PyPy is Fast (And So Can You)


Posted

in


I’ve known for some time that PyPy (Python implemented in a subset of the language called RPython) is fast. The PyPy speed charts show just how fast for a lot of benchmarks (and it’s a little slower in a few areas too).

After seeing a lot of PyPy chatter while PyCon was going on, I thought I’d check it out. On OS X it’s as simple as brew install pypy. After that, just use pypy instead of python.

The first thing I did was throw PyPy at a couple of Project Euler problems. They’re great because they’re computationally expensive and usually have lots of tight loops. For the ones I looked at, PyPy had a 50-75% speed improvement over CPython. David Ripton posted a more complete set of Euler solution runtimes using PyPy, Unladen Swallow, Jython, Psyco, and CPython. Almost all of the time, PyPy is faster, often significantly so. At this point it looks like the PyPy team is treating “slower than CPython” as a bug, or at the very least, something to improve.

The latest stable release currently targets Python 2.5, but if you build the latest version from source it looks like they’re on their way to supporting Python 2.7:

$ ./pypy-c 
Python 2.7.0 (61fefec7abc6, Mar 18 2011, 06:59:57)
[PyPy 1.5.0-alpha0] on darwin
Type "help", "copyright", "credits" or "license" for more information.
And now for something completely different: ``1.1 final released:
http://codespeak.net/pypy/dist/pypy/doc/release-1.1.0.html''
>>>> 

There are a few things to look out for when using PyPy. The entire standard library isn’t built out, though the most commonly used modules are. PyPy supports ctypes and has experimental but incomplete support for the Python C API. PyPy is built out enough to support several large non-trivial projects such as Twisted (without SSL) and Django (with sqlite).

PyPy is definitely one of many bright futures for Python, and it’s fast now. If you’ve been thinking about checking it out, perhaps now is the time to take it for a spin.