Why is PyPy3 slower at starting new Processes? - python

I would like to know why when I start new Processes (multiprocessing.Process) in PyPy3, it is actually like 2 times slower than doing the same thing on CPython. I would also like to know any solutions for this.
This is some code I wrote for illustrating this effect:
import multiprocessing as mp
from time import sleep, time
class A(object):
def __init__(self, *args, **kwargs):
# do other stuff
#self.p_conn, self.child_conn = mp.Pipe()
#self.q = mp.Queue()
def do_something(self, i):
sleep(0.1)
s = '%s * %s = %s' % (i, i, i*i)
#self.child_conn.send(s)
#self.q.put(i**2)
def run(self):
processes = []
for i in range(500):
p = mp.Process(target=self.do_something, args=(i,))
processes.append(p)
[x.start() for x in processes]
#for i in range(50):
# print(self.p_conn.recv())
#for i in range(50):
# print(self.q.get())
if __name__ == '__main__':
a = A()
s = time()
a.run()
print(f"Took {time()-s} seconds...")
CPython took around 18 seconds while PyPy3 took around 37 seconds to execute the same code. (Other tests also showed that PyPy3 is 2 times slower than CPython at starting Processes in my system...)
I would like to know how to solve the problem efficiently.

Related

Is there a way to utilize all cpu cores and power in calculations?

First, I've tried to use multithreading solution for this problem and discovered that it is not suitable for this purpose. Then I tried as community suggested to apply multiprocessing solution to bypass the GIL and even that performs poor compared to single process single thread code. Is python flawed in this domain?
Is the only solution for heavy cpu calculations is to drop python for another language?
I post my multiprocessing test code so you can get an impression.
from itertools import cycle
import random
import multiprocessing as mp
import time
# The class that represents the process
class Task(mp.Process):
def __init__(self, group=None, target=None, name=None, args=(), kwargs={}, *, daemon=None):
mp.Process.__init__(self, group=group, target=target, name=name, args=args ,kwargs=kwargs, daemon=daemon)
self.inputs = []
def run(self):
print(f"{self.name} is running")
for arr in self.inputs:
arr.sort()
def add_input(self, arr):
self.inputs.append(arr)
# A util function to cycle on iterable a finite number of times.
def finite_cycle(cycle_on, times):
infinite_cycle = cycle(cycle_on)
for _ in range(times):
yield next(infinite_cycle)
# Constants
THOUSAND = 1000
MILION = THOUSAND ** 2
PCNT = 2
TASK_CNT = 50 * THOUSAND
# Main
def main():
processes = [Task(name = f"p{pid}") for pid in range(PCNT)]
for pid in finite_cycle(range(PCNT), TASK_CNT):
processes[pid].add_input([random.randint(1,10) for _ in range(100)])
stime = time.time()
for p in processes:
p.start()
for p in processes:
p.join()
print(f"execution time: {round(time.time() - stime, 2)}")
print("finish.")
And this is the single process single thread code which is faster for every varation of the constants.
def main():
inputs = [[random.randint(1,10) for _ in range(100)] for _ in range(TASK_CNT)]
stime = time.time()
for arr in inputs:
arr.sort()
print(f"execution time: {round(time.time() - stime, 2)}")
print("finish.")
On my desktop the run methods averaged each approximately .125 seconds to run while the time elapsed between calling the first start method and the start of the first run method was approximately .23 seconds (i.e. 1628456465.1061594 - 1628456464.8741603), most of that time I believe taken by the serialization/de-serialization of self.inputs. See below, which is the original program with a few timings added.
The point is that multiprocessing has two sources of overhead that the non-multiprocessing program does not have:
Overhead in creating the processes.
Overhead in passing arguments to and getting results back from the process. This involves moving data from one address space to another (via various mechanisms) in many cases unless shared memory is being used.
Multiprocessing therefore only becomes advantageous when the processing itself (the run method in this case) is so CPU-intensive that the aforementioned costs of multiprocessing are offset by being able to "divide and conquer" the problem.
from itertools import cycle
import random
import multiprocessing as mp
import time
# The class that represents the process
class Task(mp.Process):
def __init__(self, group=None, target=None, name=None, args=(), kwargs={}, *, daemon=None):
mp.Process.__init__(self, group=group, target=target, name=name, args=args ,kwargs=kwargs, daemon=daemon)
self.inputs = []
def run(self):
t = time.time()
print(f"{self.name} is running at:", t)
for arr in self.inputs:
arr.sort()
print('elapsed time =', time.time() - t)
def add_input(self, arr):
self.inputs.append(arr)
# A util function to cycle on iterable a finite number of times.
def finite_cycle(cycle_on, times):
infinite_cycle = cycle(cycle_on)
for _ in range(times):
yield next(infinite_cycle)
# Constants
THOUSAND = 1000
MILION = THOUSAND ** 2
PCNT = 2
TASK_CNT = 50 * THOUSAND
# Main
def main():
processes = [Task(name = f"p{pid}") for pid in range(PCNT)]
for pid in finite_cycle(range(PCNT), TASK_CNT):
processes[pid].add_input([random.randint(1,10) for _ in range(100)])
stime = time.time()
print('stime =', stime)
for p in processes:
p.start()
for p in processes:
p.join()
print(f"execution time: {round(time.time() - stime, 2)}")
print("finish.")
if __name__ == '__main__':
main()
Prints:
stime = 1628456464.8741603
p0 is running at: 1628456465.1061594
elapsed time = 0.1320023536682129
p1 is running at: 1628456465.3201597
elapsed time = 0.11999750137329102
execution time: 0.62
finish.

