Related
I am pretty new to multiprocessing in python and trying to achieve something which should be a rather common thing to do. But I cannot find an easy way when searching the web.
I want to put data in a queue and then make this queue available to different consumer functions. Of course when getting an element from the queue, all consumer functions should get the same element. The following example should make clear what I want to achieve:
from multiprocessing import Process, Queue
def producer(q):
for i in range(10):
q.put(i)
q.put(None)
def consumer1(q):
while True:
data = q.get()
if data is None:
break
print(data)
def consumer2(q):
while True:
data = q.get()
if data is None:
break
print(data)
def main():
q = Queue()
p1 = Process(target=producer, args=(q,))
p2 = Process(target=consumer1, args=(q,))
p3 = Process(target=consumer2, args=(q,))
p1.start()
p2.start()
p3.start()
p1.join()
p2.join()
p3.join()
if __name__ == '__main__':
main()
Since the script is not terminating and I only get the print output of one function I guess this is not the way to do it. I think sharing a queue implies some things to consider? It works fine when using only one consumer function.
Appreciate the help!
If the values you are storing can be represented by one of the fundamental data types defined in the ctypes module, then the following could work. Here we are implementing a "queue" that can hold int values or None:
from multiprocessing import Process, Condition
import ctypes
from multiprocessing.sharedctypes import RawArray, RawValue
from threading import local
import time
my_local = local()
my_local.current = 0
class StructuredInt(ctypes.Structure):
"""
This class is necessary because we want to be able to store in the RawArray
either an int or None, which requires using ctypes.c_void_p as the array type.
But, infortunately, ctypes.c_void_p(0) is interpreted as None.
So we need a way to represent 0. Field value 'value' is the
actual int value being stored and we use an arbitrarty 'ptr'
field value that will not be interpreted as None.
To store a None value, we set 'ptr' to ctypes.c_void_p(None) and field
'value' is irrelevant.
To store an integer. we set 'ptr' to ctypes.c_void_p(1) and field
'value' has the actual value.
"""
_fields_ = [('ptr', ctypes.c_void_p), ('value', ctypes.c_int)]
class MultiIntQueue:
"""
An integer queue that can be processed by multiple threads where each thread
can retrieve all the values added to the queue.
:param maxsize: The maximum queue capacity (defaults to 20 if specified as None)
:type maxsize: int
"""
def __init__(self, maxsize=None):
if maxsize is None:
maxsize = 20
self.maxsize = maxsize
self.q = RawArray(StructuredInt, maxsize)
self.condition = Condition()
self.size = RawValue(ctypes.c_int, 0)
def get(self):
with self.condition:
while my_local.current >= self.size.value:
self.condition.wait()
i = self.q[my_local.current]
my_local.current += 1
return None if i.ptr is None else i.value
def put(self, i):
assert 0 <= self.size.value < self.maxsize
with self.condition:
self.q[self.size.value] = (ctypes.c_void_p(None), 0) if i is None else (ctypes.c_void_p(1), i)
self.size.value += 1
self.condition.notify_all()
def producer(q):
for i in range(10):
q.put(i)
time.sleep(.3) # simulate processing
q.put(None)
def consumer1(q):
while True:
data = q.get()
if data is None:
break
time.sleep(.1) # simulate processing
print('Consumer 1:', data)
def consumer2(q):
while True:
data = q.get()
if data is None:
break
time.sleep(.1) # simulate processing
print('Consumer 2:', data)
def main():
q = MultiIntQueue()
p1 = Process(target=producer, args=(q,))
p2 = Process(target=consumer1, args=(q,))
p3 = Process(target=consumer2, args=(q,))
p1.start()
p2.start()
p3.start()
p1.join()
p2.join()
p3.join()
if __name__ == '__main__':
main()
Prints:
Consumer 1: 0
Consumer 2: 0
Consumer 2: 1
Consumer 1: 1
Consumer 2: 2
Consumer 1: 2
Consumer 2: 3
Consumer 1: 3
Consumer 2: 4
Consumer 1: 4
Consumer 1: 5
Consumer 2: 5
Consumer 1: 6
Consumer 2: 6
Consumer 1: 7
