1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526
#![allow(unused)]
//! A statistics-driven micro-benchmarking library written in Rust.
//!
//! This crate is a microbenchmarking library which aims to provide strong
//! statistical confidence in detecting and estimating the size of performance
//! improvements and regressions, while also being easy to use.
//!
//! See
//! [the user guide](https://bheisler.github.io/criterion.rs/book/index.html)
//! for examples as well as details on the measurement and analysis process,
//! and the output.
//!
//! ## Features:
//! * Collects detailed statistics, providing strong confidence that changes
//! to performance are real, not measurement noise.
//! * Produces detailed charts, providing thorough understanding of your code's
//! performance behavior.
//!
//! ## Feature flags
#![cfg_attr(feature = "document-features", doc = document_features::document_features!())]
#![cfg_attr(docsrs, feature(doc_auto_cfg))]
//!
#![allow(clippy::style, clippy::complexity)]
#![warn(bare_trait_objects)]
#![cfg_attr(feature = "codspeed", allow(unused))]
#[cfg(all(feature = "rayon", target_arch = "wasm32"))]
compile_error!("Rayon cannot be used when targeting wasi32. Try disabling default features.");
use serde::{Deserialize, Serialize};
// Needs to be declared before other modules
// in order to be usable there.
#[macro_use]
mod macros_private;
#[macro_use]
mod analysis;
mod benchmark;
#[macro_use]
mod benchmark_group;
#[cfg(feature = "codspeed")]
#[macro_use]
pub mod codspeed;
pub mod async_executor;
mod bencher;
mod cli;
mod connection;
mod criterion;
mod error;
mod estimate;
mod format;
mod fs;
mod kde;
pub mod measurement;
pub mod profiler;
mod report;
mod routine;
mod stats;
#[cfg(not(feature = "codspeed"))]
#[macro_use]
mod macros;
#[cfg(feature = "codspeed")]
#[macro_use]
mod macros_codspeed;
use std::{
default::Default,
env,
net::TcpStream,
path::PathBuf,
process::Command,
sync::{Mutex, OnceLock},
time::Duration,
};
#[cfg(feature = "async")]
#[cfg(not(feature = "codspeed"))]
pub use crate::bencher::AsyncBencher;
#[cfg(not(feature = "codspeed"))]
pub use crate::bencher::Bencher;
#[cfg(not(feature = "codspeed"))]
pub use crate::benchmark_group::{BenchmarkGroup, BenchmarkId};
#[cfg(feature = "async")]
#[cfg(feature = "codspeed")]
pub use crate::codspeed::bencher::AsyncBencher;
#[cfg(feature = "codspeed")]
pub use crate::codspeed::bencher::Bencher;
#[cfg(feature = "codspeed")]
pub use crate::codspeed::benchmark_group::{BenchmarkGroup, BenchmarkId};
#[cfg(feature = "codspeed")]
pub use crate::codspeed::criterion::Criterion;
#[cfg(not(feature = "codspeed"))]
pub use crate::criterion::Criterion;
use crate::{
benchmark::BenchmarkConfig,
connection::{Connection, OutgoingMessage},
measurement::{Measurement, WallTime},
profiler::{ExternalProfiler, Profiler},
report::{BencherReport, CliReport, CliVerbosity, Report, ReportContext, Reports},
};
fn cargo_criterion_connection() -> &'static Option<Mutex<Connection>> {
static CARGO_CRITERION_CONNECTION: OnceLock<Option<Mutex<Connection>>> = OnceLock::new();
CARGO_CRITERION_CONNECTION.get_or_init(|| match std::env::var("CARGO_CRITERION_PORT") {
Ok(port_str) => {
let port: u16 = port_str.parse().ok()?;
let stream = TcpStream::connect(("localhost", port)).ok()?;
Some(Mutex::new(Connection::new(stream).ok()?))
}
Err(_) => None,
})
}
fn default_output_directory() -> &'static PathBuf {
static DEFAULT_OUTPUT_DIRECTORY: OnceLock<PathBuf> = OnceLock::new();
DEFAULT_OUTPUT_DIRECTORY.get_or_init(|| {
// Set criterion home to (in descending order of preference):
// - $CRITERION_HOME (cargo-criterion sets this, but other users could as well)
// - $CARGO_TARGET_DIR/criterion
// - the cargo target dir from `cargo metadata`
// - ./target/criterion
if let Some(value) = env::var_os("CRITERION_HOME") {
PathBuf::from(value)
} else if let Some(path) = cargo_target_directory() {
path.join("criterion")
} else {
PathBuf::from("target/criterion")
}
})
}
fn debug_enabled() -> bool {
static DEBUG_ENABLED: OnceLock<bool> = OnceLock::new();
*DEBUG_ENABLED.get_or_init(|| std::env::var_os("CRITERION_DEBUG").is_some())
}
/// Reexport of [std::hint::black_box].