Multiprocessing taking longer than single (normal) processing

How come the multi function that uses a multiprocessing pool to segment and process data on multiple "processes" is slower (8 seconds) than just calling the map function (6 seconds)?
from multiprocessing import Pool
import timeit
def timer(function):
def new_function():
start_time = timeit.default_timer()
function()
elapsed = timeit.default_timer() - start_time
print('Function "{name}" took {time} seconds to complete.'.format(name=function.__name__, time=elapsed))
return new_function
def cube(n):
return n*n*n
nums = range(20000000)
if __name__ == '__main__':
#timer
def multi():
pool = Pool()
res = pool.map(cube,nums)
pool.close()
pool.join()
#timer
def test():
a = map(cube,nums)
multi()
test()
Because all the dispatching logic behind pool.map creates an overhead.
Multiprocessing always create overhead of some sort, which heavily depends on its underlying implementation.
You are running a lot of very simple tasks here, hence the overhead caused by the threading logic is not compensated by the gain of parallel execution. Try to do the same test with a lesser number of more cpu-intensive task, you should see different results.
Exemple
See this modified test. Here, we have a silly cubes function that computes n^3 1000 times.
from multiprocessing import Pool
import timeit
def timer(function):
def new_function():
start_time = timeit.default_timer()
function()
elapsed = timeit.default_timer() - start_time
print('Function "{name}" took {time} seconds to complete.'.format(name=function.__name__, time=elapsed))
return new_function
def cubes(n):
for _ in range(999):
n * n * n
return n * n * n
nums = range(20000)
if __name__ == '__main__':
#timer
def multi():
pool = Pool()
res = pool.map(cubes, nums)
pool.close()
pool.join()
#timer
def test():
# On Python 3, simply calling map() returns an iterator
# tuple() collects its values for timing
a = tuple(map(cubes, nums))
multi()
test()
We now see multiprocessing is improving our timing:
Function "multi" took 0.6272498000000001 seconds to complete.
Function "test" took 2.130454 seconds to complete.

How to have multiple Python scripts interacting with each other [duplicate]

I am trying to understand threading in Python. I've looked at the documentation and examples, but quite frankly, many examples are overly sophisticated and I'm having trouble understanding them.
How do you clearly show tasks being divided for multi-threading?
Since this question was asked in 2010, there has been real simplification in how to do simple multithreading with Python with map and pool.
The code below comes from an article/blog post that you should definitely check out (no affiliation) - Parallelism in one line: A Better Model for Day to Day Threading Tasks. I'll summarize below - it ends up being just a few lines of code:
from multiprocessing.dummy import Pool as ThreadPool
pool = ThreadPool(4)
results = pool.map(my_function, my_array)
Which is the multithreaded version of:
results = []
for item in my_array:
results.append(my_function(item))
Description
Map is a cool little function, and the key to easily injecting parallelism into your Python code. For those unfamiliar, map is something lifted from functional languages like Lisp. It is a function which maps another function over a sequence.
Map handles the iteration over the sequence for us, applies the function, and stores all of the results in a handy list at the end.
Implementation
Parallel versions of the map function are provided by two libraries:multiprocessing, and also its little known, but equally fantastic step child:multiprocessing.dummy.
multiprocessing.dummy is exactly the same as multiprocessing module, but uses threads instead (an important distinction - use multiple processes for CPU-intensive tasks; threads for (and during) I/O):
multiprocessing.dummy replicates the API of multiprocessing, but is no more than a wrapper around the threading module.
import urllib2
from multiprocessing.dummy import Pool as ThreadPool
urls = [
'http://www.python.org',
'http://www.python.org/about/',
'http://www.onlamp.com/pub/a/python/2003/04/17/metaclasses.html',
'http://www.python.org/doc/',
'http://www.python.org/download/',
'http://www.python.org/getit/',
'http://www.python.org/community/',
'https://wiki.python.org/moin/',
]
# Make the Pool of workers
pool = ThreadPool(4)
# Open the URLs in their own threads
# and return the results
results = pool.map(urllib2.urlopen, urls)
# Close the pool and wait for the work to finish
pool.close()
pool.join()
And the timing results:
Single thread: 14.4 seconds
4 Pool: 3.1 seconds
8 Pool: 1.4 seconds
13 Pool: 1.3 seconds
Passing multiple arguments (works like this only in Python 3.3 and later):
To pass multiple arrays:
results = pool.starmap(function, zip(list_a, list_b))
Or to pass a constant and an array:
results = pool.starmap(function, zip(itertools.repeat(constant), list_a))
If you are using an earlier version of Python, you can pass multiple arguments via this workaround).
(Thanks to user136036 for the helpful comment.)
Here's a simple example: you need to try a few alternative URLs and return the contents of the first one to respond.
import Queue
import threading
import urllib2
# Called by each thread
def get_url(q, url):
q.put(urllib2.urlopen(url).read())
theurls = ["http://google.com", "http://yahoo.com"]
q = Queue.Queue()
for u in theurls:
t = threading.Thread(target=get_url, args = (q,u))
t.daemon = True
t.start()
s = q.get()
print s
This is a case where threading is used as a simple optimization: each subthread is waiting for a URL to resolve and respond, to put its contents on the queue; each thread is a daemon (won't keep the process up if the main thread ends -- that's more common than not); the main thread starts all subthreads, does a get on the queue to wait until one of them has done a put, then emits the results and terminates (which takes down any subthreads that might still be running, since they're daemon threads).
