Benchmarks

While not mandatory for most pull requests, we ask that performance-related PRs include a benchmark in order to clearly depict the use case that is being optimized for.

In this section we will review how to setup the benchmarks, and three commands spin asv -- dev, spin asv -- run and spin asv -- continuous.

Prerequisites

Begin by installing airspeed velocity in your development environment. Prior to installation, be sure to activate your development environment, then if using venv you may install the requirement with:

source skimage-dev/bin/activate
pip install asv

If you are using conda, then the command:

conda activate skimage-dev
conda install asv

is more appropriate. Once installed, it is useful to run the command:

spin asv -- machine

To let airspeed velocity know more information about your machine.

Writing a benchmark

To write benchmark, add a file in the benchmarks directory which contains a a class with one setup method and at least one method prefixed with time_.

The time_ method should only contain code you wish to benchmark. Therefore it is useful to move everything that prepares the benchmark scenario into the setup method. This function is called before calling a time_ method and its execution time is not factored into the benchmarks.

Take for example the TransformSuite benchmark:

import numpy as np
from skimage import transform

class TransformSuite:
    """Benchmark for transform routines in scikit-image."""

    def setup(self):
        self.image = np.zeros((2000, 2000))
        idx = np.arange(500, 1500)
        self.image[idx[::-1], idx] = 255
        self.image[idx, idx] = 255

    def time_hough_line(self):
        result1, result2, result3 = transform.hough_line(self.image)

Here, the creation of the image is completed in the setup method, and not included in the reported time of the benchmark.

It is also possible to benchmark features such as peak memory usage. To learn more about the features, please refer to the official airspeed velocity documentation.

Also, the benchmark files need to be importable when benchmarking old versions of scikit-image. So if anything from scikit-image is imported at the top level, it should be done as:

try:
    from skimage import metrics
except ImportError:
    pass

The benchmarks themselves don’t need any guarding against missing features, only the top-level imports.

To allow tests of newer functions to be marked as “n/a” (not available) rather than “failed” for older versions, the setup method itself can raise a NotImplemented error. See the following example for the registration module:

try:
    from skimage import registration
except ImportError:
    raise NotImplementedError("registration module not available")

Testing the benchmarks locally

Prior to running the true benchmark, it is often worthwhile to test that the code is free of typos. To do so, you may use the command:

spin asv -- dev -b TransformSuite

Where the TransformSuite above will be run once in your current environment to test that everything is in order.

Running your benchmark

The command above is fast, but doesn’t test the performance of the code adequately. To do that you may want to run the benchmark in your current environment to see the performance of your change as you are developing new features. The command asv run -E existing will specify that you wish to run the benchmark in your existing environment. This will save a significant amount of time since building scikit-image can be a time consuming task:

spin asv -- run -E existing -b TransformSuite

Comparing results to main

Often, the goal of a PR is to compare the results of the modifications in terms speed to a snapshot of the code that is in the main branch of the scikit-image repository. The command asv continuous is of help here:

spin asv -- continuous main -b TransformSuite

This call will build out the environments specified in the asv.conf.json file and compare the performance of the benchmark between your current commit and the code in the main branch.

The output may look something like:

$ spin asv -- continuous main -b TransformSuite
· Creating environments
· Discovering benchmarks
·· Uninstalling from conda-py3.7-cython-numpy1.15-scipy
·· Installing 544c0fe3 <benchmark_docs> into conda-py3.7-cython-numpy1.15-scipy.
· Running 4 total benchmarks (2 commits * 2 environments * 1 benchmarks)
[  0.00%] · For scikit-image commit 37c764cb <benchmark_docs~1> (round 1/2):
[...]
[100.00%] ··· ...ansform.TransformSuite.time_hough_line           33.2±2ms

BENCHMARKS NOT SIGNIFICANTLY CHANGED.

In this case, the differences between HEAD and main are not significant enough for airspeed velocity to report.

It is also possible to get a comparison of results for two specific revisions for which benchmark results have previously been run via the asv compare command:

spin asv -- compare v0.14.5 v0.17.2

Finally, one can also run ASV benchmarks only for a specific commit hash or release tag by appending ^! to the commit or tag name. For example to run the skimage.filter module benchmarks on release v0.17.2:

spin asv -- run -b Filter v0.17.2^!