Speed

../../_images/33175625804_e225b90f3e_k_d.jpg

CPython, the most commonly used implementation of Python, is slow for CPU bound tasks. PyPy is fast.

Using a slightly modified version of David Beazley’s CPU bound test code (added loop for multiple tests), you can see the difference between CPython and PyPy’s processing.

# PyPy
$ ./pypy -V
Python 2.7.1 (7773f8fc4223, Nov 18 2011, 18:47:10)
[PyPy 1.7.0 with GCC 4.4.3]
$ ./pypy measure2.py
0.0683999061584
0.0483210086823
0.0388588905334
0.0440690517426
0.0695300102234
# CPython
$ ./python -V
Python 2.7.1
$ ./python measure2.py
1.06774401665
1.45412397385
1.51485204697
1.54693889618
1.60109114647

Context

The GIL

The GIL (Global Interpreter Lock) is how Python allows multiple threads to operate at the same time. Python’s memory management isn’t entirely thread-safe, so the GIL is required to prevent multiple threads from running the same Python code at once.

David Beazley has a great guide on how the GIL operates. He also covers the new GIL in Python 3.2. His results show that maximizing performance in a Python application requires a strong understanding of the GIL, how it affects your specific application, how many cores you have, and where your application bottlenecks are.

C Extensions

The GIL

Special care must be taken when writing C extensions to make sure you register your threads with the interpreter.

C Extensions

Cython

Cython implements a superset of the Python language with which you are able to write C and C++ modules for Python. Cython also allows you to call functions from compiled C libraries. Using Cython allows you to take advantage of Python’s strong typing of variables and operations.

Here’s an example of strong typing with Cython:

def primes(int kmax):
"""Calculation of prime numbers with additional
Cython keywords"""

    cdef int n, k, i
    cdef int p[1000]
    result = []
    if kmax > 1000:
        kmax = 1000
    k = 0
    n = 2
    while k < kmax:
        i = 0
        while i < k and n % p[i] != 0:
            i = i + 1
        if i == k:
            p[k] = n
            k = k + 1
            result.append(n)
        n = n + 1
    return result

This implementation of an algorithm to find prime numbers has some additional keywords compared to the next one, which is implemented in pure Python:

def primes(kmax):
"""Calculation of prime numbers in standard Python syntax"""

    p = range(1000)
    result = []
    if kmax > 1000:
        kmax = 1000
    k = 0
    n = 2
    while k < kmax:
        i = 0
        while i < k and n % p[i] != 0:
            i = i + 1
        if i == k:
            p[k] = n
            k = k + 1
            result.append(n)
        n = n + 1
    return result

Notice that in the Cython version you declare integers and integer arrays to be compiled into C types while also creating a Python list:

def primes(int kmax):
    """Calculation of prime numbers with additional
    Cython keywords"""

    cdef int n, k, i
    cdef int p[1000]
    result = []
def primes(kmax):
    """Calculation of prime numbers in standard Python syntax"""

    p = range(1000)
    result = []

What is the difference? In the upper Cython version you can see the declaration of the variable types and the integer array in a similar way as in standard C. For example cdef int n,k,i in line 3. This additional type declaration (i.e. integer) allows the Cython compiler to generate more efficient C code from the second version. While standard Python code is saved in *.py files, Cython code is saved in *.pyx files.

What’s the difference in speed? Let’s try it!

import time
# Activate pyx compiler
import pyximport
pyximport.install()
import primesCy  # primes implemented with Cython
import primes  # primes implemented with Python

print("Cython:")
t1 = time.time()
print(primesCy.primes(500))
t2 = time.time()
print("Cython time: %s" % (t2 - t1))
print("")
print("Python")
t1 = time.time()
print(primes.primes(500))
t2 = time.time()
print("Python time: %s" % (t2 - t1))

These lines both need a remark:

import pyximport
pyximport.install()

The pyximport module allows you to import *.pyx files (e.g., primesCy.pyx) with the Cython-compiled version of the primes function. The pyximport.install() command allows the Python interpreter to start the Cython compiler directly to generate C code, which is automatically compiled to a *.so C library. Cython is then able to import this library for you in your Python code, easily and efficiently. With the time.time() function you are able to compare the time between these 2 different calls to find 500 prime numbers. On a standard notebook (dual core AMD E-450 1.6 GHz), the measured values are:

Cython time: 0.0054 seconds

Python time: 0.0566 seconds

And here is the output of an embedded ARM beaglebone machine:

Cython time: 0.0196 seconds

Python time: 0.3302 seconds

Pyrex

Shedskin?

Concurrency

Concurrent.futures

The concurrent.futures module is a module in the standard library that provides a “high-level interface for asynchronously executing callables”. It abstracts away a lot of the more complicated details about using multiple threads or processes for concurrency, and allows the user to focus on accomplishing the task at hand.