Proper use of threads in Python is invariably connected to I/O operations (since CPython doesn't use multiple cores to run CPU-bound tasks anyway, the only reason for threading is not blocking the process while there's a wait for some I/O). Queues are almost invariably the best way to farm out work to threads and/or collect the work's results, by the way, and they're intrinsically threadsafe, so they save you from worrying about locks, conditions, events, semaphores, and other inter-thread coordination/communication concepts.
NOTE: For actual parallelization in Python, you should use the multiprocessing module to fork multiple processes that execute in parallel (due to the global interpreter lock, Python threads provide interleaving, but they are in fact executed serially, not in parallel, and are only useful when interleaving I/O operations).
However, if you are merely looking for interleaving (or are doing I/O operations that can be parallelized despite the global interpreter lock), then the threading module is the place to start. As a really simple example, let's consider the problem of summing a large range by summing subranges in parallel:
import threading
class SummingThread(threading.Thread):
def __init__(self,low,high):
super(SummingThread, self).__init__()
self.low=low
self.high=high
self.total=0
def run(self):
for i in range(self.low,self.high):
self.total+=i
thread1 = SummingThread(0,500000)
thread2 = SummingThread(500000,1000000)
thread1.start() # This actually causes the thread to run
thread2.start()
thread1.join() # This waits until the thread has completed
thread2.join()
# At this point, both threads have completed
result = thread1.total + thread2.total
print result
Note that the above is a very stupid example, as it does absolutely no I/O and will be executed serially albeit interleaved (with the added overhead of context switching) in CPython due to the global interpreter lock.
Like others mentioned, CPython can use threads only for I/O waits due to GIL.
If you want to benefit from multiple cores for CPU-bound tasks, use multiprocessing:
from multiprocessing import Process
def f(name):
print 'hello', name
if __name__ == '__main__':
p = Process(target=f, args=('bob',))
p.start()
p.join()
Just a note: A queue is not required for threading.
This is the simplest example I could imagine that shows 10 processes running concurrently.
import threading
from random import randint
from time import sleep
def print_number(number):
# Sleeps a random 1 to 10 seconds
rand_int_var = randint(1, 10)
sleep(rand_int_var)
print "Thread " + str(number) + " slept for " + str(rand_int_var) + " seconds"
thread_list = []
for i in range(1, 10):
# Instantiates the thread
# (i) does not make a sequence, so (i,)
t = threading.Thread(target=print_number, args=(i,))
# Sticks the thread in a list so that it remains accessible
thread_list.append(t)
# Starts threads
for thread in thread_list:
thread.start()
# This blocks the calling thread until the thread whose join() method is called is terminated.
# From http://docs.python.org/2/library/threading.html#thread-objects
for thread in thread_list:
thread.join()
# Demonstrates that the main process waited for threads to complete
print "Done"
The answer from Alex Martelli helped me. However, here is a modified version that I thought was more useful (at least to me).
Updated: works in both Python 2 and Python 3
try:
# For Python 3
import queue
from urllib.request import urlopen
except:
# For Python 2
import Queue as queue
from urllib2 import urlopen
import threading
worker_data = ['http://google.com', 'http://yahoo.com', 'http://bing.com']
# Load up a queue with your data. This will handle locking
q = queue.Queue()
for url in worker_data:
q.put(url)
# Define a worker function
def worker(url_queue):
queue_full = True
while queue_full:
try:
# Get your data off the queue, and do some work
url = url_queue.get(False)
data = urlopen(url).read()
print(len(data))
except queue.Empty:
queue_full = False
# Create as many threads as you want
thread_count = 5
for i in range(thread_count):
t = threading.Thread(target=worker, args = (q,))
t.start()
Given a function, f, thread it like this:
import threading
threading.Thread(target=f).start()
To pass arguments to f
threading.Thread(target=f, args=(a,b,c)).start()
I found this very useful: create as many threads as cores and let them execute a (large) number of tasks (in this case, calling a shell program):
import Queue
import threading
import multiprocessing
import subprocess
q = Queue.Queue()
for i in range(30): # Put 30 tasks in the queue
q.put(i)
def worker():
while True:
item = q.get()
# Execute a task: call a shell program and wait until it completes
subprocess.call("echo " + str(item), shell=True)
q.task_done()
cpus = multiprocessing.cpu_count() # Detect number of cores
print("Creating %d threads" % cpus)
for i in range(cpus):
t = threading.Thread(target=worker)
t.daemon = True
t.start()
q.join() # Block until all tasks are done
Python 3 has the facility of launching parallel tasks. This makes our work easier.
It has thread pooling and process pooling.
