Inconsistency between macos and linux for neuron under multiprocessing

[AurelienJaquier]

Context

Overview of the issue

When in a subprocess, neuron dt is the same as in the main process in linux, while it is reset to its default value in macos

Expected result/behavior

I would expect neuron to behave the same on both macos and linux. I would also like for neuron dt not to be reset to its default value when in a subprocess.

NEURON setup

• Version: [8.2.1]
• Installation method [pip]
• OS + Version: [macos Monterey 12.6 / Red Hat Enterprise Linux Server release 7.9]

Minimal working example - MWE

• python script:

import neuron
import multiprocessing


def show_dt():
    print(neuron.h.dt)


if __name__ == '__main__':
    neuron.h.dt = 0.1
    show_dt()

    proc = multiprocessing.Process(target=show_dt)
    proc.start()
    proc.join()
    print("end")

Logs

output on linux:

0.1
0.1
end

ouptut on macos:

0.1
0.025
end

[pramodk]

thanks @AurelienJaquier for the report with a reproducer!

I believe this is related to the differences in multi-processing module itself on Linux vs macOS:

• fork mode (default for Unix): all resources of the parent are inherited by the child process
• spawn mode (default for MacOS): child process will only inherit those resources necessary to run the process objects run() method

Fork is the default on Linux, while Windows and MacOS use spawn by default.

If you explicitly set the start method (e.g. multiprocessing.set_start_method('fork')) then you should see the same behaviour.

[AurelienJaquier]

Thank you very much for your detailed response @pramodk !
Using the fork start method indeed solved my problem.

I'll close this issue.

[anilbey]

    fork mode (default for Unix): all resources of the parent are inherited by the child process
    spawn mode (default for MacOS): child process will only inherit those resources necessary to run the process objects run() method

Hi, I think the underlying issue is the following: NEURON does not inform multiprocessing that v_init/dt/celsius are necessary resources to run the process. As a result spawn mode cannot get the correct values, only the fork mode runs correctly.

[1uc]

Since the docs for multiprocessing advise [1] against stetting the startup method twice and only from within:

if __name__ == "__main__":

using multiprocessing.get_context might be easier to use [2][3] because it allows to configure the startup method locally.