Difference between -n=auto and -n=12 (I have 12 logical cores)

[samuelwhiskeyjohnson]

-n=auto uses as many CPU cores as possible. But why is there difference with manually specifying the number by yourself e.g.) -n=12 when you have 12 logical cores.

[RonnyPfannschmidt]

It's not clear what the ask is

[samuelwhiskeyjohnson]

Why is there difference in performance between -n=auto and -n=12 (12 being the maximum cores I have)?

[RonnyPfannschmidt]

What n is getting picked automatically?

[samuelwhiskeyjohnson]

I'm not sure what the confusion is? Isn't this the pytest plugin where you can run pytest in parallel?

In the book, Python Testing with Pytest,
pytest --count=10 -n=auto test_parallel.py
This is how you would run pytest on test_parallel.py 10 times. -n=auto makes the pytest run in parallel to run on as many CPU cores as possible.

But you can also manually specify the CPU cores by -n=<what ever number> such as -n=6

[RonnyPfannschmidt]

closing as the question left out details on whats happening and we cant answer without guessing everything

[webknjaz]

One more point: spawning subprocesses is expensive. So if you only have a few tests, oftentimes it's faster without xdist at all. You need to observe what's happening in your case.

[RonnyPfannschmidt]

@samuelwhiskeyjohnson if the same number of processes is used, there shouldn't be a actual difference - if there is, its a bug that needs a reproducer

[samuelwhiskeyjohnson]

What's the output of cat /proc/cpuinfo and nproc?

I'm on windows so this command doesnt work. But I have Ryzen 3600 which has 6 Physical cores and 12 Logical cores.
so pytest -n auto and pytest -n 12 gives you the same time.

I think I made a mistake. What I meant to say is there is a difference between pytest -n auto and pytest -n 6.
I did 6 because when I read the documentation is says With ``-n auto``, pytest-xdist will use as many processes as your computer has physical CPU cores.

But isn't this wrong? You specify the number of logical cores, not physical cores. When I run pytest -n auto it gives you 12 workers like there are 12 logical cores as Ryzen 3600 has.
If the documentation is right then pytest -n auto should give you 6 workers as there are 6 physical cores for Ryzen 3600.

[webknjaz]

@RonnyPfannschmidt I suppose this might be a docs bug, right? Having looked at https://pytest-xdist.rtfd.io/en/latest/distribution.html#running-tests-across-multiple-cpus, it seems that in this scenario, -n auto would have to return 6 and -n logical would be 12. However, it's probably the opposite.

I tried the same detection techniques locally, and here's what I got:

$ python
Python 3.11.4 (main, Jul 18 2023, 13:30:02) [GCC 12.2.1 20230428] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import psutil
>>> psutil.cpu_count()
12
>>> psutil.cpu_count(logical=True)
12
>>> psutil.cpu_count(logical=False)
6
>>> from os import sched_getaffinity
>>> sched_getaffinity(0)
{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}
>>> len(sched_getaffinity(0))
12
>>> from os import cpu_count
>>> cpu_count()
12
>>> from multiprocessing import cpu_count
>>> cpu_count()
12

This essentially means that -n auto returns whatever the kernel/OS sees, which is usually the number of the logical cores. Except for when psutil is installed — in this case, config.option.numprocesses == "logical" will be False and then, -n auto would return the number of physical cores (https://github.com/pytest-dev/pytest-xdist/blob/d8efeb9/src/xdist/plugin.py#L31).
The -n logical path seems to hit the right branch and return the number of logical cores always.

So I think that the docs might indeed be wrong. I wonder if it'd make sense to have an -n physical too.

[nicoddemus]

So I think that the docs might indeed be wrong.

I think you are right, thanks for the detailed look.