Consumer 2: 7
Consumer 2: 8
Consumer 1: 8
Consumer 1: 9
Consumer 2: 9
Your question exemplifies the misunderstanding
"all consumer functions should get the same element"
That's just not how queues work. Queues are automatically managed (there's quite a lot under the hood) such if one item is put in, only one item can be taken out. That item is not duplicated to all consumers. It seems like you actually need two separate queues to guarantee that each consumer gets each input without competing against the other consumer:
from multiprocessing import Process, Queue
def producer(q1, q2):
for i in range(10):
q1.put(i)
q2.put(i)
q1.put(None)
q2.put(None)
def consumer1(q):
while True:
data = q.get()
if data is None:
break
print(data)
def consumer2(q):
while True:
data = q.get()
if data is None:
break
print(data)
def main():
q1 = Queue()
q2 = Queue()
p1 = Process(target=producer, args=(q1, q2))
p2 = Process(target=consumer1, args=(q1,))
p3 = Process(target=consumer2, args=(q2,))
p1.start()
p2.start()
p3.start()
p1.join()
p2.join()
p3.join()
if __name__ == '__main__':
main()
The end goal is to execute a method in background, but not in parallel : when multiple objects are calling this method, each should wait for their turn to proceed. To achieve running in background, I have to run the method in a subprocess (not a thread), and I need to start it using spawn (not fork). To prevent parallel executions, the obvious solution is to have a global lock shared between processes.
When processes are forked, which is the default on Unix, it is easy to achieve, as highlighted in both of the following codes.
We can share it as a class variable :
import multiprocessing as mp
from time import sleep
class OneAtATime:
l = mp.Lock()
def f(self):
with self.l:
sleep(1)
print("Hello")
if __name__ == "__main__":
a = OneAtATime()
b = OneAtATime()
p1 = mp.Process(target = a.f)
p2 = mp.Process(target = b.f)
p1.start()
p2.start()
Or we can pass it to the method :
import multiprocessing as mp
from time import sleep
class OneAtATime:
def f(self, l):
with l:
sleep(1)
print("Hello")
if __name__ == "__main__":
a = OneAtATime()
b = OneAtATime()
m = mp.Manager()
l = mp.Lock()
p1 = mp.Process(target = a.f, args = (l,))
p2 = mp.Process(target = b.f, args = (l,))
p1.start()
p2.start()
Both of these codes have the appropriate behaviour of printing "hello" at one second of interval.
However, when changing the start method to 'spawn', they become broken.
The first one (1) prints both "hello"s at the same time. This is because the internal state of a class is not pickled, so they do not have the same lock.
The second one (2) fails with FileNotFoundError at runtime. I think it has to do with the fact that locks cannot be pickled : see Python sharing a lock between processes.
In this answer, two fixes are suggested (side note : I cannot use a pool because I want to randomly create an arbitrary number of processes).
I haven't found a way to adapt the second fix, but I tried to implement the first one :
import multiprocessing as mp
from time import sleep
if __name__ == "__main__":
mp.set_start_method('spawn')
class OneAtATime:
def f(self, l):
with l:
sleep(1)
print("Hello")
if __name__ == "__main__":
a = OneAtATime()
b = OneAtATime()
m = mp.Manager()
l = m.Lock()
p1 = mp.Process(target = a.f, args = (l,))
p2 = mp.Process(target = b.f, args = (l,))
p1.start()
p2.start()
This fails with AttributeError and FileNotFoundError (3). In fact it also fails (BrokenPipe) when the fork method is used (4).
What is the proper way of sharing a lock between spawned processes ?
A quick explanation of the four fails I numbered would be nice, too.
I'm running Python 3.6 under Archlinux.
Congratulations, you got yourself 90% of the way there. The last step is actually not very hard to do.
Yes, your final code block fails with an AttributeError, but what specifically is the error? "Can't get attribute 'OneAtATime' on ". This is very similar to a problem you've already encountered - it's not pickling the class OneAtATime.