The following gives an insight:
ThreadPoolExecutor Example (source)
import concurrent.futures
import urllib.request
URLS = ['http://www.foxnews.com/',
'http://www.cnn.com/',
'http://europe.wsj.com/',
'http://www.bbc.co.uk/',
'http://some-made-up-domain.com/']
# Retrieve a single page and report the URL and contents
def load_url(url, timeout):
with urllib.request.urlopen(url, timeout=timeout) as conn:
return conn.read()
# We can use a with statement to ensure threads are cleaned up promptly
with concurrent.futures.ThreadPoolExecutor(max_workers=5) as executor:
# Start the load operations and mark each future with its URL
future_to_url = {executor.submit(load_url, url, 60): url for url in URLS}
for future in concurrent.futures.as_completed(future_to_url):
url = future_to_url[future]
try:
data = future.result()
except Exception as exc:
print('%r generated an exception: %s' % (url, exc))
else:
print('%r page is %d bytes' % (url, len(data)))
ProcessPoolExecutor (source)
import concurrent.futures
import math
PRIMES = [
112272535095293,
112582705942171,
112272535095293,
115280095190773,
115797848077099,
1099726899285419]
def is_prime(n):
if n % 2 == 0:
return False
sqrt_n = int(math.floor(math.sqrt(n)))
for i in range(3, sqrt_n + 1, 2):
if n % i == 0:
return False
return True
def main():
with concurrent.futures.ProcessPoolExecutor() as executor:
for number, prime in zip(PRIMES, executor.map(is_prime, PRIMES)):
print('%d is prime: %s' % (number, prime))
if __name__ == '__main__':
main()
I saw a lot of examples here where no real work was being performed, and they were mostly CPU-bound. Here is an example of a CPU-bound task that computes all prime numbers between 10 million and 10.05 million. I have used all four methods here:
import math
import timeit
import threading
import multiprocessing
from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor
def time_stuff(fn):
"""
Measure time of execution of a function
"""
def wrapper(*args, **kwargs):
t0 = timeit.default_timer()
fn(*args, **kwargs)
t1 = timeit.default_timer()
print("{} seconds".format(t1 - t0))
return wrapper
def find_primes_in(nmin, nmax):
"""
Compute a list of prime numbers between the given minimum and maximum arguments
"""
primes = []
# Loop from minimum to maximum
for current in range(nmin, nmax + 1):
# Take the square root of the current number
sqrt_n = int(math.sqrt(current))
found = False
# Check if the any number from 2 to the square root + 1 divides the current numnber under consideration
for number in range(2, sqrt_n + 1):
# If divisible we have found a factor, hence this is not a prime number, lets move to the next one
if current % number == 0:
found = True
break
# If not divisible, add this number to the list of primes that we have found so far
if not found:
primes.append(current)
# I am merely printing the length of the array containing all the primes, but feel free to do what you want
print(len(primes))
#time_stuff
def sequential_prime_finder(nmin, nmax):
"""
Use the main process and main thread to compute everything in this case
"""
find_primes_in(nmin, nmax)
#time_stuff
def threading_prime_finder(nmin, nmax):
"""
If the minimum is 1000 and the maximum is 2000 and we have four workers,
1000 - 1250 to worker 1
1250 - 1500 to worker 2
1500 - 1750 to worker 3
1750 - 2000 to worker 4
so let’s split the minimum and maximum values according to the number of workers
"""
nrange = nmax - nmin
threads = []
for i in range(8):
start = int(nmin + i * nrange/8)
end = int(nmin + (i + 1) * nrange/8)
# Start the thread with the minimum and maximum split up to compute
# Parallel computation will not work here due to the GIL since this is a CPU-bound task
t = threading.Thread(target = find_primes_in, args = (start, end))
threads.append(t)
t.start()
# Don’t forget to wait for the threads to finish
for t in threads:
t.join()
#time_stuff
def processing_prime_finder(nmin, nmax):
"""
Split the minimum, maximum interval similar to the threading method above, but use processes this time
"""
nrange = nmax - nmin
processes = []
for i in range(8):
start = int(nmin + i * nrange/8)
end = int(nmin + (i + 1) * nrange/8)
p = multiprocessing.Process(target = find_primes_in, args = (start, end))
processes.append(p)
p.start()
for p in processes:
p.join()
#time_stuff
def thread_executor_prime_finder(nmin, nmax):
"""
Split the min max interval similar to the threading method, but use a thread pool executor this time.
This method is slightly faster than using pure threading as the pools manage threads more efficiently.
This method is still slow due to the GIL limitations since we are doing a CPU-bound task.
"""
nrange = nmax - nmin
with ThreadPoolExecutor(max_workers = 8) as e:
for i in range(8):
start = int(nmin + i * nrange/8)
end = int(nmin + (i + 1) * nrange/8)
e.submit(find_primes_in, start, end)
#time_stuff
def process_executor_prime_finder(nmin, nmax):
"""
Split the min max interval similar to the threading method, but use the process pool executor.
This is the fastest method recorded so far as it manages process efficiently + overcomes GIL limitations.
RECOMMENDED METHOD FOR CPU-BOUND TASKS
"""
nrange = nmax - nmin
with ProcessPoolExecutor(max_workers = 8) as e:
for i in range(8):
start = int(nmin + i * nrange/8)
end = int(nmin + (i + 1) * nrange/8)
e.submit(find_primes_in, start, end)
def main():
nmin = int(1e7)
nmax = int(1.05e7)
print("Sequential Prime Finder Starting")
sequential_prime_finder(nmin, nmax)
print("Threading Prime Finder Starting")
threading_prime_finder(nmin, nmax)
print("Processing Prime Finder Starting")
processing_prime_finder(nmin, nmax)
print("Thread Executor Prime Finder Starting")
thread_executor_prime_finder(nmin, nmax)
print("Process Executor Finder Starting")
process_executor_prime_finder(nmin, nmax)
if __name__ == "__main__":
main()
Here are the results on my Mac OS X four-core machine
Sequential Prime Finder Starting
9.708213827005238 seconds
Threading Prime Finder Starting
9.81836523200036 seconds
Processing Prime Finder Starting
3.2467174359990167 seconds
Thread Executor Prime Finder Starting
10.228896902000997 seconds
Process Executor Finder Starting
2.656402041000547 seconds
Using the blazing new concurrent.futures module
def sqr(val):
import time
time.sleep(0.1)
return val * val
def process_result(result):
print(result)
def process_these_asap(tasks):
import concurrent.futures
with concurrent.futures.ProcessPoolExecutor() as executor:
futures = []
for task in tasks:
futures.append(executor.submit(sqr, task))
for future in concurrent.futures.as_completed(futures):
process_result(future.result())
# Or instead of all this just do:
# results = executor.map(sqr, tasks)
# list(map(process_result, results))
def main():
tasks = list(range(10))
print('Processing {} tasks'.format(len(tasks)))
process_these_asap(tasks)
print('Done')
return 0
if __name__ == '__main__':
import sys
sys.exit(main())
The executor approach might seem familiar to all those who have gotten their hands dirty with Java before.
