These are some notes taken during the keynote talk of Prof. Hans Petter Langtangen at the EuroScipy 2010 conference in Paris.
Teaching science and engineering
Case study : University of Oslo, Norway
Mathematics and engineering sciences should be taught with a very hands-on approach. Students need to launch simulations, tinker with parameters, vizualize outcomes from day one. This is the future of science education.
Undergrads mathematics and mathematical modeling courses were recently redesigned at University of Oslo. The current, revamped cursus includes calculus, numerical calculus, scientific programming using python, vector calculus and linear algebra, mechanics with simulation.
There is a very strong focus on experimentation. An example of first year project is the numerical simulation of a pendulum with friction. This is fairly advanced material that wasn't covered by previous undergrad courses. This approach was really well received by the students.
Why python is the language of choice?
Python is a great tool for quick prototyping, unlike Java. It also is a general purpose programming language, suitable for many kinds of computing tasks, unlike matlab. In previous years, students were doing matlab after an introduction course with Java. It was hard to establish a connection between what was learned in the programming course, and all the apparent magic going on in matlab. Python offers a natural transition from matlab-style flat scripts, to full fledged object-oriented applications with complex GUIs, data processing and results reporting.
Python is executable pseudo-code. This helps students focus on the problem to solve, rather than deal with the syntax. Producing an algorithmic solution improves how student effectively understands the material. This is what should matter first.
Fighting deployment hell
One of the biggest problem with a python scientific stack, is the pain involved when deploying an application with its dependencies, or just setting up the right development environment for scientific computing. There are several distributions (Enthought Python Distribution, python(x,y), SAGE, linux distribution package systems, ...). This is a lot to maintain in parallel. Throw in some C and FORTRAN extensions (which are ubiquituous as soon as you're using python for science) and you're in for hell of a ride.
To prevent spending too much time on this, the course officially supports Ubuntu as the only platform (dualboot or virtualbox). Students can then get all the required software through a dedicated package repository (e.g.
apt-get install mycourse).
Install parties are organized to help beginners get their setup ready. Experience shows that people are more happy with this approach.
People who don't like Ubuntu will need to figure out the installation by themselves.
Research & High Performance Computing
"Python is slow"
Large scale projects
Scientists often need to launch long computations on computer clusters. mpi4py are python bindings for the popular message passing library MPI, which is one way to enable exploitation of multiprocessor architectures.
A typical scenario would be using python to control the pipeline, set the simulation parameters, and let the high-performance, often legacy FORTRAN code, do the number crunching. Numpy and scipy can also be used for that kind of heavy lifting.
Further improving the adoption of python in science
One of the strong points of matlab is the ubiquity of code samples available. We need that for the python scientific stack.
Unification of interfaces between ODE solvers, linear algebra packages, visualization packages is much needed.
There are still skeptics about whether a "scripting language" can be a fitting choice for high-performance computations. Finding and referencing real-world examples in which python was succesfully used helps. The bottom line being that one doesn't need to write the whole application in C or C++ for it to be fast.