SciPy benchmarks

Benchmarking SciPy with Airspeed Velocity.

Usage

Airspeed Velocity manages building and Python virtualenvs by itself, unless told otherwise. Some of the benchmarking features in runtests.py also tell ASV to use the SciPy compiled by runtests.py. To run the benchmarks, you do not need to install a development version of SciPy to your current Python environment.

Run a benchmark against currently checked-out SciPy version (don't record the result):

python runtests.py --bench sparse.Arithmetic

Compare change in benchmark results with another branch:

python runtests.py --bench-compare main sparse.Arithmetic

Run benchmarks against the system-installed SciPy rather than rebuild:

python runtests.py -n --bench sparse.Arithmetic

Run ASV commands directly (note, this will not set env vars for ccache and disabling BLAS/LAPACK multi-threading, as runtests.py does):

cd benchmarks
asv run --skip-existing-commits --steps 10 ALL
asv publish
asv preview

More on how to use asv can be found in ASV documentation Command-line help is available as usual via asv --help and asv run --help.

Writing benchmarks

See ASV documentation for the basics on how to write benchmarks.

Some things to consider:

• When importing things from SciPy on the top of the test files, do it as:

  from .common import safe_import
  
  with safe_import():
      from scipy.sparse.linalg import onenormest
  

  The benchmark files need to be importable also when benchmarking old versions of SciPy. The benchmarks themselves don't need any guarding against missing features --- only the top-level imports.

• Try to keep the runtime of the benchmark reasonable.

• Use ASV's time_ methods for benchmarking times rather than cooking up time measurements via time.clock, even if it requires some juggling when writing the benchmark.

• Preparing arrays etc., should generally be put in the setup method rather than the time_ methods, to avoid counting preparation time together with the time of the benchmarked operation.

• Use run_monitored from common.py if you need to measure memory usage.

• Benchmark versioning: by default asv invalidates old results when there is any code change in the benchmark routine or in setup/setup_cache.

  This can be controlled manually by setting a fixed benchmark version number, using the version attribute. See ASV documentation for details.

  If set manually, the value needs to be changed manually when old results should be invalidated. In case you want to preserve previous benchmark results when the benchmark did not previously have a manual version attribute, the automatically computed default values can be found in results/benchmark.json.

• Benchmark attributes such as params and param_names must be the same regardless of whether some features are available, or e.g. SCIPY_XSLOW=1 is set.

  Instead, benchmarks that should not be run can be skipped by raising NotImplementedError in setup().