Overview

pytest-benchmark

docs
tests
package

A py.test fixture for benchmarking code. It will group the tests into rounds that are calibrated to the chosen timer. See: calibration.

• Free software: BSD license

Installation

pip install pytest-benchmark

Usage

This plugin provides a benchmark fixture. This fixture is a callable object that will benchmark any function passed to it.

Example:

def something(duration=0.000001):
    # Code to be measured
    return time.sleep(duration)

def test_my_stuff(benchmark):
    # benchmark something
    result = benchmark(something)

    # Extra code, to verify that the run completed correctly.
    # Note: this code is not measured.
    assert result is None

You can also pass extra arguments:

def test_my_stuff(benchmark):
    # benchmark something
    result = benchmark(something, 0.02)

If you need to do some wrapping (like special setup), you can use it as a decorator around a wrapper function:

def test_my_stuff(benchmark):
    @benchmark
    def result():
        # Code to be measured
        return something(0.0002)

    # Extra code, to verify that the run completed correctly.
    # Note: this code is not measured.
    assert result is None

py.test command-line options:

--benchmark-min-time=BENCHMARK_MIN_TIME
	Minimum time per round. Default: 25.00us
--benchmark-max-time=BENCHMARK_MAX_TIME
	Maximum time to spend in a benchmark. Default: 1.00s
--benchmark-min-rounds=BENCHMARK_MIN_ROUNDS
	Minimum rounds, even if total time would exceed –max-time. Default: 5
--benchmark-sort=BENCHMARK_SORT
	Column to sort on. Can be one of: ‘min’, ‘max’, ‘mean’ or ‘stddev’. Default: min
--benchmark-timer=BENCHMARK_TIMER
	Timer to use when measuring time. Default: time.perf_counter
--benchmark-warmup
	Runs the benchmarks two times. Discards data from the first run.
--benchmark-warmup-iterations=BENCHMARK_WARMUP_ITERATIONS
	Max number of iterations to run in the warmup phase. Default: 100000
--benchmark-verbose
	Dump diagnostic and progress information.
--benchmark-disable-gc
	Disable GC during benchmarks.
--benchmark-skip
	Skip running any benchmarks.
--benchmark-only
	Only run benchmarks.

Setting per-test options:

@pytest.mark.benchmark(
    group="group-name",
    min_time=0.1,
    max_time=0.5,
    min_rounds=5,
    timer=time.time,
    disable_gc=True,
    warmup=False
)
def test_my_stuff(benchmark):
    @benchmark
    def result():
        # Code to be measured
        return time.sleep(0.000001)

    # Extra code, to verify that the run
    # completed correctly.
    # Note: this code is not measured.
    assert result is None

Glossary

Iteration

A single run of your benchmarked function.

Round

A set of iterations. The size of a round is computed in the calibration phase.

Stats are computed with rounds, not with iterations. The duration for a round is an average of all the iterations in that round.

See: calibration for an explanation of why it’s like this.

Features

Calibration

pytest-benchmark will run your function multiple times between measurements. A round`is that set of runs done between measurements. This is quite similar to the builtin ``timeit` module but it’s more robust.

The problem with measuring single runs appears when you have very fast code. To illustrate:

In other words, a round is a set of runs that are averaged together, those resulting numbers are then used to compute the result tables. The default settings will try to keep the round small enough (so that you get to see variance), but not too small, because then you have the timer calibration issues illustrated above (your test function is faster than or as fast as the resolution of the timer).

Patch utilities

Suppose you want to benchmark an internal function from a class:

class Foo(object):
    def __init__(self, arg=0.01):
        self.arg = arg

    def run(self):
        self.internal(self.arg)

    def internal(self, duration):
        time.sleep(duration)

With the benchmark fixture this is quite hard to test if you don’t control the Foo code or it has very complicated construction.

For this there’s an experimental benchmark_weave fixture that can patch stuff using aspectlib (make sure you pip install apectlib or pip install pytest-benchmark[aspect]):

def test_foo(benchmark_weave):
    with benchmark_weave(Foo.internal, lazy=True):
        f = Foo()
        f.run()

Documentation

https://pytest-benchmark.readthedocs.org/

Obligatory screenshot

Development

To run the all tests run:

tox

Credits

• Timing code and ideas taken from: https://bitbucket.org/haypo/misc/src/tip/python/benchmark.py