Translation to bytecode happens when a module is first imported, and it is avoided when possible to speed program startup: bytecode is automatically saved in .pyc files and, unless you change the corresponding source file, loaded directly the next time your program runs.

This bytecode compilation model makes Python scripts portable and faster than a pure interpreter that runs raw source code lines. But it also makes Python slower than true compilers that translate source code to binary machine code. Bytecode is not machine code and is ultimately run by the Python (or other) virtual machine program, not directly by your computer's hardware.

Keep in mind, though, that some of these details are specific to the standard Python implementation. For instance, the Jython system compiles Python scripts to Java bytecode, and the IronPython implementation compiles Python source code to the bytecode used by the C#/.NET environment. In addition, Python compiler-related projects have been spawned in the past and will likely continue into the future. For more details on this front, see the following:

• The Psyco just-in-time compiler for Python, which replaces portions of a running program's bytecode with optimized binary machine code tailored to specific datatypes. Psyco can speed Python programs by any factor from 2 to 100. The high end is more likely for heavily algorithmic code, whereas I/O-bound programs don't improve as much. (In my own experience, a 3x-5x speedup is common for typical programsamazing for a simple install.)

• A related project, PyPy, which aims to reimplement the Python virtual machine to better support optimizations. The PyPy project may incorporate and subsume Psyco's techniques.

• The Parrot project, which seeks to develop a bytecode and virtual machine that will be shared by many languages, including Python.

• The Installer, Py2Exe, and Freeze systems, which package Python programs as standalone executables known as "frozen binaries"a combination of your bytecode and the Python virtual machine. Frozen binaries do not require that Python be installed on the receiving end.

• Other program distribution formats, including zip archives (with modules automatically extracted on imports); Python eggs (an emerging package format); Distutils (an installation script system); and encrypted bytecode (for instance, using PyCrypto and the import hooks).

• The emerging Shed Skin system, which translates Python source code to C++. This system assumes that your code will not use the full range of Python's dynamic typing, but this constraint allows highly efficient code to be generated, which is by some accounts faster than Psyco and much faster than standard Python. Shed Skin's own website reports speedups of 12 and 45 times faster on average than Psyco and standard CPython, respectively, though results can vary greatly.

Psyco may provide a simpler optimization path for some programs than linked-in C libraries, especially for algorithm-intensive code. Although Python's extreme dynamic nature makes compilation complex (the behavior of "x + 1" cannot be easily predicted until runtime).