I made the following change and it worked as you'd like:
file ooat.py:
from time import sleep
class OneAtATime:
def f(self, l):
with l:
sleep(1)
print("Hello")
interactive shell:
import multiprocessing as mp
from oaat import OneAtATime
if __name__ == "__main__":
mp.set_start_method('spawn')
a = OneAtATime()
b = OneAtATime()
m = mp.Manager()
l = m.Lock()
p1 = mp.Process(target = a.f, args = (l,))
p2 = mp.Process(target = b.f, args = (l,))
p1.start()
p2.start()
You may notice, I didn't really do anything - just split your code into two separate files. Try it out, you'll see it works fine. (At least, it did for me, using python 3.5 on ubuntu.)
The last code snippet works, provided the script does not exit prematurely. Joining processes is enough :
import multiprocessing as mp
from time import sleep
class OneAtATime:
def f(self, l):
with l:
sleep(1)
print("Hello")
if __name__ == "__main__":
mp.set_start_method('spawn')
a = OneAtATime()
b = OneAtATime()
m = mp.Manager()
l = m.Lock()
p1 = mp.Process(target = a.f, args = (l,))
p2 = mp.Process(target = b.f, args = (l,))
p1.start()
p2.start()
p1.join()
p2.join()
More info on the error it was causing here https://stackoverflow.com/a/25456494/8194503.
In the example code below, I'd like to get the return value of the function worker. How can I go about doing this? Where is this value stored?
Example Code:
import multiprocessing
def worker(procnum):
'''worker function'''
print str(procnum) + ' represent!'
return procnum
if __name__ == '__main__':
jobs = []
for i in range(5):
p = multiprocessing.Process(target=worker, args=(i,))
jobs.append(p)
p.start()
for proc in jobs:
proc.join()
print jobs
Output:
0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
[<Process(Process-1, stopped)>, <Process(Process-2, stopped)>, <Process(Process-3, stopped)>, <Process(Process-4, stopped)>, <Process(Process-5, stopped)>]
I can't seem to find the relevant attribute in the objects stored in jobs.
Use shared variable to communicate. For example like this:
import multiprocessing
def worker(procnum, return_dict):
"""worker function"""
print(str(procnum) + " represent!")
return_dict[procnum] = procnum
if __name__ == "__main__":
manager = multiprocessing.Manager()
return_dict = manager.dict()
jobs = []
for i in range(5):
p = multiprocessing.Process(target=worker, args=(i, return_dict))
jobs.append(p)
p.start()
for proc in jobs:
proc.join()
print(return_dict.values())
I think the approach suggested by #sega_sai is the better one. But it really needs a code example, so here goes:
import multiprocessing
from os import getpid
def worker(procnum):
print('I am number %d in process %d' % (procnum, getpid()))
return getpid()
if __name__ == '__main__':
pool = multiprocessing.Pool(processes = 3)
print(pool.map(worker, range(5)))
Which will print the return values:
I am number 0 in process 19139
I am number 1 in process 19138
I am number 2 in process 19140
I am number 3 in process 19139
I am number 4 in process 19140
[19139, 19138, 19140, 19139, 19140]
If you are familiar with map (the Python 2 built-in) this should not be too challenging. Otherwise have a look at sega_Sai's link.
Note how little code is needed. (Also note how processes are re-used).
For anyone else who is seeking how to get a value from a Process using Queue:
import multiprocessing
ret = {'foo': False}
def worker(queue):
ret = queue.get()
ret['foo'] = True
queue.put(ret)
if __name__ == '__main__':
queue = multiprocessing.Queue()
queue.put(ret)
p = multiprocessing.Process(target=worker, args=(queue,))
p.start()
p.join()
print(queue.get()) # Prints {"foo": True}
Note that in Windows or Jupyter Notebook, with multithreading you have to save this as a file and execute the file. If you do it in a command prompt you will see an error like this:
AttributeError: Can't get attribute 'worker' on <module '__main__' (built-in)>
For some reason, I couldn't find a general example of how to do this with Queue anywhere (even Python's doc examples don't spawn multiple processes), so here's what I got working after like 10 tries:
from multiprocessing import Process, Queue
def add_helper(queue, arg1, arg2): # the func called in child processes
ret = arg1 + arg2
queue.put(ret)
def multi_add(): # spawns child processes
q = Queue()
processes = []
rets = []
for _ in range(0, 100):
p = Process(target=add_helper, args=(q, 1, 2))
processes.append(p)
p.start()
for p in processes:
ret = q.get() # will block
rets.append(ret)
for p in processes:
p.join()
return rets
Queue is a blocking, thread-safe queue that you can use to store the return values from the child processes. So you have to pass the queue to each process. Something less obvious here is that you have to get() from the queue before you join the Processes or else the queue fills up and blocks everything.
