Open Stack

On 05/04/2013 07:59 PM, Matthew Sherborne wrote:
> So I tried to speed up the running of my test suite and failed.
> 
> I thought I'd share my experience in case someone else in the future
> googles before
>  trying the same thing, or in case some genius out there sees some fatal
> flaw in my plan.
> 
> Here's what I did:
> 
> I'm on Ubuntu Raring. I have both gcc and clang compilers installed.
> 
> I did some experiments with a c++ test program and found that clang with
> the -O3 optimization flag creates a faster program than gcc with -O6.
> 
> So I thought, I'll compile my own python and venv with python and clang
> and -O3.
> 
> So I compiled my python from source with clang and -O3 and installed it
> in ~/python.
> 
> I deleted my venv for nova and, rebuilt it all under the new python.
> 
> Here are the *results* of: time ./run_tests.sh
> 
> *standard python:*
> 
> real    7m4.077s
> user    21m0.080s
> sys 0m24.596s
> 
> *with clang compiled python:*
> 
> real    8m11.396s
> user    24m28.824s
> sys 0m25.408s
> 
> I tried it a few times and got similar results. So it seems that the
> stock ubuntu python is faster than my cmake -O3 compiled one.
> 
> All comments and suggestions welcomed.

Nice experiment! I do love seeing actual numbers.

BTW - we've got a back-burner project to add support for the new python
wheel file format to our pypi mirror, and then to build wheels of things
when we pull them down from upstream. While this will do nothing to
speed up the actual test runs, it should make virtualenv creation WAY
quicker, since we won't have to continually re-compile libxml...

thanks for experimenting with that!