Python Performance Boost by using Profile Guided Optimization

By using GCC PGO when building the Python interpreter, I've managed to achieve a healthy 10% speedup for CPU bound Python programs.

About PGO

PGO works by compiling the program (in my case Python) with special flags, running that program on a benchmark test suite and then re-compiling, which will take into account the past performance runs to better optimize the resulting program.

The Specs

First up - here's what I am using: * Python 2.7.7 * GCC 4.8.2 * Ubuntu 14.04 (64 bit) * Intel Q6600 @ 2.40GHz CPU

Benchmarking programs used: * pybench (part of Python source tree) * pyDes (encryption algorithm in pure Python)

Compilation steps

This is method I've used to build Python on Linux using GCC and PGO:

# unpack it
tar xf Python-2.7.7.tgz
cd Python-2.7.7

# build it with PGO generation
export CFLAGS="-fprofile-generate"
export CXXFLAGS="-fprofile-generate"
export LDFLAGS="-fprofile-generate"

# run profile script
make run_profile_task  # I had to tweak the makefile to add LD_LIBRARY_PATH

# build it again with PGO collection
make clean
export CFLAGS="-fprofile-use -fprofile-correction"
export CXXFLAGS="-fprofile-use -fprofile-correction"
export LDFLAGS="-fprofile-use -fprofile-correction"

# profit!
make install


Here the results of PGO v's non PGO builds:

and the raw chart data:

Script Without PGO With PGO Increase
pybench 4285ms 3792ms 13%
pyDes 5219ms 4733ms 10%

You may notice that pyDes had a smaller improvement than pybench, and that's caused by pybench being the application that the optimization was derived from - so pybench is the best case scenario. If your using PGO yourself, you should be running (optimizing) using your real world application, in order to get the most benefit.


So PGO is more work during the compilation step - but it does have a nice performance benefit one it's fully compiled.

Does this mean Komodo will be using PGO builds sometime in the future? We certainly hope so, as the underlying Firefox code base does support it... we should be able to add the other relevant parts to get a complete PGO Komodo build... and I am excited to take on this project!