NetworkX uses asv to monitor for performance regressions and evaluate proposals for improved performance. Available benchmarks are available in the benchmarks/benchmarks subdirectory and are loosely organized/named to reflect either the algorithms or data structures that they focus on.

The two most generic sets of benchmarks are:

• benchmarks_classes.py: This covers benchmarks for performance of core NetworkX classes (Graph, DiGraph, MultiGraph, MultiDiGraph). We test performance of basic graph operations (add_node, add_edge, subgraph and others). Code that touches the core classes should run these benchmarks to make sure if there is a performance hit or gain.
• benchmark_algorithms.py: This runs a tiny subset of algorithms which cover different parts of our codebase. We use graphs with different density, and a real world dataset from SNAP datasets to run the algorithms.

There are many other algorithms captured in other files/benchmarking classes within the suite.

Benchmarking runs are controlled by the asv.conf.json configuration file. For a full listing of configuration options, see the asv documentation

Typically, the most important configuration options for a successful ASV configuration are:

1. environment_type which controls the environment management tool used by ASV,
2. pythons, specifying which version(s) of Python will be used during benchmarking, and
3. matrix which dictates what dependencies (and their versions) will be tested against.

environment_type is set to "conda" by default - the "virtualenv" option is recommended for simpler building/setup.

When using `environment_type: "virtualenv"`, it is recommended to limit the
`pythons` field to only those versions of Python that are installed on the
system.

The matrix field dictates which optional dependencies will be installed in the benchmarking environment. To get an accurate picture of NetworkX performance with "default" soft dependencies installed, be sure to include numpy and scipy in this listing. To evaluate pure-Python code branches, remove the dependencies from matrix.

Some benchmarks themselves depend on `numpy`, `scipy` and/or `pandas`; so removing
dependencies may cause some benchmarks to be skipped.

The default configuration builds benchmarking environments for 3 different versions of Python, all with the latest stable version of default dependencies, using conda for managing the build environment and Python versions.

To switch to built-in environment management:

"environment_type": "virtualenv"

To limit the benchmarking run to a single version of Python:

"pythons": ["3.13"]

To run the benchmark on the current HEAD commit:

asv run

Run the benchmark to compare two commits:

asv continuous base_commit_hash test_commit_hash

The --bench flag can be used to limit a run to a subset of benchmarks. For example, the following will only run the algorithm benchmarks for performance comparison between two commits:

asv continuous --bench AlgorithmBenchmarks <sha1> <sha2>

The same pattern can be used to specify individual benchmarks:

asv continuous --bench AlgorithmBenchmarks.time_pagerank <sha1> <sha2>

To know more visit - https://asv.readthedocs.io/en/stable/commands.html