NumPy SciPy BenchmarksJul 07, 2013 by Andy R. Terrel
At PyCon2013, a group of about 30 developers got together to discuss NumPy and SciPy. The room mostly included folks that use NumPy and SciPy extensively, such as myself, a few who were just learning, and several core contributors (if not currently in the past), such as Eric Jones, Travis Oliphant, Anthony Scopatz, and Jake Vanderplas. It was an impromptu meeting, as most birds of a feather sessions are, so I don't think we even got the full number of interested participants at PyCon2013. The scientific large support at PyCon was surprising, since most of us scientists think the wider Python ecosystem leaves us behind in many of their activities.
There were three topics we discussed: briefly about the development and maintenance of the libraries, about the use of newer technologies, e.g. Cython, PyPy, Copperhead, Numba, and finally about having a benchmark for new implementations of NumPy to see how complete they are. Since there were not very many core developers in the room, it was hard to really push a topic on the development side but the other two topics were the most discussed.
To briefly summarize the maintenance and future plans section, I asked about how the community was doing since I've been organizing this whole conference thing and haven't been on the mailing lists in a while. Travis mentioned that NumFOCUS was trying to raise money for the projects, especially around the topic of paying a maintainer.
Next we turn to the topic of benchmarks. At the end we decided that there are at least two types of benchmarks that are really needed by the community. First, is a set of miniature applications that show off the tools well. Since there is such a wide variety of tools available right now, it is difficult to know what to invest in as a developer. Ian Osvald mentioned his book that includes one of the only major comparisons I've seen in the tools. I suggested we look at how the Mantevo project implemented its mini-apps for the US DOE. Another suggestion was to organize the applications in the seven dwarfs of scientific computing suggested by Phil Collela. I believe these have been renamed to motifs and extended to eleven by the Berkeley crowd.
The idea moving forward will be to collect a number of applications, give a reference implementation, and accept contributions showing how to implement them in a Python runnable way. The idea is not to make everyone use Python for everything but to show off the capabilities of Python as an integration (and implementation) platform. One participant brought up that we could have a site like the Computer Language Benchmarks, while we agreed in spirit enough of us do not like that benchmark because of its purist ideas that applications should stick to a single language. To be friendly to vendor solutions, I'm willing to accept closed runtimes but since the point is to highlight the code, all code must be free and open, preferably BSD. I am sure we will refine the rules as we go.
Matti, a contributor to NumPyPy, brought up the idea of creating a suite of benchmarks to compare different implementations of NumPy. The idea here is once again not to compete ruthlessly for fame and glory, but rather provide a baseline for developer to hit when implementing alternatives and users to know how complete and single implementation is. Right now I know of several NumPy implementations (or close to implementations), e.g., NumPyPy, Numba, Copperhead, and a benchmark will be a welcome addition to the tools.
The plan to start off will be to grab out some tests from the NumPy regression tests and grab a benchmark runner. Matti suggested trying to port the PyPy runners for our needs. Once we get something up and going we can try to put it up somewhere for people to make contributions.
There were many concerns about the benchmarks not being productive and bringing out too much competition. My view is that we should try to be as open and welcoming as the rest of the Python ecosystem. That is let the numbers speak for themselves and make these benchmarks something anyone can contribute to.
A final aspect of the benchmark would be to show off that NumPy is useful for more then numeric algorithms. Many people see "Num" and assume there is no use for their specific needs, but in actuality, NumPy should be thought of as more a typed array. This allows fast access to contiguous blocks (although the element access is often slower than lists).
This post is intended to start a wider discussion among the community. I'm posting here because there are enough places in the community and I wanted a single place rather than four or five mailing lists. The hope is we can start a repository and get some tests rolling.