Multiprocess pool initialization with sequential initializer argument

Question

I have some code like the following:

import multiprocessing as mp

connection: module.Connection

def client_id():
    for i in range(mp.cpu_count*2):
        yield i

def initproc(host: str, port: int, client_id: int):
    global connection
    connection.connect(host, port, client_id)

def main():
    host = "something"
    port = 12345
    mp.get_context("spawn").Pool(processes=mp.cpu_count()*2,
                                 initializer=initproc,
                                 initargs=(host, port, client_id())) as p:
        res = p.starmap(processing_function, arg_list)

for the purposes of the question processing_function and arg_list are not relevant.

The issue is that I get an error with this:

    ForkingPickler(file, protocol).dump(obj)
TypeError: cannot pickle 'generator' object

Is there any way to create an initialize a process in the pool in such a way that on of the arguments to initialize it would be the next number in a sequence?

P.S. In the code as written it may be possible to initialize all connection objects outside of the initializer function but in my particular instance it is not. I need to pass arguments for connection into the initializer.

There are several things you should look into, because they likely do not do what you think they do. ``global`` and ``multiprocessing`` do not mix – you will have a *new* ``connection`` in each process. ``initproc`` says it expects ``client_id: int``, but you are passing ``client_id: Iterable[int]`` instead (on account of ``client_id()`` being a generator). Finally, what do you intend the trailing ``as p`` to do? — MisterMiyagi, Aug 24 '20 at 17:05
`as p` allows me to address the pool object as a variable `p`. `global` is necessary because variable `connection` otherwise will be local within the function and won't be accessible from data processing functions. `Iterable` or not I am not even getting to the point of where `initproc` is getting the value to process — Karlson, Aug 24 '20 at 17:23

score 2 · Accepted Answer · answered Aug 24 '20 at 16:58

A simple solution for your case would be to use the sequential number of the child-process which is contained in the Process.name. You could extract it with...

mp.current_process().name.split('-')[1]

If you need more control over where the sequence starts, you could use multiprocessing.Value as a counter from which workers get their unique number.

import multiprocessing as mp
import time


def init_p(client_id):
    with client_id.get_lock():
        globals()['client_id'] = client_id.value
        print(f"{mp.current_process().name},"
              f" {mp.current_process().name.split('-')[1]},"  # alternative
              f" client_id:{globals()['client_id']}")
        client_id.value += 1


if __name__ == "__main__":

    ctx = mp.get_context("spawn")
    client_ids = ctx.Value('i', 0)

    with ctx.Pool(
            processes=4,
            initializer=init_p,
            initargs=(client_ids,)
    ) as pool:

        time.sleep(3)

Output:

SpawnPoolWorker-2, 2, client_id:0
SpawnPoolWorker-3, 3, client_id:1
SpawnPoolWorker-1, 1, client_id:2
SpawnPoolWorker-4, 4, client_id:3

Process finished with exit code 0

Multiprocess pool initialization with sequential initializer argument

1 Answers1