libgit2/tests/benchmarks

libgit2 benchmarks

This folder contains the individual benchmark tests for libgit2, meant for understanding the performance characteristics of libgit2, comparing your development code to the existing libgit2 code, or comparing libgit2 to the git reference implementation.

Running benchmark tests

Benchmark tests can be run in several different ways: running all benchmarks, running one (or more) suite of benchmarks, or running a single individual benchmark. You can target either an individual version of a CLI, or you can A/B test a baseline CLI against a test CLI.

Specifying the command-line interface to test

By default, the git in your path is benchmarked. Use the -c (or --cli) option to specify the command-line interface to test.

Example: libgit2_bench --cli git2_cli will run the tests against git2_cli.

Running tests to compare two different implementations

You can compare a baseline command-line interface against a test command-line interface using the -b (or --baseline-cli`) option.

Example: libgit2_bench --baseline-cli git --cli git2_cli will run the tests against both git and git2_cli.

Running individual benchmark tests

Similar to how a test suite or individual test is specified in clar, the -s (or --suite) option may be used to specify the suite or individual test to run. Like clar, the suite and test name are separated by ::, and like clar, this is a prefix match.

Examples:

• libgit2_bench -shash_object will run the tests in the hash_object suite.
• libgit2_bench -shash_object::random_1kb will run the hash_object::random_1kb test.
• libgit2_bench -shash_object::random will run all the tests that begin with hash_object::random.

Writing benchmark tests

Benchmark tests are meant to be easy to write. Each individual benchmark is a shell script that allows it to do set up (eg, creating or cloning a repository, creating temporary files, etc), then running benchmarks, then teardown.

The benchmark_helpers.sh script provides many helpful utility functions to allow for cross-platform benchmarking, as well as a wrapper for hyperfine that is suited to testing libgit2. Note that the helper script must be included first, at the beginning of the benchmark test.

Benchmark example

This simplistic example compares the speed of running the git help command in the baseline CLI to the test CLI.

#!/bin/bash -e

# include the benchmark library
. "$(dirname "$0")/benchmark_helpers.sh"

# run the "help" command; this will benchmark `git2_cli help`
gitbench help

Naming

The filename of the benchmark itself is important. A benchmark's filename should be the name of the benchmark suite, followed by two underscores, followed by the name of the benchmark. For example, hash-object__random_1kb is the random_1kb test in the hash-object suite.

Options

The gitbench function accepts several options.

• --sandbox <path> The name of a test resource (in the tests/resources directory). This will be copied as-is to the sandbox location before test execution. This is copied before the prepare script is run. This option may be specified multiple times.

• --repository <path>
  The name of a test resource repository (in the tests/resources directory). This repository will be copied into a sandbox location before test execution, and your test will run in this directory. This is copied before the prepare script is run.

• --prepare <script>
  A script to run before each invocation of the test is run. This can set up data for the test that will not be timed. This script is run in bash on all platforms.

  Several helper functions are available within the context of a prepare script:

  • flush_disk_cache
    Calling this will flush the disk cache before each test run. This should probably be run at the end of the prepare script.
  • create_random_file <path> [<size>] Calling this will populate a file at the given path with size bytes of random data.
  • create_text_file <path> [<size>] Calling this will populate a file at the given path with size bytes of predictable text, with the platform line endings. This is preferred over random data as it's reproducible.

• --warmup <n>
  Specifies that the test should run n times before actually measuring the timing; useful for "warming up" a cache.