Also on a side note: To keep the universe sane, don't forget to close your pools/executors if you don't use with context (which is so awesome that it does it for you)
For me, the perfect example for threading is monitoring asynchronous events. Look at this code.
# thread_test.py
import threading
import time
class Monitor(threading.Thread):
def __init__(self, mon):
threading.Thread.__init__(self)
self.mon = mon
def run(self):
while True:
if self.mon[0] == 2:
print "Mon = 2"
self.mon[0] = 3;
You can play with this code by opening an IPython session and doing something like:
>>> from thread_test import Monitor
>>> a = [0]
>>> mon = Monitor(a)
>>> mon.start()
>>> a[0] = 2
Mon = 2
>>>a[0] = 2
Mon = 2
Wait a few minutes
>>> a[0] = 2
Mon = 2
Most documentation and tutorials use Python's Threading and Queue module, and they could seem overwhelming for beginners.
Perhaps consider the concurrent.futures.ThreadPoolExecutor module of Python 3.
Combined with with clause and list comprehension it could be a real charm.
from concurrent.futures import ThreadPoolExecutor, as_completed
def get_url(url):
# Your actual program here. Using threading.Lock() if necessary
return ""
# List of URLs to fetch
urls = ["url1", "url2"]
with ThreadPoolExecutor(max_workers = 5) as executor:
# Create threads
futures = {executor.submit(get_url, url) for url in urls}
# as_completed() gives you the threads once finished
for f in as_completed(futures):
# Get the results
rs = f.result()
With borrowing from this post we know about choosing between the multithreading, multiprocessing, and async/asyncio and their usage.
Python 3 has a new built-in library in order to make concurrency and parallelism — concurrent.futures
So I'll demonstrate through an experiment to run four tasks (i.e. .sleep() method) by Threading-Pool:
from concurrent.futures import ThreadPoolExecutor, as_completed
from time import sleep, time
def concurrent(max_worker):
futures = []
tic = time()
with ThreadPoolExecutor(max_workers=max_worker) as executor:
futures.append(executor.submit(sleep, 2)) # Two seconds sleep
futures.append(executor.submit(sleep, 1))
futures.append(executor.submit(sleep, 7))
futures.append(executor.submit(sleep, 3))
for future in as_completed(futures):
if future.result() is not None:
print(future.result())
print(f'Total elapsed time by {max_worker} workers:', time()-tic)
concurrent(5)
concurrent(4)
concurrent(3)
concurrent(2)
concurrent(1)
Output:
Total elapsed time by 5 workers: 7.007831811904907
Total elapsed time by 4 workers: 7.007944107055664
Total elapsed time by 3 workers: 7.003149509429932
Total elapsed time by 2 workers: 8.004627466201782
Total elapsed time by 1 workers: 13.013478994369507
[NOTE]:
As you can see in the above results, the best case was 3 workers for those four tasks.
If you have a process task instead of I/O bound or blocking (multiprocessing instead of threading) you can change the ThreadPoolExecutor to ProcessPoolExecutor.
I would like to contribute with a simple example and the explanations I've found useful when I had to tackle this problem myself.
In this answer you will find some information about Python's GIL (global interpreter lock) and a simple day-to-day example written using multiprocessing.dummy plus some simple benchmarks.
Global Interpreter Lock (GIL)
Python doesn't allow multi-threading in the truest sense of the word. It has a multi-threading package, but if you want to multi-thread to speed your code up, then it's usually not a good idea to use it.
Python has a construct called the global interpreter lock (GIL).
The GIL makes sure that only one of your 'threads' can execute at any one time. A thread acquires the GIL, does a little work, then passes the GIL onto the next thread.
This happens very quickly so to the human eye it may seem like your threads are executing in parallel, but they are really just taking turns using the same CPU core.
All this GIL passing adds overhead to execution. This means that if you want to make your code run faster then using the threading
package often isn't a good idea.
There are reasons to use Python's threading package. If you want to run some things simultaneously, and efficiency is not a concern,
then it's totally fine and convenient. Or if you are running code that needs to wait for something (like some I/O) then it could make a lot of sense. But the threading library won't let you use extra CPU cores.
Multi-threading can be outsourced to the operating system (by doing multi-processing), and some external application that calls your Python code (for example, Spark or Hadoop), or some code that your Python code calls (for example: you could have your Python code call a C function that does the expensive multi-threaded stuff).
Why This Matters
Because lots of people spend a lot of time trying to find bottlenecks in their fancy Python multi-threaded code before they learn what the GIL is.
