Pyston, Pandas and scientific computing in Python

[Pham Cong Dinh]

Hi,

There is an interesting discussion on Pyston, a JIT-based Python implementation, backed by Dropbox, and its position on Python scientific computing in Python.

https://github.com/dropbox/pyston/issues/634

When most of people agrees that CPython may not as fast as expected for scientific computing tasks, they tend to realize that PyPy is not relevant when they seek for a as-fast-as-C++ Python implementation which is possible to work well with Pandas and/or any other components in PyData. Pyston's target is to "push Python into domains dominated by traditional systems languages like C++" (1), be compatible to CPython and run unmodified Python code: " It uses modern JIT techniques on top of LLVM, and natively supports many CPython C extension modules via a recompile." (2)

https://blogs.dropbox.com/tech/2014/04/introducing-pyston-an-upcoming-jit-based-python-implementation/

They actually reuse CPython code to move faster

http://blog.pyston.org/2015/02/24/pyston-0-3-self-hosting-sufficiency/

It is great if Pandas, NumPy devs can group around Pyston to push its development forward. Dropbox is a well-funded firm. But I don't know if their upper managers agree to add Python scientific computing libraries support as a priority in their roadmap.

PS: I am no expert in this field.

Dinh

1. https://blogs.dropbox.com/tech/2014/04/introducing-pyston-an-upcoming-jit-based-python-implementation/
2. http://blog.pyston.org

[Ivan Ogasawara]

Very interesting! :)

--
You received this message because you are subscribed to the Google Groups "PyData" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pydata+un...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Big Stone]

Apparently, they still target Python2.7, not Python3.

see https://github.com/dropbox/pyston/blob/master/README.md

[Nathaniel Smith]

Sure, and also they haven't even finished implementing the language, and what they do implement mostly isn't faster than cpython yet, for the very sensible reason that they're focusing on functionality before speed. Definitely not something you'll be deploying in production in the next 6 months; these things take time.

Still, it's pretty exciting that there's someone working on a full python jit and that "run numpy" is not just part of their test suite but IIUC already somewhat functional. If you look at their local patch to numpy, the only really functionality-breaking hack they needed to get it to build/import was to disable numpy.random. Everything else is either working around temporary bugs in pyston, arguable bugs in numpy that cpython lets us get away with but that could be fixed, or minor stuff that would be reasonable to #ifdef in numpy upstream.

-n

On Aug 14, 2015 13:08, "Big Stone" <stone...@gmail.com> wrote:

Apparently, they still target Python2.7, not Python3.

see https://github.com/dropbox/pyston/blob/master/README.md

[Big Stone]

"working on a full python jit" is the fashion of the year. See Microsoft

https://www.youtube.com/watch?v=_5vLWe4d8X8