Skip to content

The PyARV package

The fastest random numbers in Python.

What is the package's mission?

The PyARV package was designed to:

  1. Provide a clean C library implementation of approximate random variables, to either be used directly in C applications, indirectly through wrappers and bindings, or purely as a reference implementation for porting to other languages.
  2. Make most of this performance accessible for Python applications.

Should I be using the package?

You should be using the package if:

  • You have an optimised application spending most of its time generating random numbers.
  • You are a scientist or a researcher investigating applications which make extensive use of random numbers.

You probably should not be using the package if:

  • You are spending a negligible amount of time generating random numbers.
  • You require cryptographically secure random numbers.
  • You don't care about performance and haven't optimised your application already.

Where does the package originate from?

Generating random numbers from desired distributions can be a bottleneck in many applications (e.g. in Monte Carlo simulations). Approximate random variables (the "ARV" in PyARV) were developed and analysed by Prof. Mike Giles and Dr Oliver Sheridan-Methven as a method of alleviating this problem, offering considerable speed improvements. (They further showed that using nested multilevel Monte Carlo can allow full accuracy to be all-the-while maintained). These were implemented in C by Dr Oliver Sheridan-Methven with the assistance of Prof. Mike Giles and the Performance Libraries team of Arm led by Dr Christopher Goodyer.

The code was originally developed as part of Dr Oliver Sheridan-Methven's doctoral thesis. After a few iterations of cleaning the code for various publications and hand-overs, the code has been pruned and improved, and has transitioned from a proof-of-concept, messy, and "rough around the edges" codebase, into a much cleaner, better documented, better tested code base. There is still much room for improvement, but the code has certainly come a long way.

The repository preceeding PyArv was: approximate_random_variables, and this contains various "how-to" and "getting started" guides, alongside implementations specialised for Intel AVX-512, Nvidia, and Arm ThunderX2 VLA SVE hardwares. The core C libraries in PyARV are lifted from this repo.

Why does the package exist?

This package is a selection of cherry-picked implementations which are suitable for a broad range of hardwares. (There are other hardware specific specialisations which are even faster, targetting Arm and Intel, which are not included in this package).

The underlying C implementations are the core of the package, and arguably the most valuable part of the package. However, to promote the use of approximate random numbers more broadly, we created this package to expose the core functionality to the Python community. While this won't of course make Python a HPC language, it will bring considerable HPC capabilities into the reach of Python applications, at least in so far as random number generation is concerned.