Update for those who are object-oriented (tested in Python 3.4):
from multiprocessing import Process, Queue
class Multiprocessor():
def __init__(self):
self.processes = []
self.queue = Queue()
#staticmethod
def _wrapper(func, queue, args, kwargs):
ret = func(*args, **kwargs)
queue.put(ret)
def run(self, func, *args, **kwargs):
args2 = [func, self.queue, args, kwargs]
p = Process(target=self._wrapper, args=args2)
self.processes.append(p)
p.start()
def wait(self):
rets = []
for p in self.processes:
ret = self.queue.get()
rets.append(ret)
for p in self.processes:
p.join()
return rets
# tester
if __name__ == "__main__":
mp = Multiprocessor()
num_proc = 64
for _ in range(num_proc): # queue up multiple tasks running `sum`
mp.run(sum, [1, 2, 3, 4, 5])
ret = mp.wait() # get all results
print(ret)
assert len(ret) == num_proc and all(r == 15 for r in ret)
This example shows how to use a list of multiprocessing.Pipe instances to return strings from an arbitrary number of processes:
import multiprocessing
def worker(procnum, send_end):
'''worker function'''
result = str(procnum) + ' represent!'
print result
send_end.send(result)
def main():
jobs = []
pipe_list = []
for i in range(5):
recv_end, send_end = multiprocessing.Pipe(False)
p = multiprocessing.Process(target=worker, args=(i, send_end))
jobs.append(p)
pipe_list.append(recv_end)
p.start()
for proc in jobs:
proc.join()
result_list = [x.recv() for x in pipe_list]
print result_list
if __name__ == '__main__':
main()
Output:
0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
['0 represent!', '1 represent!', '2 represent!', '3 represent!', '4 represent!']
This solution uses fewer resources than a multiprocessing.Queue which uses
a Pipe
at least one Lock
a buffer
a thread
or a multiprocessing.SimpleQueue which uses
a Pipe
at least one Lock
It is very instructive to look at the source for each of these types.
It seems that you should use the multiprocessing.Pool class instead and use the methods .apply() .apply_async(), map()
http://docs.python.org/library/multiprocessing.html?highlight=pool#multiprocessing.pool.AsyncResult
You can use the exit built-in to set the exit code of a process. It can be obtained from the exitcode attribute of the process:
import multiprocessing
def worker(procnum):
print str(procnum) + ' represent!'
exit(procnum)
if __name__ == '__main__':
jobs = []
for i in range(5):
p = multiprocessing.Process(target=worker, args=(i,))
jobs.append(p)
p.start()
result = []
for proc in jobs:
proc.join()
result.append(proc.exitcode)
print result
Output:
0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
[0, 1, 2, 3, 4]
The pebble package has a nice abstraction leveraging multiprocessing.Pipe which makes this quite straightforward:
from pebble import concurrent
#concurrent.process
def function(arg, kwarg=0):
return arg + kwarg
future = function(1, kwarg=1)
print(future.result())
Example from: https://pythonhosted.org/Pebble/#concurrent-decorators
Thought I'd simplify the simplest examples copied from above, working for me on Py3.6. Simplest is multiprocessing.Pool:
import multiprocessing
import time
def worker(x):
time.sleep(1)
return x
pool = multiprocessing.Pool()
print(pool.map(worker, range(10)))
You can set the number of processes in the pool with, e.g., Pool(processes=5). However it defaults to CPU count, so leave it blank for CPU-bound tasks. (I/O-bound tasks often suit threads anyway, as the threads are mostly waiting so can share a CPU core.) Pool also applies chunking optimization.
(Note that the worker method cannot be nested within a method. I initially defined my worker method inside the method that makes the call to pool.map, to keep it all self-contained, but then the processes couldn't import it, and threw "AttributeError: Can't pickle local object outer_method..inner_method". More here. It can be inside a class.)
