Custom IterDataPipe for multiprocessing

Hello everyone.

I’m not sure if i understand MultiProcessingReadingService correctly.
Below is a small test with a custom IterDataPipe which should be spawned to 4 workers.

Without MultiProcessingReadingService it takes ~ 110ms on my PC.
With MultiProcessingReadingService it takes minutes!
I mean it’s not a little bit slower, its like… crazy :slight_smile:

My versions: torchdata 0.6.0, Python 3.9.16
I’m testing on a Win11 machine.

Am i doing it wrong?
Is the slow down caused by IPC and multiprocessing should only be used for very heavy-CPU related stuff?
Is it Windows related?

Thank you for any help!

from io import IOBase
import time
from typing import IO, Tuple
from torchdata.dataloader2 import DataLoader2, MultiProcessingReadingService
from torchdata.datapipes.iter import IterDataPipe

class TestIterDataPipe(IterDataPipe[Tuple[str, IOBase]]):
    def __init__(self, source_datapipe: IterDataPipe[Tuple[str, IO]]) -> None:
        self.source_datapipe: IterDataPipe[Tuple[str, IO]] = source_datapipe

    def __iter__(self):
        for i in range(100000):
            yield 'foo', i

if __name__ == '__main__':

    tStep1 = time.perf_counter() * 1000.0
    tStart = tStep1

    pipe = TestIterDataPipe(None)
    mpReader = MultiProcessingReadingService(num_workers=4)
    dl2 = DataLoader2(pipe, reading_service=mpReader)

    tStep2 = time.perf_counter() * 1000.0
    tDiff1 = tStep2 - tStep1
    tStep1 = tStep2

    ndx = 0
    for dt1, dt2 in dl2:
        ndx += 1
        if not ndx % 123:
            tStep2 = time.perf_counter() * 1000.0
            tDiff2 = tStep2 - tStep1
            print(f'{ndx}: {dt2}, {tDiff2} ms')

    tStep2 = time.perf_counter() * 1000.0
    tDiff2 = tStep2 - tStep1
    tStep1 = tStep2

    print(f'pipeline: {tDiff1} ms, loop: {tDiff2} ms')
    print(f'total: {tStep2 - tStart} ms')


Small update: looks like a Windows (native) problem.
I setup WSL2 and recreated my conda-env in WSL2 (Ubuntu).
It works much better now.
I worked with Threads before on Windows, but not multiprocessing, and never had any problems.
So i’m not sure if it’s related to torchdata oooor python itself?