#[inline]
pub fn black_box<T>(dummy: T) -> T {
std::hint::black_box(dummy)
}
/// Argument to [`Bencher::iter_batched`] and [`Bencher::iter_batched_ref`] which controls the
/// batch size.
///
/// Generally speaking, almost all benchmarks should use `SmallInput`. If the input or the result
/// of the benchmark routine is large enough that `SmallInput` causes out-of-memory errors,
/// `LargeInput` can be used to reduce memory usage at the cost of increasing the measurement
/// overhead. If the input or the result is extremely large (or if it holds some
/// limited external resource like a file handle), `PerIteration` will set the number of iterations
/// per batch to exactly one. `PerIteration` can increase the measurement overhead substantially
/// and should be avoided wherever possible.
///
/// Each value lists an estimate of the measurement overhead. This is intended as a rough guide
/// to assist in choosing an option, it should not be relied upon. In particular, it is not valid
/// to subtract the listed overhead from the measurement and assume that the result represents the
/// true runtime of a function. The actual measurement overhead for your specific benchmark depends
/// on the details of the function you're benchmarking and the hardware and operating
/// system running the benchmark.
///
/// With that said, if the runtime of your function is small relative to the measurement overhead
/// it will be difficult to take accurate measurements. In this situation, the best option is to use
/// [`Bencher::iter`] which has next-to-zero measurement overhead.
#[derive(Debug, Eq, PartialEq, Copy, Hash, Clone)]
pub enum BatchSize {
/// `SmallInput` indicates that the input to the benchmark routine (the value returned from
/// the setup routine) is small enough that millions of values can be safely held in memory.
/// Always prefer `SmallInput` unless the benchmark is using too much memory.
///
/// In testing, the maximum measurement overhead from benchmarking with `SmallInput` is on the
/// order of 500 picoseconds. This is presented as a rough guide; your results may vary.
SmallInput,
/// `LargeInput` indicates that the input to the benchmark routine or the value returned from
/// that routine is large. This will reduce the memory usage but increase the measurement
/// overhead.
///
/// In testing, the maximum measurement overhead from benchmarking with `LargeInput` is on the
/// order of 750 picoseconds. This is presented as a rough guide; your results may vary.
LargeInput,
/// `PerIteration` indicates that the input to the benchmark routine or the value returned from
/// that routine is extremely large or holds some limited resource, such that holding many values
/// in memory at once is infeasible. This provides the worst measurement overhead, but the
/// lowest memory usage.
///
/// In testing, the maximum measurement overhead from benchmarking with `PerIteration` is on the
/// order of 350 nanoseconds or 350,000 picoseconds. This is presented as a rough guide; your
/// results may vary.
PerIteration,
/// `NumBatches` will attempt to divide the iterations up into a given number of batches.
/// A larger number of batches (and thus smaller batches) will reduce memory usage but increase
/// measurement overhead. This allows the user to choose their own tradeoff between memory usage
/// and measurement overhead, but care must be taken in tuning the number of batches. Most
/// benchmarks should use `SmallInput` or `LargeInput` instead.
NumBatches(u64),
/// `NumIterations` fixes the batch size to a constant number, specified by the user. This
/// allows the user to choose their own tradeoff between overhead and memory usage, but care must
/// be taken in tuning the batch size. In general, the measurement overhead of `NumIterations`
/// will be larger than that of `NumBatches`. Most benchmarks should use `SmallInput` or
/// `LargeInput` instead.
NumIterations(u64),
#[doc(hidden)]
__NonExhaustive,
}
impl BatchSize {
/// Convert to a number of iterations per batch.
///
/// We try to do a constant number of batches regardless of the number of iterations in this
/// sample. If the measurement overhead is roughly constant regardless of the number of
/// iterations the analysis of the results later will have an easier time separating the
/// measurement overhead from the benchmark time.
fn iters_per_batch(self, iters: u64) -> u64 {
match self {
BatchSize::SmallInput => (iters + 10 - 1) / 10,
BatchSize::LargeInput => (iters + 1000 - 1) / 1000,
BatchSize::PerIteration => 1,
BatchSize::NumBatches(batches) => (iters + batches - 1) / batches,
BatchSize::NumIterations(size) => size,
BatchSize::__NonExhaustive => panic!("__NonExhaustive is not a valid BatchSize."),
}
}
}
/// Baseline describes how the `baseline_directory` is handled.
#[derive(Debug, Clone, Copy)]
pub enum Baseline {
/// `CompareLenient` compares against a previous saved version of the baseline.
/// If a previous baseline does not exist, the benchmark is run as normal but no comparison occurs.
