Skip to content

Latest commit

 

History

History
198 lines (157 loc) · 5.86 KB

windows_benchmarks.md

File metadata and controls

198 lines (157 loc) · 5.86 KB
Raw

Running PyPerformance Benchmarks on Windows

Background links:

Quick notes:

  • Don't bother installing from PyPI, that version is out of date

Goals:

  • Get reference benchmark numbers for CPython 3.10
  • Get reliable benchmark numbers for CPython main (to become 3.11)
  • Compare the two

Software to install:

  • git (hopefully you have it installed already :-)
  • CPython 3.10 from python.org (I know there's a 3.10 on the Windows store, I don't know if it was built the same way; it might be worth comparing the two.)

Repos to clone:

  • CPython: git clone https://github.com/python/cpython
  • PyPerformance: git clone https://github.com/python/pyperformance
  • Make sure you have the main branch for each repo

Silencing your machine:

  • This should be done to run the benchmarks
  • Close all apps except for a DOS box
  • Do whatever else you can to shutdown background processes (Stop Windows Defender???)
  • Enter airplane mode or disable networking once the first benchmark is running (I had to leave Bluetooth on since my mouse and keyboard are wireless)
  • Note that on Linux, pyperf system tune will do a bunch of this; sadly this doesn't support Windows.

Getting 3.10 reference benchmark numbers

(Note: maybe we should build 3.10 from source as well? For that, you could follow the instructions below for 3.11 but use git checkout 3.10 to check out the 3.10 branch. But according to Steve Dower it shouldn't make a difference except possibly for startup time -- not sure in which direction.)

  1. Check that you can run CPython 3.10 using the py runner
py -3.10 --version
  1. Check that you have the latest version of pip
py -3.10 -m pip install -U pip setuptools
  1. Install pyperformance from source

This (~\pyperformance\) should point to the pyperformance repo you cloned during the preliminary steps above (best by absolute path).

py -3.10 -m pip install ~\pyperformance\
  1. Check that you can run one benchmark
py -3.10 -m pyperformance run --affinity 1 -b chaos

This should end by printing something like this:

.....................
chaos: Mean +- std dev: 121 ms +- 5 ms

Performance version: 1.0.4
Python version: 3.10.3 (64-bit) revision a342a49
Report on Windows-10-10.0.22000-SP0
Number of logical CPUs: 8
Start date: 2022-04-04 10:52:41.374884
End date: 2022-04-04 10:52:56.707208

### chaos ###
Mean +- std dev: 121 ms +- 5 ms

To be honest I have no idea if the --affinity flag has any effect. Its argument is one or more CPU numbers (comma-separated).

  1. Create the venv(s) for all benchmarks
py -3.10 -m pyperformance venv recreate

This will also seed the pip cache with all the downloads you need.

  1. Run all benchmarks

Before starting this, silence the machine as much as you can (see above). This step will take 30-60 minutes.

py -3.10 -m pyperformance run --affinity 1 -o 310a.json

The -o 310a.json specifies the JSON file where the results are written. The file must not already exist.

  1. Repeat step 6 a few times with different JSON output files

This is so that we get some idea of how stable these numbers are. Let's say you end up with files 310a.json, 310b.json, 310c.json.

It might also be interesting to experiment with different values for --affinity.

  1. Compare the 3.10 results to each other

To compare JSON files I prefer to use pyperf, not pyperformance.

py -m pyperf compare_to 310a.json 310b.json

This prints a geometric mean at the end. Ideally that should be 1.00 faster/slower. I suppose if we see 1.01 faster/slower that's acceptable too. On my machine, alas, I see 1.06 faster, which I interpret as the machine being too noisy to run decent benchmarks.

Try all three pairwise, or specify all JSON files together (see pyperf docs for more about this).

Getting 3.11 benchmark numbers

Here you have to build CPython first.

  1. Build CPython in the cpython repo with PGO+LTO

This will take some time since the PGO training is running a subset of the test suite.

cd ~\cpython
.\PCbuild\build.bat --pgo
  1. Create an installable CPython tree
.\python.bat .\PC\layout\ --preset-default --copy installed

This creates a directory installed containing the python.exe you should use to run PyPerformance.

  1. Ensure you have the latest pip for 3.11
.\installed\python.exe -m ensurepip
.\installed\python.exe -m pip install -U pip setuptools

If this fails, I've probably forgotten some steps.

  1. Install pyperformance from source

This (~\pyperformance\) should point to the pyperformance repo you cloned during the preliminary steps above (best by absolute path).

.\installed\python.exe -m pip install ~\pyperformance\
  1. Verify that one benchmark runs
.\installed\python.exe -m pyperformance run --affinity 1 -b chaos

Output should be similar as for 3.10.

  1. Create the venv(s) for all benchmarks
.\installed\python.exe -m pyperformance venv recreate

This will also seed the pip cache with all the downloads you need (which might be different than for 3.10 in some cases).

  1. Run all benchmarks

Again, silence the machine as much as you can (see above).

.\installed\python.exe -m pyperformance run --affinity 1 -o 311a.json
  1. Run a second time

Again, I'd like to have multiple results so a single fluke doesn't scare us -- or to confirm that things really are as bad is they seem. Compare them as shown above (it doesn't matter which Python binary is used to do this part).

  1. Compare 3.10 and 3.11 results

This should be something as simple as

py -3.10 -m pyperf compare_to 310a.json 311a.json