Wednesday 3 June 2015

Latest version of the JIT-enhanced Python engine boosts the speed of Python programs by seven times or more

PyPy, the JIT-powered implementation of the Python language built for speed, continues to press the pedal to the metal and remains the easiest way to make Python fast and convenient.

Version 2.6, released yesterday, runs Python applications an average of seven times faster than CPython, the default implementation provided by the Python Software Foundation. In some cases the speedup is even greater; for the popular Django Web framework, the latest set of PyPy benchmarks clocks in at 20 times faster than CPython.




Both fast and easy 



Aside from speed, PyPy's big appeal is its ease of use. Unlike Cython or Numba, which require rewrites of Python apps, PyPy is a drop-in replacement for CPython. Admins can accelerate existing software as is, and developers can continue to develop in Python as they have always done.
PyPy has grown faster over time thanks to constant work on its JITting compiler, with most of the performance improvements in the latest build attributed to "internal refactoring and cleanups." Other changes in PyPy include making the interpreter more useful to the culture of tooling that already exists for Python -- like preliminary support forvmprof, a profiler that reveals bottlenecks in Python code and is specifically designed to work with a JIT compiler like PyPy.
Calls to external C libraries, another speed-up technique common in Python, also received a boost in PyPy via an upgrade to the most recent version of Python's C foreign function interface system, which is used by libraries like the popular stats package Numpy. (Numpy itself also received a feature-support boost in PyPy 2.6.)

Other ways to rev up

Nuitka and Pyston, two other recently minted Python-acceleration projects in the works, stand in contrast to PyPy. Nuitka compiles Python to C++; Pyston, a project sponsored by Dropbox, uses JITting in conjunction with LLVM. The former doesn't yet provide much of a speed boost -- around 2.58 times according to the last set of benchmarks -- and the latter is still under heavy development. For the time being, PyPy remains the best production-ready solution.
Where PyPy faces its biggest challenge is not in solving technical problems, but in finding funding to keep development going. Most of the work on the PyPy project is funded by donations, and efforts to push forward with development for the Python 3.x, 64-bit Windows and ARM versions of PyPy are stalled, according to the most recent blog post.

Another donation-funded project, the Software Transactional Memory effort, aims to do away with one of CPython's longest-standing performance bugaboos, the Global Interpreter Lock (GIL). No easy solution exists, and any solution developed for PyPy might not be rolled into CPython as the bar for doing so is quite high: Python creator Guido van Rossum has said the GIL will stay in CPython "until someone other than me goes through the effort of removing it, and showing that its removal doesn't slow down single-threaded Python code."

No comments:

Post a Comment