How to use python multiprocessing Pool.map within loop

Question

I am running a simulation using Runge-Kutta. At every time step two FFT of two independent variables are necessary which can be parallelized. I implemented the code like this:

from multiprocessing import Pool
import numpy as np

pool = Pool(processes=2)    # I like to calculate only 2 FFTs parallel 
                            # in every time step, therefor 2 processes

def Splitter(args):
    '''I have to pass 2 arguments'''
    return makeSomething(*args):

def makeSomething(a,b):
    '''dummy function instead of the one with the FFT'''
    return a*b

def RungeK():
    # ...
    # a lot of code which create the vectors A and B and calculates 
    # one Kunge-Kutta step for them 
    # ...

    n = 20                         # Just something for the example
    A = np.arange(50000)
    B = np.ones_like(A)

    for i in xrange(n):                  # loop over the time steps
        A *= np.mean(B)*B - A
        B *= np.sqrt(A)
        results = pool.map(Splitter,[(A,3),(B,2)])
        A = results[0]
        B = results[1]

    print np.mean(A)                                 # Some output
    print np.max(B)

if __name__== '__main__':
    RungeK()

Unfortunately python generates a unlimited number of processes after reaching the loop. Before it seems that only two processes are running. Also my memory fills up. Adding a

pool.close()
pool.join()

behind the loop does not solve my problem, and to put it inside the loop makes no sense for me. Hope you can help.

Answer 1

Move the creation of the pool into the RungeK function;

def RungeK():
    # ...
    # a lot of code which create the vectors A and B and calculates
    # one Kunge-Kutta step for them
    # ...

    pool = Pool(processes=2)
    n = 20                         # Just something for the example
    A = np.arange(50000)
    B = np.ones_like(A)

    for i in xrange(n):  # loop over the time steps
        A *= np.mean(B)*B - A
        B *= np.sqrt(A)
        results = pool.map(Splitter, [(A, 3), (B, 2)])
        A = results[0]
        B = results[1]
    pool.close()
    print np.mean(A)  # Some output
    print np.max(B)

Alternatively, put it in the main block.

This is probably a side effect of how multiprocessing works. E.g. on MS windows, you need to be able to import the main module without side effects (like creating new processes).