Writing a parallel loop

Question

I am trying to run a parallel loop on a simple example.
What am I doing wrong?

from joblib import Parallel, delayed  
import multiprocessing

def processInput(i):  
        return i * i

if __name__ == '__main__':

    # what are your inputs, and what operation do you want to 
    # perform on each input. For example...
    inputs = range(1000000)      

    num_cores = multiprocessing.cpu_count()

    results = Parallel(n_jobs=4)(delayed(processInput)(i) for i in inputs) 

    print(results)

The problem with the code is that when executed under Windows environments in Python 3, it opens num_cores instances of python to execute the parallel jobs but only one is active. This should not be the case since the activity of the processor should be 100% instead of 14% (under i7 - 8 logic cores).

Why are the extra instances not doing anything?

Answer 1

Continuing on your request to provide a working multiprocessing code, I suggest that you use pool_map (if the delayed functionality is not important), I'll give you an example, if your'e working on python3 its worth to mention you can use starmap. Also worth mentioning that you can use map_sync/starmap_async if the order of the returned results does not have to correspond to the order of inputs.

import multiprocessing as mp

def processInput(i):
        return i * i

if __name__ == '__main__':

    # what are your inputs, and what operation do you want to
    # perform on each input. For example...
    inputs = range(1000000)
    #  removing processes argument makes the code run on all available cores
    pool = mp.Pool(processes=4)
    results = pool.map(processInput, inputs)
    print(results)

Answer 2

On Windows, the multiprocessing module uses the 'spawn' method to start up multiple python interpreter processes. This is relatively slow. Parallel tries to be smart about running the code. In particular, it tries to adjust batch sizes so a batch takes about half a second to execute. (See the batch_size argument at https://pythonhosted.org/joblib/parallel.html)

Your processInput() function runs so fast that Parallel determines that it is faster to run the jobs serially on one processor than to spin up multiple python interpreters and run the code in parallel.

If you want to force your example to run on multiple cores, try setting batch_size to 1000 or making processInput() more complicated so it takes longer to execute.

Edit: Working example on windows that shows multiple processes in use (I'm using windows 7):

from joblib import Parallel, delayed
from os import getpid

def modfib(n):
    # print the process id to see that multiple processes are used, and
    # re-used during the job.
    if n%400 == 0:
        print(getpid(), n)  

    # fibonacci sequence mod 1000000
    a,b = 0,1
    for i in range(n):
        a,b = b,(a+b)%1000000
    return b

if __name__ == "__main__":
    Parallel(n_jobs=-1, verbose=5)(delayed(modfib)(j) for j in range(1000, 4000))