Embracing the GIL

• by David Beazley
• slides: http://dabeaz.com/talks/EmbraceGIL/

Embracing the GIL could be better

• People love to talk about it
• Rant about it
• Godwin’s Law of Python?

premise

• People love to hate on them
• That’s because threads are useful
• Threads make great stuff work
• Even if you don’t see them directly

The Gil in a Nutshell

• Every Python file gets compiled into VM instructions
• In cpython, it is unsafe to execute instruction concurrently
• Hence: Locking

What GIL protects

Note

duh missed this

Major GIL issues

• Threads using multiple CPUs

The Challenge

• The GIL is unlikely to go away anytime soon
• Can it be improved? Yes!
• How can it be done?
• How about Python 3!

Experiment

• request/reply server for size-prefixed messages
• each method has payload/header

Why this experiment?

• Comes up in a lot of contexts
• Involves IO
• Used as a foundation for a lot of other things

Five different implementations

• 1000 iterations of some simple code

• Done on EC2 with nothing else running

• implementations

  • C + 0mq
  • Python + 0mq
  • Python + multiprocessing
  • Python + blocking sockets
  • Python + nonblocking sockets

• Results

  • All finish in about 13 seconds

What happens when you introduce a thread?

• What does it do to the performance?

  • C + 0mq (samish seed)
  • Python + 0mq (7x slower)
  • Python + multiprocessing (8.9x slower)
  • Python + blocking sockets (approx 10.x slower)
  • Python + nonblocking sockets (approx 10.x slower)

Commentary

• Simple test

• Not a hard-core realistic talk

• How about PyPI?

  • What? Older version was 567 slower!
  • New version is much faster. .. note:: Get results!

Warning

Distracted by some work stuff. Missed some awesome stuff here.

Performance Explained - thread priorities

• To fix this, you need priorities
• The original “Fix GIL” patch had priorities
• That should be revisited

Another experiment

• David’s 3.2 fork with priorities
• Not suitable for real work
• Interesting for testing
• Lets you set the priorities manually

import sys
import threading
def spin(value):

    sys.set_priority(-1)

Some thoughts

• Huge boost in Python with only minor changes to a few files

• Is this the only GIL improvement?

  • No
  • There are other ways to do it
  • GIL released on non-blocking I/O operations

PyDanny take away

• Now I think I grok the GIL issues finally
• Ya, me is slow. :)