Is shared readonly data copied to different proces

问题:

The piece of code that I have looks some what like this:

glbl_array = # a 3 Gb array

def my_func( args, def_param = glbl_array):
    #do stuff on args and def_param

if __name__ == \'__main__\':
  pool = Pool(processes=4)
  pool.map(my_func, range(1000))

Is there a way to make sure (or encourage) that the different processes does not get a copy of glbl_array but shares it. If there is no way to stop the copy I will go with a memmapped array, but my access patterns are not very regular, so I expect memmapped arrays to be slower. The above seemed like the first thing to try. This is on Linux. I just wanted some advice from Stackoverflow and do not want to annoy the sysadmin. Do you think it will help if the the second parameter is a genuine immutable object like glbl_array.tostring().

回答1:

You can use the shared memory stuff from multiprocessing together with Numpy fairly easily:

import multiprocessing
import ctypes
import numpy as np

shared_array_base = multiprocessing.Array(ctypes.c_double, 10*10)
shared_array = np.ctypeslib.as_array(shared_array_base.get_obj())
shared_array = shared_array.reshape(10, 10)

#-- edited 2015-05-01: the assert check below checks the wrong thing
#   with recent versions of Numpy/multiprocessing. That no copy is made
#   is indicated by the fact that the program prints the output shown below.
## No copy was made
##assert shared_array.base.base is shared_array_base.get_obj()

# Parallel processing
def my_func(i, def_param=shared_array):
    shared_array[i,:] = i

if __name__ == \'__main__\':
    pool = multiprocessing.Pool(processes=4)
    pool.map(my_func, range(10))

    print shared_array

which prints

[[ 0.  0.  0.  0.  0.  0.  0.  0.  0.  0.]
 [ 1.  1.  1.  1.  1.  1.  1.  1.  1.  1.]
 [ 2.  2.  2.  2.  2.  2.  2.  2.  2.  2.]
 [ 3.  3.  3.  3.  3.  3.  3.  3.  3.  3.]
 [ 4.  4.  4.  4.  4.  4.  4.  4.  4.  4.]
 [ 5.  5.  5.  5.  5.  5.  5.  5.  5.  5.]
 [ 6.  6.  6.  6.  6.  6.  6.  6.  6.  6.]
 [ 7.  7.  7.  7.  7.  7.  7.  7.  7.  7.]
 [ 8.  8.  8.  8.  8.  8.  8.  8.  8.  8.]
 [ 9.  9.  9.  9.  9.  9.  9.  9.  9.  9.]]

However, Linux has copy-on-write semantics on fork(), so even without using multiprocessing.Array, the data will not be copied unless it is written to.

回答2:

The following code works on Win7 and Mac (maybe on linux, but not tested).

import multiprocessing
import ctypes
import numpy as np

#-- edited 2015-05-01: the assert check below checks the wrong thing
#   with recent versions of Numpy/multiprocessing. That no copy is made
#   is indicated by the fact that the program prints the output shown below.
## No copy was made
##assert shared_array.base.base is shared_array_base.get_obj()

shared_array = None

def init(shared_array_base):
    global shared_array
    shared_array = np.ctypeslib.as_array(shared_array_base.get_obj())
    shared_array = shared_array.reshape(10, 10)

# Parallel processing
def my_func(i):
    shared_array[i, :] = i

if __name__ == \'__main__\':
    shared_array_base = multiprocessing.Array(ctypes.c_double, 10*10)

    pool = multiprocessing.Pool(processes=4, initializer=init, initargs=(shared_array_base,))
    pool.map(my_func, range(10))

    shared_array = np.ctypeslib.as_array(shared_array_base.get_obj())
    shared_array = shared_array.reshape(10, 10)
    print shared_array

回答3:

For those stuck using Windows, which does not support fork() (unless using CygWin), pv\'s answer does not work. Globals are not made available to child processes.

Instead, you must pass the shared memory during the initializer of the Pool/Process as such:

#! /usr/bin/python

import time

from multiprocessing import Process, Queue, Array

def f(q,a):
    m = q.get()
    print m
    print a[0], a[1], a[2]
    m = q.get()
    print m
    print a[0], a[1], a[2]

if __name__ == \'__main__\':
    a = Array(\'B\', (1, 2, 3), lock=False)
    q = Queue()
    p = Process(target=f, args=(q,a))
    p.start()
    q.put([1, 2, 3])
    time.sleep(1)
    a[0:3] = (4, 5, 6)
    q.put([4, 5, 6])
    p.join()

(it\'s not numpy and it\'s not good code but it illustrates the point ;-)

回答4:

If you are looking for an option that works efficiently on Windows, and works well for irregular access patterns, branching, and other scenarios where you might need to analyze different matrices based on a combination of a shared-memory matrix and process-local data, the mathDict toolkit in the ParallelRegression package was designed to handle this exact situation.