(Appreciate the original question specified printing 'represent!' rather than time.sleep(), but without it I thought some code was running concurrently when it wasn't.)
Py3's ProcessPoolExecutor is also two lines (.map returns a generator so you need the list()):
from concurrent.futures import ProcessPoolExecutor
with ProcessPoolExecutor() as executor:
print(list(executor.map(worker, range(10))))
With plain Processes:
import multiprocessing
import time
def worker(x, queue):
time.sleep(1)
queue.put(x)
queue = multiprocessing.SimpleQueue()
tasks = range(10)
for task in tasks:
multiprocessing.Process(target=worker, args=(task, queue,)).start()
for _ in tasks:
print(queue.get())
Use SimpleQueue if all you need is put and get. The first loop starts all the processes, before the second makes the blocking queue.get calls. I don't think there's any reason to call p.join() too.
If you are using Python 3, you can use concurrent.futures.ProcessPoolExecutor as a convenient abstraction:
from concurrent.futures import ProcessPoolExecutor
def worker(procnum):
'''worker function'''
print(str(procnum) + ' represent!')
return procnum
if __name__ == '__main__':
with ProcessPoolExecutor() as executor:
print(list(executor.map(worker, range(5))))
Output:
0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
[0, 1, 2, 3, 4]
A simple solution:
import multiprocessing
output=[]
data = range(0,10)
def f(x):
return x**2
def handler():
p = multiprocessing.Pool(64)
r=p.map(f, data)
return r
if __name__ == '__main__':
output.append(handler())
print(output[0])
Output:
[0, 1, 4, 9, 16, 25, 36, 49, 64, 81]
You can use ProcessPoolExecutor to get a return value from a function as shown below:
from concurrent.futures import ProcessPoolExecutor
def test(num1, num2):
return num1 + num2
with ProcessPoolExecutor() as executor:
feature = executor.submit(test, 2, 3)
print(feature.result()) # 5
I modified vartec's answer a bit since I needed to get the error codes from the function. (Thanks vertec!!! its an awesome trick)
This can also be done with a manager.list but I think is better to have it in a dict and store a list within it. That way, way we keep the function and the results since we can't be sure of the order in which the list will be populated.
from multiprocessing import Process
import time
import datetime
import multiprocessing
def func1(fn, m_list):
print 'func1: starting'
time.sleep(1)
m_list[fn] = "this is the first function"
print 'func1: finishing'
# return "func1" # no need for return since Multiprocess doesnt return it =(
def func2(fn, m_list):
print 'func2: starting'
time.sleep(3)
m_list[fn] = "this is function 2"
print 'func2: finishing'
# return "func2"
def func3(fn, m_list):
print 'func3: starting'
time.sleep(9)
# if fail wont join the rest because it never populate the dict
# or do a try/except to get something in return.
raise ValueError("failed here")
# if we want to get the error in the manager dict we can catch the error
try:
raise ValueError("failed here")
m_list[fn] = "this is third"
except:
m_list[fn] = "this is third and it fail horrible"
# print 'func3: finishing'
# return "func3"
def runInParallel(*fns): # * is to accept any input in list
start_time = datetime.datetime.now()
proc = []
manager = multiprocessing.Manager()
m_list = manager.dict()
for fn in fns:
# print fn
# print dir(fn)
p = Process(target=fn, name=fn.func_name, args=(fn, m_list))
p.start()
proc.append(p)
for p in proc:
p.join() # 5 is the time out
print datetime.datetime.now() - start_time
return m_list, proc
if __name__ == '__main__':
manager, proc = runInParallel(func1, func2, func3)
# print dir(proc[0])
# print proc[0]._name
# print proc[0].name
# print proc[0].exitcode
# here you can check what did fail
for i in proc:
print i.name, i.exitcode # name was set up in the Process line 53
# here will only show the function that worked and where able to populate the
# manager dict
for i, j in manager.items():
print dir(i) # things you can do to the function
print i, j
I have a 3 processes running in one script. Process 1 passes data to Process 2, and then Process 2 passes data to Process 3. When I put data to queue2, error occurs that "Global name "queue2" is not defined", I am stuck on this error now...
if __name__ == '__main__':
queue1 = mp.Queue()
queue2 = mp.Queue()
p1 = mp.Process(target=f2, args=(queue1,))
p1.start()
p2 = mp.Process(target=f3, args=(queue2,))
p2.start()
f1()
def f1():
# do something to a get x
queue1.put(x)
def f2(q):
a = q.get()
# do something to a, to produce b
queue2.put(b) # error happens here: Global name "queue2" is not defined
def f3(q):
c = q.get()
# keeping processing c...
Just as you passed queue1 to f2, you also need to pass queue2.
