A harness and a set of benchmarks for measuring Django's performance over time.
Running the benchmarks
Here's the short version:
mkvirtualenv djangobench pip install -e git://github.com/django/djangobench.git#egg=djangobench git clone git://github.com/django/django.git cd django djangobench --control=1.2 --experiment=master
Okay, so what the heck's going on here?
First, djangobench doesn't test a single Django version in isolation -- that wouldn't be very useful. Instead, it benchmarks an "experiment" Django against a "control", reporting on the difference between the two and measuring for statistical significance.
Because a Git clone can contain all the project development history, you can test against a single repository specifying individual commit IDs, tag (as we've done above) and even possibly branches names with the --control and --experiment options.
Before djangobench 0.10 you had to use --vcs=git to get this behavior. Now it's the default. There is also support for Mercurial (--vcs=hg).
Another way to use djangobench, is to run it against two complete Django source trees, you can specify this mode by using --vcs=none. By default it looks for directories named django-control and django-experiment in the current working directory:
but you can change that by using the --control or --experiment options:
djangobench --vcs=none --control pristine --experiment work
Now, it's impractical to install the Django source code trees under test (this is particularly true in the two-trees scenario): djangobench works its magic by mucking with PYTHONPATH.
However, the benchmarks themselves need access to the djangobench module, so you'll need to install it.
You can specify the benchmarks to run by passing their names on the command line.
This is an example of not-statistically-significant results:
Running 'startup' benchmark ... Min: 0.138701 -> 0.138900: 1.0014x slower Avg: 0.139009 -> 0.139378: 1.0027x slower Not significant Stddev: 0.00044 -> 0.00046: 1.0382x larger
Writing new benchmarks
Benchmarks are very simple: they're a Django app, along with a settings file, and an executable benchmarks.py that gets run by the harness. The benchmark script needs to honor a simple contract:
It's an executable Python script, run as __main__ (e.g. python path/to/benchmark.py). The subshell environment will have PYTHONPATH set up to point to the correct Django; it'll also have DJANGO_SETTINGS_MODULE set to <benchmark_dir>.settings.
The benchmark script needs to accept a --trials argument giving the number of trials to run.
The output should be simple RFC 822-ish text -- a set of headers, followed by data points:
Title: some benchmark Description: whatever the benchmark does 1.002 1.003 ...
The list of headers is TBD.
There's a couple of utility functions in djangobench.utils that assist with honoring this contract; see those functions' docstrings for details.
The existing benchmarks should be pretty easy to read for inspiration. The query_delete benchmark is probably a good place to start.
Please write new benchmarks and send us pull requests on Github!