Once this information is clear, here's my code:
#!/bin/python
from multiprocessing.dummy import Pool
from subprocess import PIPE,Popen
import time
import os
# In the variable pool_size we define the "parallelness".
# For CPU-bound tasks, it doesn't make sense to create more Pool processes
# than you have cores to run them on.
#
# On the other hand, if you are using I/O-bound tasks, it may make sense
# to create a quite a few more Pool processes than cores, since the processes
# will probably spend most their time blocked (waiting for I/O to complete).
pool_size = 8
def do_ping(ip):
if os.name == 'nt':
print ("Using Windows Ping to " + ip)
proc = Popen(['ping', ip], stdout=PIPE)
return proc.communicate()[0]
else:
print ("Using Linux / Unix Ping to " + ip)
proc = Popen(['ping', ip, '-c', '4'], stdout=PIPE)
return proc.communicate()[0]
os.system('cls' if os.name=='nt' else 'clear')
print ("Running using threads\n")
start_time = time.time()
pool = Pool(pool_size)
website_names = ["www.google.com","www.facebook.com","www.pinterest.com","www.microsoft.com"]
result = {}
for website_name in website_names:
result[website_name] = pool.apply_async(do_ping, args=(website_name,))
pool.close()
pool.join()
print ("\n--- Execution took {} seconds ---".format((time.time() - start_time)))
# Now we do the same without threading, just to compare time
print ("\nRunning NOT using threads\n")
start_time = time.time()
for website_name in website_names:
do_ping(website_name)
print ("\n--- Execution took {} seconds ---".format((time.time() - start_time)))
# Here's one way to print the final output from the threads
output = {}
for key, value in result.items():
output[key] = value.get()
print ("\nOutput aggregated in a Dictionary:")
print (output)
print ("\n")
print ("\nPretty printed output: ")
for key, value in output.items():
print (key + "\n")
print (value)
Here is the very simple example of CSV import using threading. (Library inclusion may differ for different purpose.)
Helper Functions:
from threading import Thread
from project import app
import csv
def import_handler(csv_file_name):
thr = Thread(target=dump_async_csv_data, args=[csv_file_name])
thr.start()
def dump_async_csv_data(csv_file_name):
with app.app_context():
with open(csv_file_name) as File:
reader = csv.DictReader(File)
for row in reader:
# DB operation/query
Driver Function:
import_handler(csv_file_name)
Here is multi threading with a simple example which will be helpful. You can run it and understand easily how multi threading is working in Python. I used a lock for preventing access to other threads until the previous threads finished their work. By the use of this line of code,
tLock = threading.BoundedSemaphore(value=4)
you can allow a number of processes at a time and keep hold to the rest of the threads which will run later or after finished previous processes.
import threading
import time
#tLock = threading.Lock()
tLock = threading.BoundedSemaphore(value=4)
def timer(name, delay, repeat):
print "\r\nTimer: ", name, " Started"
tLock.acquire()
print "\r\n", name, " has the acquired the lock"
while repeat > 0:
time.sleep(delay)
print "\r\n", name, ": ", str(time.ctime(time.time()))
repeat -= 1
print "\r\n", name, " is releaseing the lock"
tLock.release()
print "\r\nTimer: ", name, " Completed"
def Main():
t1 = threading.Thread(target=timer, args=("Timer1", 2, 5))
t2 = threading.Thread(target=timer, args=("Timer2", 3, 5))
t3 = threading.Thread(target=timer, args=("Timer3", 4, 5))
t4 = threading.Thread(target=timer, args=("Timer4", 5, 5))
t5 = threading.Thread(target=timer, args=("Timer5", 0.1, 5))
t1.start()
t2.start()
t3.start()
t4.start()
t5.start()
print "\r\nMain Complete"
if __name__ == "__main__":
Main()
None of the previous solutions actually used multiple cores on my GNU/Linux server (where I don't have administrator rights). They just ran on a single core.
I used the lower level os.fork interface to spawn multiple processes. This is the code that worked for me:
from os import fork
values = ['different', 'values', 'for', 'threads']
for i in range(len(values)):
p = fork()
if p == 0:
my_function(values[i])
break
As a python3 version of the second anwser:
import queue as Queue
import threading
import urllib.request
# Called by each thread
def get_url(q, url):
q.put(urllib.request.urlopen(url).read())
theurls = ["http://google.com", "http://yahoo.com", "http://www.python.org","https://wiki.python.org/moin/"]
q = Queue.Queue()
def thread_func():
for u in theurls:
t = threading.Thread(target=get_url, args = (q,u))
t.daemon = True
t.start()
s = q.get()
def non_thread_func():
for u in theurls:
get_url(q,u)
s = q.get()
And you can test it:
start = time.time()
thread_func()
end = time.time()
print(end - start)
start = time.time()
non_thread_func()
end = time.time()
print(end - start)
non_thread_func() should cost 4 times the time spent than thread_func()
import threading
import requests
def send():
r = requests.get('https://www.stackoverlow.com')
thread = []
t = threading.Thread(target=send())
thread.append(t)
t.start()
It's very easy to understand. Here are the two simple ways to do threading.
import time
from concurrent.futures import ThreadPoolExecutor, as_completed
import threading
def a(a=1, b=2):
print(a)
time.sleep(5)
print(b)
return a+b
def b(**kwargs):
if "a" in kwargs:
print("am b")
else:
print("nothing")
to_do=[]
executor = ThreadPoolExecutor(max_workers=4)
ex1=executor.submit(a)
to_do.append(ex1)
ex2=executor.submit(b, **{"a":1})
to_do.append(ex2)
for future in as_completed(to_do):
print("Future {} and Future Return is {}\n".format(future, future.result()))
print("threading")
to_do=[]
to_do.append(threading.Thread(target=a))
to_do.append(threading.Thread(target=b, kwargs={"a":1}))
for threads in to_do:
threads.start()
for threads in to_do:
threads.join()
This code below can run 10 threads concurrently printing the numbers from 0 to 99:
from threading import Thread
def test():
for i in range(0, 100):
print(i)
thread_list = []
for _ in range(0, 10):
thread = Thread(target=test)
thread_list.append(thread)
for thread in thread_list:
thread.start()
for thread in thread_list:
thread.join()
And, this code below is the shorthand for loop version of the above code running 10 threads concurrently printing the numbers from 0 to 99:
from threading import Thread
def test():
[print(i) for i in range(0, 100)]
thread_list = [Thread(target=test) for _ in range(0, 10)]
[thread.start() for thread in thread_list]
[thread.join() for thread in thread_list]
This is the result below:
...
