Share Large, Read-Only Numpy Array Between Multiprocessing Processes

@Velimir Mlaker gave a great answer. I thought I could add some bits of comments and a tiny example.

(I couldn't find much documentation on sharedmem - these are the results of my own experiments.)

  1. Do you need to pass the handles when the subprocess is starting, or after it has started? If it's just the former, you can just use the target and args arguments for Process. This is potentially better than using a global variable.
  2. From the discussion page you linked, it appears that support for 64-bit Linux was added to sharedmem a while back, so it could be a non-issue.
  3. I don't know about this one.
  4. No. Refer to example below.

Example

#!/usr/bin/env python
from multiprocessing import Process
import sharedmem
import numpy

def do_work(data, start):
    data[start] = 0;

def split_work(num):
    n = 20
    width  = n/num
    shared = sharedmem.empty(n)
    shared[:] = numpy.random.rand(1, n)[0]
    print "values are %s" % shared

    processes = [Process(target=do_work, args=(shared, i*width)) for i in xrange(num)]
    for p in processes:
        p.start()
    for p in processes:
        p.join()

    print "values are %s" % shared
    print "type is %s" % type(shared[0])

if __name__ == '__main__':
    split_work(4)

Output

values are [ 0.81397784  0.59667692  0.10761908  0.6736734   0.46349645  0.98340718
  0.44056863  0.10701816  0.67167752  0.29158274  0.22242552  0.14273156
  0.34912309  0.43812636  0.58484507  0.81697513  0.57758441  0.4284959
  0.7292129   0.06063283]
values are [ 0.          0.59667692  0.10761908  0.6736734   0.46349645  0.
  0.44056863  0.10701816  0.67167752  0.29158274  0.          0.14273156
  0.34912309  0.43812636  0.58484507  0.          0.57758441  0.4284959
  0.7292129   0.06063283]
type is <type 'numpy.float64'>

This related question might be useful.


If you are on Linux (or any POSIX-compliant system), you can define this array as a global variable. multiprocessing is using fork() on Linux when it starts a new child process. A newly spawned child process automatically shares the memory with its parent as long as it does not change it (copy-on-write mechanism).

Since you are saying "I don't need any kind of locks, since the array (actually a matrix) will be read-only" taking advantage of this behavior would be a very simple and yet extremely efficient approach: all child processes will access the same data in physical memory when reading this large numpy array.

Don't hand your array to the Process() constructor, this will instruct multiprocessing to pickle the data to the child, which would be extremely inefficient or impossible in your case. On Linux, right after fork() the child is an exact copy of the parent using the same physical memory, so all you need to do is making sure that the Python variable 'containing' the matrix is accessible from within the target function that you hand over to Process(). This you can typically achieve with a 'global' variable.

Example code:

from multiprocessing import Process
from numpy import random


global_array = random.random(10**4)


def child():
    print sum(global_array)


def main():
    processes = [Process(target=child) for _ in xrange(10)]
    for p in processes:
        p.start()
    for p in processes:
        p.join()


if __name__ == "__main__":
    main()

On Windows -- which does not support fork() -- multiprocessing is using the win32 API call CreateProcess. It creates an entirely new process from any given executable. That's why on Windows one is required to pickle data to the child if one needs data that has been created during runtime of the parent.


You may be interested in a tiny piece of code I wrote: github.com/vmlaker/benchmark-sharedmem

The only file of interest is main.py. It's a benchmark of numpy-sharedmem -- the code simply passes arrays (either numpy or sharedmem) to spawned processes, via Pipe. The workers just call sum() on the data. I was only interested in comparing the data communication times between the two implementations.

I also wrote another, more complex code: github.com/vmlaker/sherlock.

Here I use the numpy-sharedmem module for real-time image processing with OpenCV -- the images are NumPy arrays, as per OpenCV's newer cv2 API. The images, actually references thereof, are shared between processes via the dictionary object created from multiprocessing.Manager (as opposed to using Queue or Pipe.) I'm getting great performance improvements when compared with using plain NumPy arrays.

Pipe vs. Queue:

In my experience, IPC with Pipe is faster than Queue. And that makes sense, since Queue adds locking to make it safe for multiple producers/consumers. Pipe doesn't. But if you only have two processes talking back-and-forth, it's safe to use Pipe, or, as the docs read:

... there is no risk of corruption from processes using different ends of the pipe at the same time.

sharedmem safety:

The main issue with sharedmem module is the possibility of memory leak upon ungraceful program exit. This is described in a lengthy discussion here. Although on Apr 10, 2011 Sturla mentions a fix to memory leak, I have still experienced leaks since then, using both repos, Sturla Molden's own on GitHub (github.com/sturlamolden/sharedmem-numpy) and Chris Lee-Messer's on Bitbucket (bitbucket.org/cleemesser/numpy-sharedmem).