You can declare the queues as global:
def f2(q):
global queue2
a = q.get()
queue2.put(b)
This works :
import multiprocessing as mp
queue1 = mp.Queue()
queue2 = mp.Queue()
def f1(q):
x = 5
# do something to a get x
q.put(x)
def f2(in_queue, out_queue):
a = in_queue.get()
b = a + 2
# do something to a, to produce b
out_queue.put(b)
def f3(q):
c = q.get()
print c
f1(queue1)
p1 = mp.Process(target=f2, args=(queue1, queue2))
p1.start()
p2 = mp.Process(target=f3, args=(queue2,))
p2.start()
Your code doesn't return the error you seem to have, it returns "f2 not defined" since you when you spawn the process p1, f2 is not a defined variable yet. The rule when you fork is that at creation time your processes must see the variables they use, i.e. they must be in the current scope.
To put it clearly, at spawning process time you inherit the current namespace from the parent process.
In the example code below, I'd like to get the return value of the function worker. How can I go about doing this? Where is this value stored?
Example Code:
import multiprocessing
def worker(procnum):
'''worker function'''
print str(procnum) + ' represent!'
return procnum
if __name__ == '__main__':
jobs = []
for i in range(5):
p = multiprocessing.Process(target=worker, args=(i,))
jobs.append(p)
p.start()
for proc in jobs:
proc.join()
print jobs
Output:
0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
[<Process(Process-1, stopped)>, <Process(Process-2, stopped)>, <Process(Process-3, stopped)>, <Process(Process-4, stopped)>, <Process(Process-5, stopped)>]
I can't seem to find the relevant attribute in the objects stored in jobs.
Use shared variable to communicate. For example like this:
import multiprocessing
def worker(procnum, return_dict):
"""worker function"""
print(str(procnum) + " represent!")
return_dict[procnum] = procnum
if __name__ == "__main__":
manager = multiprocessing.Manager()
return_dict = manager.dict()
jobs = []
for i in range(5):
p = multiprocessing.Process(target=worker, args=(i, return_dict))
jobs.append(p)
p.start()
for proc in jobs:
proc.join()
print(return_dict.values())
I think the approach suggested by #sega_sai is the better one. But it really needs a code example, so here goes:
import multiprocessing
from os import getpid
def worker(procnum):
print('I am number %d in process %d' % (procnum, getpid()))
return getpid()
if __name__ == '__main__':
pool = multiprocessing.Pool(processes = 3)
print(pool.map(worker, range(5)))
Which will print the return values:
I am number 0 in process 19139
I am number 1 in process 19138
I am number 2 in process 19140
I am number 3 in process 19139
I am number 4 in process 19140
[19139, 19138, 19140, 19139, 19140]
If you are familiar with map (the Python 2 built-in) this should not be too challenging. Otherwise have a look at sega_Sai's link.
Note how little code is needed. (Also note how processes are re-used).
For anyone else who is seeking how to get a value from a Process using Queue:
import multiprocessing
ret = {'foo': False}
def worker(queue):
ret = queue.get()
ret['foo'] = True
queue.put(ret)
if __name__ == '__main__':
queue = multiprocessing.Queue()
queue.put(ret)
p = multiprocessing.Process(target=worker, args=(queue,))
p.start()
p.join()
print(queue.get()) # Prints {"foo": True}
Note that in Windows or Jupyter Notebook, with multithreading you have to save this as a file and execute the file. If you do it in a command prompt you will see an error like this:
AttributeError: Can't get attribute 'worker' on <module '__main__' (built-in)>
For some reason, I couldn't find a general example of how to do this with Queue anywhere (even Python's doc examples don't spawn multiple processes), so here's what I got working after like 10 tries:
from multiprocessing import Process, Queue
def add_helper(queue, arg1, arg2): # the func called in child processes
ret = arg1 + arg2
queue.put(ret)
def multi_add(): # spawns child processes
q = Queue()
processes = []
rets = []
for _ in range(0, 100):
p = Process(target=add_helper, args=(q, 1, 2))
processes.append(p)
p.start()
for p in processes:
ret = q.get() # will block
rets.append(ret)
for p in processes:
p.join()
return rets
Queue is a blocking, thread-safe queue that you can use to store the return values from the child processes. So you have to pass the queue to each process. Something less obvious here is that you have to get() from the queue before you join the Processes or else the queue fills up and blocks everything.