99
83
97
84
98
99
85
86
87
88
...
The easiest way of using threading/multiprocessing is to use more high level libraries like autothread.
import autothread
from time import sleep as heavyworkload
#autothread.multithreaded() # <-- This is all you need to add
def example(x: int, y: int):
heavyworkload(1)
return x*y
Now, you can feed your functions lists of ints. Autothread will handle everything for you and just give you the results computed in parallel.
result = example([1, 2, 3, 4, 5], 10)

Counting the dead pixel by using multiprocessing in Python

I have no idea how counting the dead pixels with multiprocessing works I got it so far without multiprocessing and analyzing the 10 pictures we need to analyze takes around 7 minutes...
import random
import time
from multiprocessing import Process, Queue, current_process, freeze_support
from PIL import Image, ImageDraw
image1 = Image.open('MA_HA1_drawing_0.png')
image2 = Image.open('MA_HA1_drawing_1.png')
image2 = Image.open('MA_HA1_drawing_2.png')
image3 = Image.open('MA_HA1_drawing_3.png')
image4 = Image.open('MA_HA1_drawing_4.png')
image5 = Image.open('MA_HA1_drawing_5.png')
image6 = Image.open('MA_HA1_drawing_6.png')
image7 = Image.open('MA_HA1_drawing_7.png')
image8 = Image.open('MA_HA1_drawing_8.png')
image9 = Image.open('MA_HA1_drawing_9.png')
def analyze_picture(image):
time.sleep(0.5*random.random())
counter = 0
for x in range(616,6446):
for y in range(756,3712):
r,g,b = image.getpixel((x,y))
if r != 1 and g != 1 and b != 1:
counter += 1
return counter
def test():
NUMBER_OF_PROCESSES = 4
TASKS1 = [(analyze_picture(image1))]
TASKS2 = [(analyze_picture(image2))]
TASKS3 = [(analyze_picture(image2))]
TASKS4 = [(analyze_picture(image3))]
TASKS5 = [(analyze_picture(image4))]
TASKS6 = [(analyze_picture(image5))]
TASKS7 = [(analyze_picture(image6))]
TASKS8 = [(analyze_picture(image7))]
TASKS9 = [(analyze_picture(image8))]
TASKS10 = [(analyze_picture(image9))]
print TASKS1
if __name__ == '__main__':
freeze_support()
test()
and they gave us some functions to understand multiprocessing and use it for our task but I don't understand them and have no idea how to use them.
def worker(input, output):
for func, args in iter(input.get, 'STOP'):
result = calculate(func, args)
output.put(result)
def calculate(func, args):
result = func(*args)
return '%s says that %s%s = %s' % \
(current_process().name, func.__name__, args, result)
def mul(a, b):
time.sleep(0.5*random.random())
return a * b
def plus(a, b):
time.sleep(0.5*random.random())
return a + b
# Create queues
task_queue = Queue()
done_queue = Queue()
# Submit tasks
for task in TASKS1:
task_queue.put(task)
# Start worker processes
for i in range(NUMBER_OF_PROCESSES):
Process(target=worker, args=(task_queue, done_queue)).start()
print i
# Get and print results
print 'Unordered results:'
for i in range(len(TASKS1)):
print '\t', done_queue.get()
# Add more tasks using `put()`
for task in TASKS2:
task_queue.put(task)
# Get and print some more results
for i in range(len(TASKS2)):
print '\t', done_queue.get()
# Tell child processes to stop
for i in range(NUMBER_OF_PROCESSES):
task_queue.put('STOP')
print 'process ', i, ' is stopped'
EDIT: New Code
import random
import time
from multiprocessing import Process, Queue, current_process, freeze_support
from PIL import Image, ImageDraw
def worker(input, output):
for func, args in iter(input.get, 'STOP'):
result = calculate(func, args)
output.put(result)
def calculate(func, args):
result = func(args)
return '%s says that %s%s has %s dead pixels\n' % \
(current_process().name, func.__name__, args, result)
def analyze_picture(image_name):
t1 = time.clock()
image = Image.open(image_name)
time.sleep(0.5*random.random())
counter = 0
for x in range(616,6446):
for y in range(756,3712):
r,g,b = image.getpixel((x,y))
if r != 1 and g != 1 and b != 1:
counter += 1
t2 = time.clock()
dt = t2 - t1
print '\tThe process takes ',dt,' seconds.\n Result:\n'
return counter
def test():
NUMBER_OF_PROCESSES = 4
TASKS1 = [(analyze_picture, image_names[i]) for i in range(10)]
print TASKS1
# Create queues
task_queue = Queue()
done_queue = Queue()
# Submit tasks
for task in TASKS1:
task_queue.put(task)
# Start worker processes
for i in range(NUMBER_OF_PROCESSES):
Process(target=worker, args=(task_queue, done_queue)).start()
print i
# Get and print results
print 'Unordered results:'
for i in range(len(TASKS1)):
print '\t', done_queue.get()
# Tell child processes to stop
for i in range(NUMBER_OF_PROCESSES):
task_queue.put('STOP')
print 'process ', i, ' is stopped'
if __name__ == '__main__':
image_names =[('MA_HA1_drawing_'+str(i)+'.png') for i in range(10)]
freeze_support()
test()
The idea behind multiprocessing:
Create several workers which can be distributed to different cores for execution in parallel.
