unladen-swallow with numpy/scipy

has anyone used unladen-swallow with numpy/scipy for numeric/scientific applications? Is it significantly faster in y开发者_StackOverflow中文版our experience? Any opinions would be great.

Nobody has extensive experience with Unladen Swallow yet (except the developers), so it's going to be difficult to find many people who can discuss it. Also, with the talk of merging Unladen Swallow (which is built using LLVM) with the CPython runtime, things are going to be something of a moving target until everything's more stable.

There are benchmarks available for Unladen Swallow, but numpy and scipy aren't included. As the developers themselves explain: "... the performance of extension modules like numpy is uninteresting since numpy's core routines are implemented in C".

In short, if you're writing good code for numpy and scipy, your code won't run "significantly faster" under Unladen Swallow, since it's already running below the virtual machine level. If you're writing bad code for numpy and scipy, you need to fix your code, then refer back to the first sentence.

It should be faster. I have not tested it myself, but i just got back from pycon and they had a talk about unladen-swallow in which they mentioned the performance increase with numpy and other packages. You can watch the talk here.

On the question, not an answer:

Total runtime = python + numpy + interface,  
                cpython/unladenswallow + mostlyC + interface.

Without real data on how these 3 split -- 20 70 10, 40 40 20 ? and that for > 1 benchmark,
there's no way of telling which way is up.