Python is an extremely powerful language for writing research software. It comes with a huge ecosystem of modules that enable it to be used to write software across almost all fields of research.

One reason for its success is that Python is a scripting language (also called an interpreted language). It is easier and quicker to write programs in a scripting language than using a compiled language, such as C, C++ or Fortran. To understand why, we need to understand the difference between a script and compiled language, and how exactly a program is executed by a computer processor.

Computer processors don’t understand Python, C, Fortran or any other programming language. Computer processors only understand (and can execute) machine code. All programs, regardless of language, have to be converted to machine code so that they can be executed on your computer.

Python scripts are simple text files that are interpreted and (ultimately) converted to machine code at runtime. This is different to compiled languages, such as C, C++ and Fortran, which have to be compiled into executables before they can be run. These executables are files containing the machine code that is then given to the computer processor to be executed.

The Python interpreter works by interpreting the Python script into an intermediate representation, called bytecode. Bytecode is an intermediate representation of the script that can be executed by the Python Virtual Machine.

The Python Virtual Machine executes that bytecode, converting it dynamically into machine code that is understandable by the processor on your computer. This is why Python scripts are portable - while every different architecture of computer processor has its own machine code, the Python Virtual Machine has only a single Python bytecode. As long as there is a Python Virtual Machine available for your computer, then it can execute Python bytecode on your computer, and your Python scripts can run.

The Python Virtual Machine is more flexible than a physical (real) computer processor, and bytecode is more expressive and capable than machine code. This increased expressiveness and flexibility over a real processor executing machine code is why Python scripts are easier to write and understand by humans, and why you need to write fewer lines of code compared to compiled programming languages like C, C++ or Fortran.

However, a downside of Python is that software written purely in Python can be very slow. Executing bytecode on a virtual machine is inherently more expensive than directly executing machine code on a computer processor. Also, bytecode operating on a virtual machine sees a virtual representation of the computer - it can’t directly access memory, but instead has to access memory via the virtual machine.

Programs written in scripting languages are slower and consume more energy than those written in compiled languages. This is a problem for two reasons:

  1. Who wants to wait longer than needed for your calculation to complete?
  2. Energy is precious and should not be wasted. Programs written in Python can consume significantly more electrical energy than equivalent programs written in compiled languages (10s-100s of times more energy, in many cases).

Much of this can be mitigated by making use of optimised Python modules, such as NumPy, Scikit-Learn and TensorFlow. The “slow” parts of these are written such that they are directly compiled to machine code, and so can execute as quickly, and using as little energy, as a program written in a compiled programming language.

However, it is often the case that you need to do something that is not available in an existing, well-optimised module. In these cases, how could we speed up our code, and write it so that it consumes less electricity?

In this workshop we will be covering two different technologies that both solve this problem by directly compiling Python into machine code.

The first technology, numba works by compiling your Python to machine code at run time. The parts of your script that you want to accelerate are compiled automatically just before they are run, and the resulting machine code is then executed by your computer processor.

The second technology is an alternative to numba, and is used manually before your Python script is run. This technology, cython, parses your Python and converts it into the compiled programming language, C. This C code has to be compiled into machine code. This machine code is then automatically loaded when you execute your Python script, and is used instead of going through bytecode and the Python Virtual Machine.

Both of these technologies enable you to write Python that executes as quickly and uses as little power as any compiled programming language. We at BristolRSE have had great success using both technologies, e.g.

  1. Using numba to accelerate minimalmarkers
  2. Using cython to accelerate metawards