For multiprocessing these workers are processes with their separate memory space (in contrast to threads).
Due to the separate memory space they can not communicate (receive a task and send its result) via the memory. As a result, queues for inter-process communications are needed.
Now, the tasks are distributed to the workers via the queues.
Finally, the results are collected which the workers have sent via the queues.
If it is mandatory to use the posted code you can do it as follows:
Create queues
Start worker processes
Submit tasks
It is very import to submit the tasks after the workers were created. It is possible, that the buffer of the queue gets full and blocks until something was taken from it, but as long as there is no worker yet, nothing is taken from the queue --> DEADLOCK.
Since you want to do all the images in parallel, your TASKS1 (recognize the plural) has to be [(analyze_picture, (analyze_picture(image1),), (analyze_picture, (analyze_picture(image2),), ...] (worker expects a tuple of the function and the arguments as a tuple itself).
Get and print results
Tell child processes to stop
Probably that's what you was asking about.
After all, there are three further aspects of improving performance (and code readability):
Inter-process communication is quite expensive. Hence, you should minimize the data which is transfered from and to the workers.
In your case, this means to pass the image name only instead of the entire image. Additionally, this leads to a parallel read in of all the images because the workers read the images.
All the worker stuff is already implemented in multiprocessing.Pool, which reduces the lines of code for the multiprocessing to two:
pool = multiprocessing.Pool(processes=NUMBER_OF_PROCESSES)
result = pool.map(analyze_picture, [image1, image2, ...])
Last but not least, iterating pixel by pixel is rather slow. With NumPy (or the higher level SciPy) you can accelerate it a lot.
Finally, your script could look like the following and will be much faster than 7 minutes:
import multiprocessing as mp
import numpy as np
from scipy import misc
def analyze_picture(imagename):
image = misc.imread(imagename) # image[y, x, r/g/b]
return len(np.argwhere( (a[756:,616:,0]!=1) & (a[756:,616:,1]!=1) & (a[756:,616:,2]!=1) ))
def main():
pool = mp.Pool() # default: number of logical cores
result = pool.map(analyze_picture, ( "MA_HA1_drawing_{}.png".format(i)
for i in range(10) ))
print(result)
if __name__ == '__main__':
mp.freeze_support()
main()
I am not sure about how your images look like (the {r,g,b}!=1 is strange) but in the reference of scipy.misc.imread you'll find the appropriate mode for you image.

Multiprocessing Class in Subprocess

I want to use python's multiprocessing module in a class, which itself uses subprocesses to not block the main call.
The minimal example looks like this:
import multiprocessing as mp
class mpo():
def __init__(self):
cpu = mp.cpu_count()
self.Pool = mp.Pool(processes = 2)
self.alive = True
self.p = mp.Process(target = self.sub,args=())
def worker():
print 'Alive'
def sub(self):
print self.alive
for i in range(2):
print i
self.Pool.apply_async(self.worker, args=())
print 'done'
self.Pool.close()
# self.Pool.join()
I commented the last line out, as it raises an assertion Error (can only join a child process).
When I do:
m =mpo()
m.p.start()
The output is
True
0
1
done
My main question is, why the print statement in the worker thread never is reached?
Update:
The updated code looks like this.
import multiprocessing as mp
class mpo():
def __init__(self):
cpu = mp.cpu_count()
self.alive = True
self.p = mp.Process(target = self.sub,args=())
self.result=[]
def worker(self):
self.result.append(1)
print 'Alive'
def sub(self):
print self.alive
Pool = mp.Pool(processes = 2)
for i in range(2):
print i
Pool.apply_async(self.worker, args=())
print 'done'
Pool.close()
Pool.join()
The pool now doesn't have to be inherited as it is created in the subprocess. Instead of the print statement the result is appended to the calling object and the pool is properly joined. Nevertheless, there is no result showing up.
so I think this may correspond to a simple example of what you are looking for:
import multiprocessing as mp
def worker(arg):
#print 'Alive'+str(arg)
return "Alive and finished {0}".format(arg)
class mpo():
def __init__(self):
cpu = mp.cpu_count()
self.alive = True
self.pool = mp.Pool(processes = 2)
def sub(self,arguments):
self.results=self.pool.map_async(worker, arguments)
return self.results
if __name__=="__main__":
s=mpo()
s.sub(range(10))
print s.results.get()
Additionally you can call
self.results.ready()
to find out whether the processes have finished their work. You do not have to put this inside of another process because the map_async call does not block the rest of your program.
EDIT:
Concerning your comment, I do not really see the value of putting the calculation in a separate process, because the function is already running in separate processes (in the pool). You only add complexity by nesting it in another subprocess, but it is possible:
import multiprocessing as mp
def worker(arg):
#print 'Alive'+str(arg)
return "Alive and finished {0}".format(arg)
class mpo():
def __init__(self):
cpu = mp.cpu_count()
self.alive = True
self.pool = mp.Pool(processes = 2)
def sub(self,arguments):
self.results=self.pool.map_async(worker, arguments)
return self.results
def run_calculation(q):
s=mpo()
results=s.sub(range(10))
q.put(results.get())
queue=mp.Queue()
proc=mp.Process(target=run_calculation,args=(queue,))
proc.start()
proc.join()
queue.get()

Categories