C Extension Benchmarks

[Timothy Baldridge]

In order to get an idea of what sort of performance gains we'd see by going with Clojure-py on LLVM via Python extension modules, I ran a simple fib benchmark:

9227465

Time 5.65887999535 # Pure Python

9227465

Time 3.04140090942 # Simple Native compiled

9227465

Time 1.86417007446 # Call optimized

9227465

Time 0.0701291561127 # Type hinted. 

Times are in sec. Source is here https://gist.github.com/4337160

So simply translating clojure code to native code using Python C APIs, should result in almost a 2x performance boost. Optimizing how native functions call each other, should result in another 2x boost. Supporting type hints will result in about a 72x boost (and is actually faster than clojure-jvm at that point, from what I can tell). 

So this looks hopeful. Even with the worst case situation, we should at least be quite a bit faster than python. 

Timothy

[Konrad Hinsen]

--On 19 décembre 2012 07:51:34 -0700 Timothy Baldridge

<tbald...@gmail.com> wrote:

> So simply translating clojure code to native code using Python C APIs,
> should result in almost a 2x performance boost. Optimizing how native

I suppose you could get less of a boost for code that does more work inside
Python objects' methods. At least that's my experience with compiling plain
Python code using Cython. Your example uses very simple Python methods
(integer arithmetic), so the interpretation overhead in your function is
relatively more important.

Konrad.