The concurrent.futures module exposes two main classes, the ThreadPoolExecutor and the ProcessPoolExecutor. The ThreadPoolExecutor will create a pool of worker threads that a user can submit jobs to. These jobs will then be executed in another thread when the next worker thread becomes available.

The ProcessPoolExecutor works in the same way, except instead of using multiple threads for its workers, it will use multiple processes. This makes it possible to side-step the GIL; however, because of the way things are passed to worker processes, only picklable objects can be executed and returned.

Because of the way the GIL works, a good rule of thumb is to use a ThreadPoolExecutor when the task being executed involves a lot of blocking (i.e. making requests over the network) and to use a ProcessPoolExecutor executor when the task is computationally expensive.

There are two main ways of executing things in parallel using the two Executors. One way is with the map(func, iterables) method. This works almost exactly like the builtin map() function, except it will execute everything in parallel.

from concurrent.futures import ThreadPoolExecutor
import requests

def get_webpage(url):
    page = requests.get(url)
    return page

pool = ThreadPoolExecutor(max_workers=5)

my_urls = ['http://google.com/']*10  # Create a list of urls

for page in pool.map(get_webpage, my_urls):
    # Do something with the result
    print(page.text)

For even more control, the submit(func, *args, **kwargs) method will schedule a callable to be executed ( as func(*args, **kwargs)) and returns a Future object that represents the execution of the callable.

The Future object provides various methods that can be used to check on the progress of the scheduled callable. These include:

cancel()
Attempt to cancel the call.
cancelled()
Return True if the call was successfully cancelled.
running()
Return True if the call is currently being executed and cannot be cancelled.
done()
Return True if the call was successfully cancelled or finished running.
result()
Return the value returned by the call. Note that this call will block until the scheduled callable returns by default.
exception()
Return the exception raised by the call. If no exception was raised then this returns None. Note that this will block just like result().
add_done_callback(fn)
Attach a callback function that will be executed (as fn(future)) when the scheduled callable returns.
from concurrent.futures import ProcessPoolExecutor, as_completed

def is_prime(n):
    if n % 2 == 0:
        return n, False

    sqrt_n = int(n**0.5)
    for i in range(3, sqrt_n + 1, 2):
        if n % i == 0:
            return n, False
    return n, True

PRIMES = [
    112272535095293,
    112582705942171,
    112272535095293,
    115280095190773,
    115797848077099,
    1099726899285419]

futures = []
with ProcessPoolExecutor(max_workers=4) as pool:
    # Schedule the ProcessPoolExecutor to check if a number is prime
    # and add the returned Future to our list of futures
    for p in PRIMES:
        fut = pool.submit(is_prime, p)
        futures.append(fut)

# As the jobs are completed, print out the results
for number, result in as_completed(futures):
    if result:
        print("{} is prime".format(number))
    else:
        print("{} is not prime".format(number))

The concurrent.futures module contains two helper functions for working with Futures. The as_completed(futures) function returns an iterator over the list of futures, yielding the futures as they complete.

The wait(futures) function will simply block until all futures in the list of futures provided have completed.

For more information, on using the concurrent.futures module, consult the official documentation.

threading

The standard library comes with a threading module that allows a user to work with multiple threads manually.

Running a function in another thread is as simple as passing a callable and its arguments to Thread’s constructor and then calling start():

from threading import Thread
import requests

def get_webpage(url):
    page = requests.get(url)
    return page

some_thread = Thread(get_webpage, 'http://google.com/')
some_thread.start()

To wait until the thread has terminated, call join():

some_thread.join()

After calling join(), it is always a good idea to check whether the thread is still alive (because the join call timed out):

if some_thread.is_alive():
    print("join() must have timed out.")
else:
    print("Our thread has terminated.")

Because multiple threads have access to the same section of memory, sometimes there might be situations where two or more threads are trying to write to the same resource at the same time or where the output is dependent on the sequence or timing of certain events. This is called a data race or race condition. When this happens, the output will be garbled or you may encounter problems which are difficult to debug. A good example is this Stack Overflow post.

The way this can be avoided is by using a Lock that each thread needs to acquire before writing to a shared resource. Locks can be acquired and released through either the contextmanager protocol (with statement), or by using acquire() and release() directly. Here is a (rather contrived) example:

from threading import Lock, Thread

file_lock = Lock()

def log(msg):
    with file_lock:
        open('website_changes.log', 'w') as f:
            f.write(changes)

def monitor_website(some_website):
    """
    Monitor a website and then if there are any changes,
    log them to disk.
    """
    while True:
        changes = check_for_changes(some_website)
        if changes:
            log(changes)

websites = ['http://google.com/', ... ]
for website in websites:
    t = Thread(monitor_website, website)
    t.start()

Here, we have a bunch of threads checking for changes on a list of sites and whenever there are any changes, they attempt to write those changes to a file by calling log(changes). When log() is called, it will wait to acquire the lock with with file_lock:. This ensures that at any one time, only one thread is writing to the file.

Spawning Processes

Multiprocessing