Python is Only Slow If You Use it Wrong

Author

Bup

  • Written in Python
  • Backup software
  • Uses Git as a data store
  • 80 megs/second

sshuttle

  • VPN that handles wireless speeds
  • Also in Python

How to Use Python Wrong

  • Tight Inner Loops
  • In compiled languages, you have these often
  • Really bad in Python
  • Line of code in Python is 80-100x slower than C
  • Keep it in a higher level

Ways to Make it Fast

  • Use Regex and C modules
    • Word based instead of char based ~5x faster
    • Will run it in C
    • Most of bup is Python, small bit in C to speed it up

  • 100% Pure is not pragmatic

  • CPython has a really good C API
    • Java doesn’t, it’s super painful
  • Python + C is winning so far
    • C is for tight inner loops
    • Python for the higher level

Threads

  • Computation threads are useless, because of GIL
    • Sometimes worse than single threaded
  • Okay for I/O
    • GIL will release for I/O
  • fork() works great for both
    • Recommend to use it all the time
    • No GIL
    • Trick is getting info from process to process
    • Bup uses this
    • No weird locking interactions
  • C modules can use threads
    • Can release GIL when you get objects
    • Run threads
    • Get GIL when computations are done
    • Can get high performance

  • CPU Bound threads in Python is doing it wrong

  • Question from audience: Scipy has Weave, which will allow you to inline C code. * Dynamic compilation

  • There are workarounds for the GIL

Garbage Collection

  • Python is both refcounting and gc

  • Refcounting
    • Whenever you use a variable, increase reference count
    • Whenever you stop, decrease the reference count
    • Terrible, terrible thing with threads
    • Need to lock on refcounts
    • GIL solves this problem
  • Shows graphs of programs memory and time
    • Allocates 10k of space a lot
    • Refcounting sematics allow Python lower mem usage than Java
  • Testing Java
    • 3 different tests
    • Shows one where it allocates as much memory as possible
  • Sometimes Python is Garbage Collected
    • Mutual referencing objects that have ref count of one
    • Backup GC finds this, and collects them
    • Shows example on how to do this
    • Pretty complicated in order to get across the GC
    • Then it relies on sucking up tons of memory, and getting it later

Advice: Stay away from GC

  • Break circular references

  • Most common, trees with reference to parents
    • Full tree need to be GC’ed

  • Better: use the weakref module

Deterministic Destructors

  • Win32 example of two writers to a file

  • Win32 doesn’t allow two writers

  • CPython allows it because it closes the writer because of refcounting

  • This causes deterministic behavior, unlike ‘real’ gc
    • In Python you don’t need to manage many resources
    • Files, database handles, etc.
  • Some people are trying to take this away
    • Pypy?
    • with statement isn’t a desirable alternative

HelloMark

  • Fork and exec “Hello World” 20x

  • Demonstrates startup times

  • Jython takes 15 seconds, slower than C+valgrind

  • Shows what you want to write command line tools in

  • pyc + CPython files are awesome for this
    • Django and Tornado can reload really quickly

  • Pypy loses in this regard

Summary