Benchmarking SciPy with airspeed velocity¶
This document introduces benchmarking, including reviewing SciPy benchmark test results online, writing a benchmark test, and running it locally. For a video run-through of writing a test and running it locally, see Benchmarking SciPy.
As written in the airspeed velocity (asv) documentation:
Airspeed velocity (asv) is a tool for benchmarking Python packages over their lifetime. Runtime, memory consumption, and even custom-computed values may be tracked. The results are displayed in an interactive web frontend that requires only a basic static webserver to host.
To see what this means, take a look at airspeed velocity of an unladen scipy. Each plot summarizes the execution time of a particular test over the commit history of the project; that is, as each commit is merged, the benchmark test is run, its execution time is measured, and the elapsed time is plotted. In addition to tracking the performance of the code, a commit is intended to affect, running all benchmarks for each commit is helpful for identifying unintentional regressions: significant increases in the execution time of one or more benchmark tests. As SciPy is a web of interconnected code, the repercussions of a small change may not be immediately obvious to a contributor, so this benchmark suite makes it easier to detect regressions and identify the commit that caused them. When you contribute a substantial new feature - or notice a feature that doesn’t already have a benchmark test - please consider writing benchmarks.
To see how benchmarks are written, take a look at
scipy/benchmarks/benchmarks/optimize_linprog.py. Each subclass of
Benchmark defines a benchmark test. For example, the
class defines a benchmark test based on the Klee-Minty hypercube
problem, a diabolical test of the simplex algorithm for linear
programming. The class has four parts:
setupprepares the benchmark to run. The execution time of this function is not counted in the benchmark results, so this is a good place to set up all variables that define the problem. In the
KleeMintyexample, this involves generating arrays
b_ubcorresponding with a Klee-Minty hypercube in
dimsdimensions and storing them as instance variables.
time_klee_mintyactually runs the benchmark test. This function executes after a
KleeMintyobject has been instantiated and
setuphas run, so it gets the arrays defining the problem from
self. Note that the prefix
timein the function name indicates to
asvthat the execution time of this function is to be counted in the benchmark results.
paramsis a list of lists defining parameters of the test. Benchmarks are run for all possible combinations of these parameters. For example, the first time the benchmark is run, the first element of
simplex) is passed into
time_klee_mintyas the first argument,
meth, and the first element of
[3, 6, 9](
3) is passed into
time_klee_mintyas the second argument,
dims. The next time the benchmark is run,
6as arguments, and so this continues until all combinations of parameters have been used.
param_namesis a list of human-readable names for each element of the
paramslist. These are used for presenting results.
Results of this benchmark over the past few years are available by clicking on the KleeMinty.time_klee_minty link at airspeed velocity of an unladen scipy. Note that each trace of the plot corresponds with a combination of benchmark parameters and environment settings (e.g., the Cython version), and that the visibility of the traces can be toggled using the control panel on the left.
Running benchmarks locally¶
Before beginning, ensure that airspeed velocity is installed.
After contributing new benchmarks, you should test them locally before submitting a pull request.
To run all benchmarks, navigate to the root SciPy directory at the command line and execute:
python runtests.py --bench
--bench activates the benchmark suite instead of the test
suite. This builds SciPy and runs the benchmarks. (Note: this could
take a while. Benchmarks often take longer to run than unit tests, and
each benchmark is run multiple times to measure the distribution in
To run benchmarks from a particular benchmark module, such as
optimize_linprog.py, simply append the filename without the
python runtests.py --bench optimize_linprog
To run a benchmark defined in a class, such as
python runtests.py --bench optimize_linprog.KleeMinty
To compare benchmark results between the active branch and another, such
python runtests.py --bench-compare master optimize_linprog.KleeMinty
All of the commands above display the results in plain text in the
console, and the results are not saved for comparison with future
commits. For greater control, a graphical view, and to have results
saved for future comparison, you can use use the
asv terminal command
To use it, navigate to
scipy/benchmarks in the console and then
This command runs the whole benchmark suite and saves the results for comparison against future commits.
To run only a single benchmark, such as
asv run --bench optimize_linprog.KleeMinty
One great feature of
asv is that it can automatically run a
benchmark not just for the current commit, but for every commit in a
method='interior-point' was merged into SciPy
7fa17f2369e0e5ad055b23cc1a5ee079f9e8ca32, so let’s
KleeMinty benchmark for 10 commits between then and now to
track its performance over time:
asv run --bench optimize_linprog.KleeMinty --steps 10 7fa17f..
This will take a while, because SciPy has to be rebuilt for each commit! For more information about specifying ranges of commits, see the git revisions documentation.
To “publish” the results (prepare them to be viewed) and “preview” them in an interactive console:
asv publish asv preview
ASV will report that it is running a server. Using any browser, you can review the results by navigating to http://127.0.0.1:8080 (local machine, port 8080).
For much more information about the
see the airspeed velocity Commands documentation. (Tip:
check out the
asv find command and the
--profile options for