Should batch_run default to running on 1 CPU?

[rht]

Defaulting to 1 CPU will simplify the tutorial on Windows.

Then, if number_processes is > 1 and on Windows, we show a warning that tells user to add freeze_support() (we pre-emptively tell the user of this issue instead of waiting for it to happen). This warning can be disabled via a flag to the function:

batch_run(
    ...
   hide_windows_freeze_support_warning=True,
   ...
)

[tpike3]

I think this is the way to go...particularly with new users this is adding cognitive overload. Better to simplify it.

[tpike3]

Looking at different setup.py for #1215, I noticed this use of OpenMP in SciKit Learn. I will explore more as I wonder if there may be solution in this approach to our multiprocessing issues.

[rht]

They seem to use it for the purpose of CI. I have never observed the use of shell env vars in Python scientific computing. The closest one is setting the port for a Flask server, but this is not scientific computing.

[clarkpetri]

I'm running on a Mac M1 and also had to set number_processes = 1 to prevent throwing the errors described in the tutorial. I originally thought I wouldn't need to do so due to working in MacOS.

[tpike3]

Also go George Mason :)

[clarkpetri]

@clarkpetri Just to make sure, are you running your model in Jupyter or an IDE (if so which one) or the Python interpreter?

Jupyter in this instance for BatchRunner.

It's not immediately clear to me what I would do with the fork method or ray library but I'll fiddle with it when I have time (projects, sick kid, very pregnant wife...).

The Mason CSS curriculum has been great to me! Very happy here.

[tpike3]

@clarkpetri For batch_run its best not to run in Jupyter. If you pull the code outside jupyter on a Mac is should run fine, otherwise you need to get the Jupyter Extension so you can run multiprocessing with Jupyter.

[clarkpetri]

Good to know. Thank you! Multiprocessing is working now, though I need to figure out how to stop the subsequent lines of code from executing until all multi-thread operations are complete.

[tpike3]

Make sure you wrap the whole model in batch_run, it is effectively just putting one model with a specific set of parameters on one core. So yes having code outside the batch_run if it is tied to the model will produce odd behaviors.

[clarkpetri]

Got it working now! I don't know what I was doing wrong other than being dense and tired with a sick toddler. Code below and it's performing the 500 runs for my training set substantially faster:

.

[rht]

This SO answer seems to suggest that putting the code inside if __name__ == "__main__": is the piece that actually resolves the problem instead of the freeze_support(). I wonder if you can test this hypothesis.

[clarkpetri]

That's it. Keeping if __name__ == "__main__": but commenting out freeze_support() worked. The inverse and tabbing the code in threw all the same errors I was experiencing.

I was averaging ~ 7it/s on single core but ~65it/s now. 👍

[rht]

Will be resolved by #1210.

[ymgongzi13]

if I want to run it in jupyter notebook, how should I solve this problem?

[ymgongzi13]

and if I use the method above, it works like followings:

[rht]

You have a typo "mian".

[ymgongzi13]

Oh, God! What a stupid error! Maybe my brain needs to take a break. Now it works. Thank you very much!!!!!!!!!!!!!!!!!!!!!!!

[ymgongzi13]

Well, I found a problem nowadays. Though my scripts can run normally, its speed is still very slow when my 'number_processes' was set to 8.
And it need about 2 hours to finish simulate.
However, the same methods in netlogo just spend a few minutes.

Why does this problem happen? And is there some methods to spend less time?
Wait for your kind advice! Thank you very much!

[tpike3]

There should not be that big a difference in the Netlogo run and your Mesa run.

First, are running the same number of iterations? What is your total number of model runs for each version (i.e. Mesa and NetLogo)?

Second, you can find inefficient code, which with this big a difference would indicate to me a possible bug with performance testing. Here is a good overview of performance testing

[ymgongzi13]

They both have 11000 iterations.
Maybe my code needs to be simplified.
I'll try to find the inefficient code by the lib you introduced.
Thanks again for your kind help!!