The State of PyPy (April 2015)

Modern programming languages are expected to offer high-level abstractions at reasonable speed. Especially scripting languages like Ruby, PHP and Python are highly dynamic and flexible, which makes performance optimizations difficult to implement.

While Python 3 offers a lot more features compared to Python 2, these features don't always improve runtime performance. In the pystone benchmark, execution speed has even decreased over the recent years.
PyPy, a relatively new implementation of Python with a focus on performance tries to improve the situation - with obvious success. Code execution is much faster thanks to the built-in just in time compiler which optimizes code at runtime.

In this talk I give an overview of PyPy and how it works internally. Also I talk a bit about PyPy STM, a new approach for parallel code execution that can take advantage of modern multi-core hardware.

Transcript

1. PyPy
   The state of
   2015-04 @matthiasendler
   

   View Slide

2. Why PyPy?
   "Scientific American" Vol. XXXVI, No. 19, May 12, 1877 p.294 New York: Munn &Company (Publisher)
   

   View Slide

3. Why PyPy?
   What keeps you from using switching to Python 3?
   

   View Slide

4. Why PyPy?
   No more support for older versions
   Major bugfixes
   New language features
   Improved performance
   Incentives to switch to a new language version
   

   View Slide

5. Why PyPy?
   More speed means…
   less servers
   new possibilities
   

   View Slide

6. Why PyPy?
   Two ways to speed up a scripting language
   •Improve the implementation
   •Improve the runtime
   

   View Slide

7. Why PyPy?
   Two ways to speed up a scripting language
   •Improve the implementation
   •Improve the runtime
   

   View Slide

8. Why PyPy?
   Warning:

   Benchmarks are always error-prone.

   Always measure yourself!

   This is just a trend.
   

   View Slide

9. 0
   35
   70
   105
   140
   1.8.7 1.9.3 2.0.0 2.1.0 2.2.0
   Relative Ruby performance
   Why PyPy?
   Source: isrubyfastyet.com
   speed (%)
   version
   baseline (100%)
   +12%
   +30%
   +35% +36%
   

   View Slide

10. Why PyPy?
    0
    65
    130
    195
    260
    5.2 5.3 5.4 5.5 5.6 7.0
    Relative PHP performance
    Source: PHPNG a new core for PHP7 by Dmitry Stogov
    baseline (100%)
    speed (%)
    version
    +13%
    +36% +39%
    +50%
    +149%
    

    View Slide

11. Why PyPy?
    25
    50
    75
    100
    2.6 2.7 3.4
    Relative Python performance
    Source: pystone benchmark on Core i7, 16 GB RAM
    baseline (100%)
    speed (%)
    version
    -10%
    -3%
    

    View Slide

12. Why PyPy?
    0
    38
    75
    113
    150
    2007 2009 2011 2013 2015
    PHP Python Ruby
    baseline (100%)
    speed (%)
    time
    Relative performance improvements per language over time
    

    View Slide

13. Why PyPy?
    The net result of the 3.0 generalizations is that 

    Python 3.0 runs the pystone benchmark around 10% slower than Python 2.5.
    Most likely the biggest cause is the removal of special-casing for small
    integers. There’s room for improvement, but it will happen after 3.0 is
    released!
    docs.python.org/3/whatsnew/3.0.html
    

    View Slide

14. Why PyPy?
    Two ways to speed up a scripting language
    •Improve the implementation
    •Improve the runtime
    

    View Slide

15. Why PyPy?
    Ruby Rubinius
    PyPy
    HHVM
    Python
    PHP
    

    View Slide

16. About PyPy
    Cottrell & Babcock's air spring printing press
    Source: Scientific American, June 22, 1872
    

    View Slide

17. What is PyPy?
    Followup of Psyco runtime
    Version 1.0 in 2007
    Python written in Python
    Just-In-Time compiler
    Compatible with Python 2.7 and 3.4
    

    View Slide

18. What is PyPy?
    Followup of Psyco runtime
    Version 1.0 in 2007
    Python written in Python RPython
    Just-In-Time compiler
    Compatible with Python 2.7 and 3.4
    

    View Slide

19. PyPy implementation
    • Variables in any given control flow path must have only a single type
    • No multiple inheritance
    • No generators
    • Most dunder-methods can't be used.
    RPython is a subset of Python
    http:/
    /www.quora.com/PyPy/How-annoying-is-writing-RPython-compared-to-Python
    

    View Slide

20. PyPy implementation
    Tracing JIT compiler
    http:/
    /www.aosabook.org/en/pypy.html
    Your Python program
    

    View Slide

21. PyPy implementation
    Tracing JIT compiler
    http:/
    /www.aosabook.org/en/pypy.html
    „Hot“ loop
    

    View Slide

22. PyPy implementation
    Tracing JIT compiler
    http:/
    /www.aosabook.org/en/pypy.html
    can_enter_jit()
    jit_merge_point()
    

