Numerical benchmarks for Julia, Python.

We try to avoid trivial and non-sens benchmarks (for people doing numerics!) like fibonacci, sorting and so on.

We put ourself in the shoes of a typical Matlab or Matlab-like programmer, writing quite short but numerically intensive programs.

Are Python and Julia easy to use and efficient? We compare them with a C++ optimized implementation (and sometimes with a Fortran one).

The benchmark(s):

• Gaussian: Gaussian elimination with partial pivoting.

• FeStiff: compute the stiffness matrix, for the Poisson equation, discretized with P2 finite elements on triangles.

• Weno: a classical solver for hyperbolic equations, in dimension 1, with application to Burghers equation and to Convection.

• 🆕 Sparse: building a sparse matrix and doing a sparse matrix x vector product.

• MicroBenchmarks: very simple benchmarks to show the importance of different programing styles.

We will add other numerical significant benchmarks in the (near) future.

Dependencies:

What you need to install:

• python3
• pip (pip3)
• g++ (and/or clang++)
• gfortran
• lapack
• openblas
• cmake
• gnuplot

You can install them using your distribution tool (apt...).

• julia

❗ Julia ❗ since 2018-10-08 programs need at least version Version 1.1 (stable version in 2018-10); note that all programs needed adaptation when moving to this version, and will not run with former ones.

Note also that the version packaged with Ubuntu 18-04 is older. Install the stable version from here. Note also that Julia is evolving, and it is possible that the codes need some adaptation to run with later versions of the language.

You also need:

• pythran
• scipy
• Numpy
• numba

to install them, you can just do:

pip install pythran

and so on...

You can also install them from conda.

TODO list

• Add a small static website (Sphinx? Pelican?) to present the results (Readthedocs?)

• Improve the quality of the figures (using Pandas + Matplotlib?)

• Automate the execution of the benchmarks and the creation of the website (Gitlab CI?).

• Add something based on the Julia micro benchmarks

• Maybe add Fortran if a Fortran developper is willing to work on this...

• Maybe add Xtensor if a xtensor developper is willing to work on this...