSS is the child's, not the parent's.

Requirements

• Linux with access to perf_event_open. If your host restricts it, check /proc/sys/kernel/perf_event_paranoid (counters here exclude the kernel, so level 2 and below are sufficient for unprivileged use).
• A recent stable Rust toolchain.

Acknowledgements

The design and output format are modeled on Andrew Kelley's poop (MIT).

License

Licensed under the MIT License.