paraproc 0.1.3

Easy parallel processing in Python

Overview

Paraproc is a simple library that helps you easily parallelize your computation (over independent chunks of data) across multiple processes in Python, especially when you want to mix callings to external command line programs and hand brew Python functions together in your data processing pipeline.

Under the hood, it combines subprocess and multiprocessing, and uses a process pool to schedule the jobs. It also provides a numpy.ndarray interface to access shared-memory across multiple processes.

Paraproc supports both Python 2 and 3, with numpy as the only external dependency. It is contained in only one Python file, so it can be easily copied into your project. (The copyright and license notice must be retained.)

Code snippets that demonstrate the basic usage of the library can be found later in this documentation, and in the demo_*.py files.

Bugs can be reported to https://github.com/herrlich10/paraproc. The code can also be found there.

Quick starts

Execute commands in parallel

You can run both Python codes and command line programs in parallel:

import os
import paraproc
def my_job():
    print(os.getpid())

pc = paraproc.PooledCaller()
for k in range(5):
    pc.check_call(my_job)
for k in range(5):
    pc.check_call('echo $$', shell=True) # For linux/mac
pc.wait()

The pc.check_call() method will return immediatedly. The actual execution of the queued commands are delayed until you call pc.wait().

Use shared-memory

You can load large data in shared-memory, and read or write them as a normal numpy array from multiple processes:

import numpy as np
import paraproc
def slow_operation(k, x):
    x.acquire()
    x[:100000,:] += 1 # Write access
    res = np.mean(x) # Read access
    x.release()
    print('#{0}: mean = {1}'.format(k, res))

a = paraproc.SharedMemoryArray.from_array(np.random.rand(1000000,500)) # About 4 GB
pc = paraproc.PooledCaller()
for k in range(pc.pool_size):
    pc.check_call(slow_operation, k, a)
pc.wait()

The data in a is shared in memory across all children processes and never copied even with write accesses.