parfor 2024.11.1

A package to mimic the use of parfor as done in Matlab.

Parfor

Used to parallelize for-loops using parfor in Matlab? This package allows you to do the same in python. Take any normal serial but parallelizable for-loop and execute it in parallel using easy syntax. Don't worry about the technical details of using the multiprocessing module, race conditions, queues, parfor handles all that.

Tested on linux, Windows and OSX with python 3.10 and 3.12.

Why is parfor better than just using multiprocessing?

• Easy to use
• Using dill instead of pickle: a lot more objects can be used when parallelizing
• Progress bars are built-in
• Automatically use multithreading instead of multiprocessing when the GIL is disabled

How it works

This depends on whether the GIL is currently disabled or not. Disabling the GIL in Python is currently an experimental feature in Python3.13, and not the standard.

Python with GIL enabled

The work you want parfor to do is divided over a number of processes. These processes are started by parfor and put together in a pool. This pool is reused when you want parfor to do more work, or shut down when no new work arrives within 10 minutes.

A handle to each bit of work is put in a queue from which the workers take work. The objects needed to do the work are stored in a memory manager in serialized form (using dill) and the manager hands out an object to a worker when the worker is requesting it. The manager deletes objects automatically when they're not needed anymore.

When the work is done the result is sent back for collection in the main process.

Python with GIL disabled

The work you want parfor to do is given to a new thread. These threads are started by parfor and put together in a pool. The threads and pool are not reused and closed automatically when done.

When the work is done a message is sent to the main thread to update the status of the pool.

Installation

pip install parfor

Usage

Parfor decorates a functions and returns the result of that function evaluated in parallel for each iteration of an iterator.

Requires

tqdm, dill

Limitations

If you're using Python with the GIL enabaled, then objects passed to the pool need to be dillable (dill needs to serialize them). Generators and SwigPyObjects are examples of objects that cannot be used. They can be used however, for the iterator argument when using parfor, but its iterations need to be dillable. You might be able to make objects dillable anyhow using dill.register or with __reduce__, __getstate__, etc.

Arguments

To functions parfor.parfor, parfor.pmap and parfor.gmap.

Required:

fun:      function taking arguments: iteration from  iterable, other arguments defined in args & kwargs
iterable: iterable or iterator from which an item is given to fun as a first argument

Optional:

args:   tuple with other unnamed arguments to fun
kwargs: dict with other named arguments to fun
total:  give the length of the iterator in cases where len(iterator) results in an error
desc:   string with description of the progress bar
bar:    bool enable progress bar,
            or a callback function taking the number of passed iterations as an argument
serial: execute in series instead of parallel if True, None (default): let pmap decide
length: deprecated alias for total
n_processes: number of processes to use,
    the parallel pool will be restarted if the current pool does not have the right number of processes
yield_ordered: return the result in the same order as the iterable
yield_index: return the index of the result too
**bar_kwargs: keyword arguments for tqdm.tqdm

Return

list with results from applying the function 'fun' to each iteration of the iterable / iterator

Examples

Normal serial for loop

<<
from time import sleep

a = 3
fun = []
for i in range(10):
    sleep(1)
    fun.append(a * i ** 2)
print(fun)

>> [0, 3, 12, 27, 48, 75, 108, 147, 192, 243]

Using parfor to parallelize

<<
from time import sleep
from parfor import parfor
@parfor(range(10), (3,))
def fun(i, a):
    sleep(1)
    return a * i ** 2
print(fun)

>> [0, 3, 12, 27, 48, 75, 108, 147, 192, 243]

<<
@parfor(range(10), (3,), bar=False)
def fun(i, a):
    sleep(1)
    return a * i ** 2
print(fun)

>> [0, 3, 12, 27, 48, 75, 108, 147, 192, 243]

Using parfor in a script/module/.py-file

Parfor should never be executed during the import phase of a .py-file. To prevent that from happening use the if __name__ == '__main__': structure:

<<
from time import sleep
from parfor import parfor

if __name__ == '__main__':
    @parfor(range(10), (3,))
    def fun(i, a):
        sleep(1)
        return a * i ** 2
    print(fun)

>> [0, 3, 12, 27, 48, 75, 108, 147, 192, 243]    

or:

<<
from time import sleep
from parfor import parfor

def my_fun(*args, **kwargs):
    @parfor(range(10), (3,))
    def fun(i, a):
        sleep(1)
        return a * i ** 2
    return fun

if __name__ == '__main__':
    print(my_fun())

>> [0, 3, 12, 27, 48, 75, 108, 147, 192, 243]

If you hate decorators not returning a function

pmap maps an iterator to a function like map does, but in parallel

<<
from parfor import pmap
from time import sleep
def fun(i, a):
    sleep(1)
    return a * i ** 2
print(pmap(fun, range(10), (3,)))

>> [0, 3, 12, 27, 48, 75, 108, 147, 192, 243]     

Using generators

If iterators like lists and tuples are too big for the memory, use generators instead. Since generators don't have a predefined length, give parfor the length (total) as an argument (optional).

<<
import numpy as np
c = (im for im in imagereader)
@parfor(c, total=len(imagereader))
def fun(im):
    return np.mean(im)
    
>> [list with means of the images]

Extra's

pmap

The function parfor decorates, it's used similarly to map, it returns a list with the results.

gmap

Same as pmap, but returns a generator. Useful to use the result as soon as it's generated.

Chunks

Split a long iterator in bite-sized chunks to parallelize

ParPool

More low-level accessibility to parallel execution. Submit tasks and request the result at any time, (although to avoid breaking causality, submit first, then request), use different functions and function arguments for different tasks.