Update for those who are object-oriented (tested in Python 3.4):
from multiprocessing import Process, Queue
class Multiprocessor():
def __init__(self):
self.processes = []
self.queue = Queue()
#staticmethod
def _wrapper(func, queue, args, kwargs):
ret = func(*args, **kwargs)
queue.put(ret)
def run(self, func, *args, **kwargs):
args2 = [func, self.queue, args, kwargs]
p = Process(target=self._wrapper, args=args2)
self.processes.append(p)
p.start()
def wait(self):
rets = []
for p in self.processes:
ret = self.queue.get()
rets.append(ret)
for p in self.processes:
p.join()
return rets
# tester
if __name__ == "__main__":
mp = Multiprocessor()
num_proc = 64
for _ in range(num_proc): # queue up multiple tasks running `sum`
mp.run(sum, [1, 2, 3, 4, 5])
ret = mp.wait() # get all results
print(ret)
assert len(ret) == num_proc and all(r == 15 for r in ret)
This example shows how to use a list of multiprocessing.Pipe instances to return strings from an arbitrary number of processes:
import multiprocessing
def worker(procnum, send_end):
'''worker function'''
result = str(procnum) + ' represent!'
print result
send_end.send(result)
def main():
jobs = []
pipe_list = []
for i in range(5):
recv_end, send_end = multiprocessing.Pipe(False)
p = multiprocessing.Process(target=worker, args=(i, send_end))
jobs.append(p)
pipe_list.append(recv_end)
p.start()
for proc in jobs:
proc.join()
result_list = [x.recv() for x in pipe_list]
print result_list
if __name__ == '__main__':
main()
Output:
0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
['0 represent!', '1 represent!', '2 represent!', '3 represent!', '4 represent!']
This solution uses fewer resources than a multiprocessing.Queue which uses
a Pipe
at least one Lock
a buffer
a thread
or a multiprocessing.SimpleQueue which uses
a Pipe
at least one Lock
It is very instructive to look at the source for each of these types.
It seems that you should use the multiprocessing.Pool class instead and use the methods .apply() .apply_async(), map()
http://docs.python.org/library/multiprocessing.html?highlight=pool#multiprocessing.pool.AsyncResult
You can use the exit built-in to set the exit code of a process. It can be obtained from the exitcode attribute of the process:
import multiprocessing
def worker(procnum):
print str(procnum) + ' represent!'
exit(procnum)
if __name__ == '__main__':
jobs = []
for i in range(5):
p = multiprocessing.Process(target=worker, args=(i,))
jobs.append(p)
p.start()
result = []
for proc in jobs:
proc.join()
result.append(proc.exitcode)
print result
Output:
0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
[0, 1, 2, 3, 4]
The pebble package has a nice abstraction leveraging multiprocessing.Pipe which makes this quite straightforward:
from pebble import concurrent
#concurrent.process
def function(arg, kwarg=0):
return arg + kwarg
future = function(1, kwarg=1)
print(future.result())
Example from: https://pythonhosted.org/Pebble/#concurrent-decorators
Thought I'd simplify the simplest examples copied from above, working for me on Py3.6. Simplest is multiprocessing.Pool:
import multiprocessing
import time
def worker(x):
time.sleep(1)
return x
pool = multiprocessing.Pool()
print(pool.map(worker, range(10)))
You can set the number of processes in the pool with, e.g., Pool(processes=5). However it defaults to CPU count, so leave it blank for CPU-bound tasks. (I/O-bound tasks often suit threads anyway, as the threads are mostly waiting so can share a CPU core.) Pool also applies chunking optimization.
