Toast Driven

← Back to November 10, 2008

A Brief Introduction To Multiprocessing

Parallel processing has been getting increasing attention lately online, and with good reason. As CPU manufacturers start adding more and more cores to their processors, creating parallel code is a great way to improve performance. There are a lot of options out there for parallel development (such as functional languages or developer toolkits). However, if you have a substantial codebase already in Python, you have a relatively new option to help you get into parallel programming.

Along with the release of Python 2.6 came a new module called multiprocessing. The multiprocessing module lets you write parallelized code using processes in relatively simple code. By leveraging system processes instead of threads, multiprocessing lets you avoid issues like the GIL.

For those of us still on Python 2.5, some kind souls backported the library (with sources available on Google Code and the Cheeseshop) to Python 2.4/2.5. It has no dependencies I am aware of and installed cleanly/easily on the first try. As a warning, there are occasional small API differences (such as 2.6's applyAsync is apply_async in 2.5) so if something's not working right, dir() and help() in a shell are your friends.

This post assumes you have a little bit of an idea about important concepts in parallel programming (avoiding global/shared state, locking, units of work, etc). Finally, please understand I am not an expert on this subject, just an enthusiast playing with interesting code. Let's cover some basics first, then we'll look at a complete working program.

Processes, Queues and Locks, Oh My!

The multiprocessing module comes with plenty of built-in options for building a parallel application. But the three most basic (and safest) are the Process, Queue and Lock classes. Since the Process is the central figure in this story, it's important to understand what it is and what it does.


The Process is an abstraction that sets up another (Python) process, provides it code to run and a way for the parent application to control execution. Common usage is pretty simple:

from multiprocessing import Process

def say_hello(name='world'):
    print "Hello, %s" % name

p = Process(target=say_hello)

We import the Process class, create a function the process will run, then instantiate a Process object with the function it should run. Nothing has happened yet and won't until we tell it to begin via p.start(). The process will run and return it's result. Finally, we tell the process to complete via p.join(). NOTE - Without the p.join(), the child process will sit idle and not terminate, becoming a zombie you must manually kill. If you want to pass arguments to the function, simply provide a args keyword argument like so:

from multiprocessing import Process

def say_hello(name='world'):
    print "Hello, %s" % name

p = Process(target=say_hello, args=('Daniel',))

You can create as many processes as you like/need, but be aware that there is a limit on which adding more processes will increase performance that is NOT constant between computer (as the number of cores, system scheduler and other factors will vary). You'll likely want to build in customizable settings in this regard.


Queue objects are exactly what they sound like: a thread/process safe, FIFO data structure. They can store any pickle-able Python object (though simple ones are best) and are extremely useful for sharing data between processes. Usage is easy:

from multiprocessing import Queue

q = Queue()

q.put('Why hello there!')
q.put(['a', 1, {'b': 'c'}])

q.get() # Returns 'Why hello there!'
q.get() # Returns ['a', 1, {'b': 'c'}]

Queues are especially useful when passed as a parameter to a Process' target function to enable the Process to consume (or return) data.


Like Queue objects, Locks are relatively straightforward. They allow your code to claim the lock, blocking other processes from executing similar code until the process has completed and release the lock. Again, simple usage:

from multiprocessing import Lock

l = Lock()

print 'Ha! Only I can write to stdout!'

With that, we have all the basic building blocks we need to write a simple multiprocessing application.

A Simple Application

Being someone who maintains multiple websites, it'd be useful to have a tool to monitor my sites' statuses (I know other better tools exist, bear with me). Combined with httplib2 (which will be involved in another near-future post), we can build a simple app that does this. First, the complete source:

#!/usr/bin/env python
import httplib2
from multiprocessing import Lock, Process, Queue, current_process

def worker(work_queue, done_queue):
        for url in iter(work_queue.get, 'STOP'):
            status_code = print_site_status(url)
            done_queue.put("%s - %s got %s." % (current_process().name, url, status_code))
    except Exception, e:
        done_queue.put("%s failed on %s with: %s" % (current_process().name, url, e.message))
    return True

def print_site_status(url):
    http = httplib2.Http(timeout=10)
    headers, content = http.request(url)
    return headers.get('status', 'no response')

def main():
    sites = (
    workers = 2
    work_queue = Queue()
    done_queue = Queue()
    processes = []
    for url in sites:
    for w in xrange(workers):
        p = Process(target=worker, args=(work_queue, done_queue))
    for p in processes:
    for status in iter(done_queue.get, 'STOP'):
        print status

if __name__ == '__main__':

We can write this script mostly as we would a single process status checker. The print_site_status function does exactly what we'd normally have it do. In our main(), we simply define a list of sites we want to check (in this case, some of my favorite webcomics).

We place these URLs in a Queue so that our processes have a common source they can remove URLs from and check. We then create and start the number of processes we want, storing them in a list for controlling in the future (you may also want to read up on Pools in the multiprocessing documentation). Finally, after we've started all of our processes and they're churning away at URLs, we add a 'STOP' sentinel for each of our processes and then tell the processes to join.

Each process places the result of its efforts in the done_queue and once all other processing is done and the child processes have exited, we iterate through the done_queue and print out the results.


Hopefully, this has been a helpful brief introduction to parallel programming with Python. I have noticed a few idiosyncrasies with the backported version of multiprocessing (an early version of the code used locks and simply printed the status to the screen, avoiding the use of a done_queue, but would periodically fail in 2.5 only), but am overall impressed with how relatively easy it is to use given the typical difficulties experienced with parallel code.

And there are tons of applications of these techniques, especially in the web world. Everything from spiders (like the one our sys admin wrote) to search indexers to API scrapers and beyond, there's lots of potential.

Toast Driven