    View Slide

23. PyPy implementation
    Tracing JIT compiler
    http:/
    /www.aosabook.org/en/pypy.html
    Optimized machine code
    

    View Slide

24. PyPy Speed
    speed.pypy.org
    

    View Slide

25. Upright boring machine, 1872
    Source: Scientific American, Nov 23,1872
    Demo
    

    View Slide

26. The future of PyPy
    Berryman's steam trap, 1872
    Source: Scientific American, Sept 14,1872
    

    View Slide

27. Project milestones
    2008 Django support
    2010 First JIT-compiler
    2011 Compatibility with CPython 2.7
    2014 Basic ARM support
    CPython 3 support
    Improve compatibility with C extensions
    NumPyPy
    Multi-threading support
    

    View Slide

28. Project milestones
    2008 Django support
    2010 First JIT-compiler
    2011 Compatibility with CPython 2.7
    2014 Basic ARM support
    CPython 3 support
    Improve compatibility with C extensions
    NumPyPy
    Multi-threading support
    

    View Slide

29. PyPy STM
    

    View Slide

30. PyPy STM
    http://dabeaz.com/GIL/gilvis/
    GIL locking
    

    View Slide

31. PyPy STM
    10 loops, best of 3: 1.2 sec per loop
    10 loops, best of 3: 822 msec per loop
    from threading import Thread
    def count(n):
    while n > 0:
    n -= 1
    def run():
    t1 = Thread(target=count, args=(10000000,))
    t1.start()
    t2 = Thread(target=count, args=(10000000,))
    t2.start()
    t1.join(); t2.join()
    def count(n):
    while n > 0:
    n -= 1
    def run():
    count(10000000)
    count(10000000)
    Inside the Python GIL - David Beazley
    

    View Slide

32. PyPy STM
    T1
    T2
    T3
    x
    y
    z
    Thread 1
    Shared memory T1
    T2
    T3
    Thread 2
    

    View Slide

33. PyPy STM
    T1
    T2
    T3
    x = 3
    x
    y
    z
    Thread 1
    Shared memory T1
    T2
    T3
    y=5
    Thread 2
    

    View Slide

34. PyPy STM
    T1
    T2
    T3
    3
    5
    x
    y
    z
    Thread 1
    Shared memory T1
    T2
    T3
    Thread 2
    

    View Slide

35. PyPy STM
    T1
    T2
    T3
    x = 18

    y= 12
    3
    5
    x
    y
    z
    Thread 1
    Shared memory T1
    T2
    T3
    x = 2
    Thread 2
    

    View Slide

36. PyPy STM
    T1
    T2
    T3
    2
    5
    x
    y
    z
    Thread 1
    Shared memory T1
    T2
    T3
    Thread 2
    

    View Slide

37. Lathrop's sewing machine, 1872
    Source: Scientific American, Oct 26,1872
    Demo
    

    View Slide

38. Conclusion
    …already awesome for sequential tasks
    …gradually improving parallel performance
    …a very important Python project
    PyPy is…
    

    View Slide

39. Further reading
    Python 3 Performance
    https:/
    /speakerdeck.com/pyconslides/python-3-dot-3-trust-me-its-
    better-than-python-2-dot-7-by-dr-brett-cannon
    https:/
    /www.youtube.com/watch?v=UJPdhx5zTaw
    

    View Slide

40. Further reading
    RPython
    http:/
    /tratt.net/laurie/blog/entries/
    fast_enough_vms_in_fast_enough_time
    

    View Slide

41. Further reading
    PyPy STM
    http:/
    /pypy.readthedocs.org/en/latest/stm.html
    http:/
    /morepypy.blogspot.de/
    Demo code from
    http:/
    /www.ripton.net/blog/?p=104
    

    View Slide

42. Blueprints provided by Marcel Douwe Dekker under Creative Commons.
    https://www.flickr.com/photos/marceldouwedekker/7564782310/in/
    photostream/
    

    Blueprint by Joy Oil Co Ltd (http://commons.wikimedia.org/wiki/
    File:Joy_Oil_gas_station_blueprints.jpg)
    Credits
    

    View Slide

43. by Simon Child
    and by Luis Prado
    by Océan Bussard
    by Luis Rodrigues
    by Mister Pixel
    by Alexander Wiefel
    by Darin S
    by spotted paint
    

    The icons used in this presentation were downloaded from The Noun Project.

    Thanks to the following artists:
    

    View Slide