First install the Archery utility to run the benchmark suite.

Running the benchmark suite

The benchmark suites can be ran with the benchmark run sub-command.

# Run benchmarks in the current git workspace
archery benchmark run
# Storing the results in a file
archery benchmark run --output=run.json

Sometimes, it is required to pass custom CMake flags, e.g.

export CC=clang-8 CXX=clang++8
archery benchmark run --cmake-extras="-DARROW_SIMD_LEVEL=NONE"


One goal with benchmarking is to detect performance regressions. To this end, archery implements a benchmark comparison facility via the benchmark diff sub-command.

In the default invocation, it will compare the current source (known as the current workspace in git) with local master branch.

For more information, invoke the archery benchmark diff --help command for multiple examples of invocation.

Iterating efficiently

Iterating with benchmark development can be a tedious process due to long build time and long run times. Multiple tricks can be used with archery benchmark diff to reduce this overhead.

First, the benchmark command supports comparing existing build directories, This can be paired with the --preserve flag to avoid rebuilding sources from zero.

# First invocation clone and checkouts in a temporary directory. The
# directory is preserved with --preserve
archery benchmark diff --preserve

# Modify C++ sources

# Re-run benchmark in the previously created build directory.
archery benchmark diff /tmp/arrow-bench*/{WORKSPACE,master}/build

Second, a benchmark run result can be saved in a json file. This also avoids rebuilding the sources, but also executing the (sometimes) heavy benchmarks. This technique can be used as a poor’s man caching.

# Run the benchmarks on a given commit and save the result
archery benchmark run --output=run-head-1.json HEAD~1
# Compare the previous captured result with HEAD
archery benchmark diff HEAD run-head-1.json

Third, the benchmark command supports filtering suites (--suite-filter) and benchmarks (--benchmark-filter), both options supports regular expressions.

# Taking over a previous run, but only filtering for benchmarks matching
# `Kernel` and suite matching `compute-aggregate`.
archery benchmark diff                                       \
  --suite-filter=compute-aggregate --benchmark-filter=Kernel \

Regression detection

Writing a benchmark

  1. The benchmark command will filter (by default) benchmarks with the regular expression ^Regression. This way, not all benchmarks are run by default. Thus, if you want your benchmark to be verified for regression automatically, the name must match.

  2. The benchmark command will run with the --benchmark_repetitions=K options for statistical significance. Thus, a benchmark should not override the repetitions in the (C++) benchmark’s arguments definition.

  3. Due to #2, a benchmark should run sufficiently fast. Often, when the input does not fit in memory (L2/L3), the benchmark will be memory bound instead of CPU bound. In this case, the input can be downsized.

  4. By default, google’s benchmark library will use the cputime metric, which is the sum of runtime dedicated on the CPU for all threads of the process. By contrast to realtime which is the wall clock time, e.g. the difference between end_time - start_time. In a single thread model, the cputime is preferable since it is less affected by context switching. In a multi thread scenario, the cputime will give incorrect result since the since it’ll be inflated by the number of threads and can be far off realtime. Thus, if the benchmark is multi threaded, it might be better to use SetRealtime(), see this example <>.


archery is written as a python library with a command line frontend. The library can be imported to automate some tasks.

Some invocation of the command line interface can be quite verbose due to build output. This can be controlled/avoided with the --quiet option, e.g.

archery --quiet benchmark diff --benchmark-filter=Kernel
{"benchmark": "BenchSumKernel/32768/0", "change": -0.6498, "regression": true, ...
{"benchmark": "BenchSumKernel/32768/1", "change": 0.01553, "regression": false, ...

or the --output=<file> can be used, e.g.

archery benchmark diff --benchmark-filter=Kernel --output=compare.json