(Note that the worker method cannot be nested within a method. I initially defined my worker method inside the method that makes the call to pool.map, to keep it all self-contained, but then the processes couldn't import it, and threw "AttributeError: Can't pickle local object outer_method..inner_method". More here. It can be inside a class.)
(Appreciate the original question specified printing 'represent!' rather than time.sleep(), but without it I thought some code was running concurrently when it wasn't.)
Py3's ProcessPoolExecutor is also two lines (.map returns a generator so you need the list()):
from concurrent.futures import ProcessPoolExecutor
with ProcessPoolExecutor() as executor:
print(list(executor.map(worker, range(10))))
With plain Processes:
import multiprocessing
import time
def worker(x, queue):
time.sleep(1)
queue.put(x)
queue = multiprocessing.SimpleQueue()
tasks = range(10)
for task in tasks:
multiprocessing.Process(target=worker, args=(task, queue,)).start()
for _ in tasks:
print(queue.get())
Use SimpleQueue if all you need is put and get. The first loop starts all the processes, before the second makes the blocking queue.get calls. I don't think there's any reason to call p.join() too.
If you are using Python 3, you can use concurrent.futures.ProcessPoolExecutor as a convenient abstraction:
from concurrent.futures import ProcessPoolExecutor
def worker(procnum):
'''worker function'''
print(str(procnum) + ' represent!')
return procnum
if __name__ == '__main__':
with ProcessPoolExecutor() as executor:
print(list(executor.map(worker, range(5))))
Output:
0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
[0, 1, 2, 3, 4]
A simple solution:
import multiprocessing
output=[]
data = range(0,10)
def f(x):
return x**2
def handler():
p = multiprocessing.Pool(64)
r=p.map(f, data)
return r
if __name__ == '__main__':
output.append(handler())
print(output[0])
Output:
[0, 1, 4, 9, 16, 25, 36, 49, 64, 81]
You can use ProcessPoolExecutor to get a return value from a function as shown below:
from concurrent.futures import ProcessPoolExecutor
def test(num1, num2):
return num1 + num2
with ProcessPoolExecutor() as executor:
feature = executor.submit(test, 2, 3)
print(feature.result()) # 5
I modified vartec's answer a bit since I needed to get the error codes from the function. (Thanks vertec!!! its an awesome trick)
This can also be done with a manager.list but I think is better to have it in a dict and store a list within it. That way, way we keep the function and the results since we can't be sure of the order in which the list will be populated.
from multiprocessing import Process
import time
import datetime
import multiprocessing
def func1(fn, m_list):
print 'func1: starting'
time.sleep(1)
m_list[fn] = "this is the first function"
print 'func1: finishing'
# return "func1" # no need for return since Multiprocess doesnt return it =(
def func2(fn, m_list):
print 'func2: starting'
time.sleep(3)
m_list[fn] = "this is function 2"
print 'func2: finishing'
# return "func2"
def func3(fn, m_list):
print 'func3: starting'
time.sleep(9)
# if fail wont join the rest because it never populate the dict
# or do a try/except to get something in return.
raise ValueError("failed here")
# if we want to get the error in the manager dict we can catch the error
try:
raise ValueError("failed here")
m_list[fn] = "this is third"
except:
m_list[fn] = "this is third and it fail horrible"
# print 'func3: finishing'
# return "func3"
def runInParallel(*fns): # * is to accept any input in list
start_time = datetime.datetime.now()
proc = []
manager = multiprocessing.Manager()
m_list = manager.dict()
for fn in fns:
# print fn
# print dir(fn)
p = Process(target=fn, name=fn.func_name, args=(fn, m_list))
p.start()
proc.append(p)
for p in proc:
p.join() # 5 is the time out
print datetime.datetime.now() - start_time
return m_list, proc
if __name__ == '__main__':
manager, proc = runInParallel(func1, func2, func3)
# print dir(proc[0])
# print proc[0]._name
# print proc[0].name
# print proc[0].exitcode
# here you can check what did fail
for i in proc:
print i.name, i.exitcode # name was set up in the Process line 53
# here will only show the function that worked and where able to populate the
# manager dict
for i, j in manager.items():
print dir(i) # things you can do to the function
print i, j