CompareLenient,
/// `CompareStrict` compares against a previous saved version of the baseline.
/// If a previous baseline does not exist, a panic occurs.
CompareStrict,
/// `Save` writes the benchmark results to the baseline directory,
/// overwriting any results that were previously there.
Save,
/// `Discard` benchmark results.
Discard,
}
#[derive(Debug, Clone)]
/// Enum representing the execution mode.
pub(crate) enum Mode {
/// Run benchmarks normally.
Benchmark,
/// List all benchmarks but do not run them.
List(ListFormat),
/// Run benchmarks once to verify that they work, but otherwise do not measure them.
Test,
/// Iterate benchmarks for a given length of time but do not analyze or report on them.
Profile(Duration),
}
impl Mode {
pub fn is_benchmark(&self) -> bool {
matches!(self, Mode::Benchmark)
}
pub fn is_terse(&self) -> bool {
matches!(self, Mode::List(ListFormat::Terse))
}
}
#[derive(Debug, Clone, Copy)]
/// Enum representing the list format.
pub(crate) enum ListFormat {
/// The regular, default format.
Pretty,
/// The terse format, where nothing other than the name of the test and ": benchmark" at the end
/// is printed out.
Terse,
}
impl Default for ListFormat {
fn default() -> Self {
Self::Pretty
}
}
/// Benchmark filtering support.
#[derive(Clone, Debug)]
pub enum BenchmarkFilter {
/// Run all benchmarks.
AcceptAll,
/// Run the benchmark matching this string exactly.
Exact(String),
/// Do not run any benchmarks.
RejectAll,
}
/// Returns the Cargo target directory, possibly calling `cargo metadata` to
/// figure it out.
fn cargo_target_directory() -> Option<PathBuf> {
#[derive(Deserialize)]
struct Metadata {
target_directory: PathBuf,
}
env::var_os("CARGO_TARGET_DIR").map(PathBuf::from).or_else(|| {
let output = Command::new(env::var_os("CARGO")?)
.args(["metadata", "--format-version", "1"])
.output()
.ok()?;
let metadata: Metadata = serde_json::from_slice(&output.stdout).ok()?;
Some(metadata.target_directory)
})
}
/// Enum representing different ways of measuring the throughput of benchmarked code.
/// If the throughput setting is configured for a benchmark then the estimated throughput will
/// be reported as well as the time per iteration.
// TODO: Remove serialize/deserialize from the public API.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
pub enum Throughput {
/// Measure throughput in terms of bytes/second. The value should be the number of bytes
/// processed by one iteration of the benchmarked code. Typically, this would be the length of
/// an input string or `&[u8]`.
Bytes(u64),
/// Equivalent to Bytes, but the value will be reported in terms of
/// kilobytes (1000 bytes) per second instead of kibibytes (1024 bytes) per
/// second, megabytes instead of mibibytes, and gigabytes instead of gibibytes.
BytesDecimal(u64),
/// Measure throughput in terms of elements/second. The value should be the number of elements
/// processed by one iteration of the benchmarked code. Typically, this would be the size of a
/// collection, but could also be the number of lines of input text or the number of values to
/// parse.
Elements(u64),
}
/// Axis scaling type
#[derive(Debug, Clone, Copy)]
pub enum AxisScale {
/// Axes scale linearly
Linear,
/// Axes scale logarithmically
Logarithmic,
}
/// This enum allows the user to control how Criterion.rs chooses the iteration count when sampling.
/// The default is Auto, which will choose a method automatically based on the iteration time during
/// the warm-up phase.
#[derive(Debug, Clone, Copy)]
pub enum SamplingMode {
/// Criterion.rs should choose a sampling method automatically. This is the default, and is
/// recommended for most users and most benchmarks.
Auto,
/// Scale the iteration count in each sample linearly. This is suitable for most benchmarks,
/// but it tends to require many iterations which can make it very slow for very long benchmarks.
Linear,
/// Keep the iteration count the same for all samples. This is not recommended, as it affects
/// the statistics that Criterion.rs can compute. However, it requires fewer iterations than
/// the Linear method and therefore is more suitable for very long-running benchmarks where
/// benchmark execution time is more of a problem and statistical precision is less important.
Flat,
}
impl SamplingMode {
pub(crate) fn choose_sampling_mode(
&self,
warmup_mean_execution_time: f64,
sample_count: u64,
target_time: f64,
) -> ActualSamplingMode {
match self {
SamplingMode::Linear => ActualSamplingMode::Linear,
SamplingMode::Flat => ActualSamplingMode::Flat,
SamplingMode::Auto => {
// Estimate execution time with linear sampling
let total_runs = sample_count * (sample_count + 1) / 2;
let d =
(target_time / warmup_mean_execution_time / total_runs as f64).ceil() as u64;
let expected_ns = total_runs as f64 * d as f64 * warmup_mean_execution_time;
if expected_ns > (2.0 * target_time) {
ActualSamplingMode::Flat
} else {
ActualSamplingMode::Linear
}
}
}
}
}
/// Enum to represent the sampling mode without Auto.
#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
pub(crate) enum ActualSamplingMode {
Linear,
Flat,
}
impl ActualSamplingMode {
pub(crate) fn iteration_counts(
&self,
warmup_mean_execution_time: f64,
sample_count: u64,
target_time: &Duration,
) -> Vec<u64> {
match self {
ActualSamplingMode::Linear => {
let n = sample_count;
let met = warmup_mean_execution_time;
let m_ns = target_time.as_nanos();
// Solve: [d + 2*d + 3*d + ... + n*d] * met = m_ns
let total_runs = n * (n + 1) / 2;
let d = ((m_ns as f64 / met / total_runs as f64).ceil() as u64).max(1);
let expected_ns = total_runs as f64 * d as f64 * met;
if d == 1 {
let recommended_sample_size =
ActualSamplingMode::recommend_linear_sample_size(m_ns as f64, met);
let actual_time = Duration::from_nanos(expected_ns as u64);
eprint!(
"\nWarning: Unable to complete {} samples in {:.1?}. You may wish to increase target time to {:.1?}",
n, target_time, actual_time
);
if recommended_sample_size != n {
eprintln!(
", enable flat sampling, or reduce sample count to {}.",
recommended_sample_size
);
} else {
eprintln!(" or enable flat sampling.");
}
}
(1..(n + 1)).map(|a| a * d).collect::<Vec<u64>>()
}
ActualSamplingMode::Flat => {
let n = sample_count;
let met = warmup_mean_execution_time;
let m_ns = target_time.as_nanos() as f64;
let time_per_sample = m_ns / (n as f64);
// This is pretty simplistic; we could do something smarter to fit into the allotted time.
let iterations_per_sample = ((time_per_sample / met).ceil() as u64).max(1);
let expected_ns = met * (iterations_per_sample * n) as f64;
if iterations_per_sample == 1 {
let recommended_sample_size =
ActualSamplingMode::recommend_flat_sample_size(m_ns, met);
let actual_time = Duration::from_nanos(expected_ns as u64);
eprint!(
"\nWarning: Unable to complete {} samples in {:.1?}. You may wish to increase target time to {:.1?}",
n, target_time, actual_time
);
if recommended_sample_size != n {
eprintln!(", or reduce sample count to {}.", recommended_sample_size);
} else {
eprintln!(".");
}
}
vec![iterations_per_sample; n as usize]
}
}
}
fn is_linear(&self) -> bool {
matches!(self, ActualSamplingMode::Linear)
}
fn recommend_linear_sample_size(target_time: f64, met: f64) -> u64 {
// Some math shows that n(n+1)/2 * d * met = target_time. d = 1, so it can be ignored.
// This leaves n(n+1) = (2*target_time)/met, or n^2 + n - (2*target_time)/met = 0
// Which can be solved with the quadratic formula. Since A and B are constant 1,
// this simplifies to sample_size = (-1 +- sqrt(1 - 4C))/2, where C = (2*target_time)/met.
// We don't care about the negative solution. Experimentation shows that this actually tends to
// result in twice the desired execution time (probably because of the ceil used to calculate
// d) so instead I use c = target_time/met.
let c = target_time / met;
let sample_size = (-1.0 + (4.0 * c).sqrt()) / 2.0;
let sample_size = sample_size as u64;
// Round down to the nearest 10 to give a margin and avoid excessive precision
let sample_size = (sample_size / 10) * 10;
// Clamp it to be at least 10, since criterion.rs doesn't allow sample sizes smaller than 10.
if sample_size < 10 {
10
} else {
sample_size
}
}
fn recommend_flat_sample_size(target_time: f64, met: f64) -> u64 {
let sample_size = (target_time / met) as u64;
// Round down to the nearest 10 to give a margin and avoid excessive precision
let sample_size = (sample_size / 10) * 10;
// Clamp it to be at least 10, since criterion.rs doesn't allow sample sizes smaller than 10.
if sample_size < 10 {
10
} else {
sample_size
}
}
}
#[derive(Debug, Serialize, Deserialize)]
pub(crate) struct SavedSample {
sampling_mode: ActualSamplingMode,
iters: Vec<f64>,
times: Vec<f64>,
}
/// Custom-test-framework runner. Should not be called directly.
#[doc(hidden)]
pub fn runner(benches: &[&dyn Fn()]) {
for bench in benches {
bench();
}
crate::criterion::Criterion::default().configure_from_args().final_summary();
}