A native runtime bootstrap for PEXes.
Warning
This set of tools is very alpha and definitely not intended for production use yet!
PEXes must meet a few basic criteria to meet the historic Pex design goals:
- Support both CPython 2.7 and PyPy 2.7 as well as 3.5 and up for both implementations.
- Support multi-platform PEXes that include platform-specific wheels for each targeted platform.
- Provide a hermetic runtime environment by default. You can only import the project packages your PEX ships with no matter the vagaries of the machines the PEX lands on.
PexCZ provides a pexcz binary that can take an existing zipapp PEX and replace its runtime
.bootstrap/ with a native code bootstrap that meets all the design goals above while also
producing smaller PEXes (CZEXes 1) that are faster to execute in both cold and warm cache
scenarios across the full range of PEX sizes.
For example:
# Given both a traditional zipapp PEX and a venv PEX:
:; pex cowsay -c cowsay -o cowsay.zipapp.pex
:; pex cowsay -c cowsay --venv -o cowsay.venv.pex
# Create CZEXes from them:
:; time zig-out/bin/native/pexcz inject cowsay.zipapp.pex
Injected pexcz runtime for cowsay.zipapp.pex in cowsay.zipapp.czex
real 0m0.090s
user 0m0.087s
sys 0m0.003s
:; zig-out/bin/native/pexcz inject cowsay.venv.pex
Injected pexcz runtime for cowsay.venv.pex in cowsay.venv.czex
# Compare PEX and CZEX sizes:
:; ls -1sh cowsay.zipapp.* cowsay.venv.*
400K cowsay.venv.czex
952K cowsay.venv.pex
400K cowsay.zipapp.czex
932K cowsay.zipapp.pex
# Compare cold cache speed:
:; hyperfine -w2 \
-p 'rm -rf ~/.cache/pex' 'python cowsay.zipapp.pex -t Moo!' \
-p 'rm -rf ~/.cache/pex' 'python cowsay.venv.pex -t Moo!' \
-p 'rm -rf ~/.cache/pexcz/' 'python cowsay.zipapp.czex -t -t Moo!' \
-p 'rm -rf ~/.cache/pexcz/' 'python cowsay.venv.czex -t -t Moo!'
Benchmark 1: python cowsay.zipapp.pex -t Moo!
Time (mean ± σ): 675.0 ms ± 5.7 ms [User: 602.1 ms, System: 73.5 ms]
Range (min … max): 665.4 ms … 683.3 ms 10 runs
Benchmark 2: python cowsay.venv.pex -t Moo!
Time (mean ± σ): 806.3 ms ± 22.3 ms [User: 714.7 ms, System: 87.5 ms]
Range (min … max): 789.0 ms … 865.5 ms 10 runs
Benchmark 3: python cowsay.zipapp.czex -t -t Moo!
Time (mean ± σ): 106.1 ms ± 0.9 ms [User: 78.8 ms, System: 34.4 ms]
Range (min … max): 104.6 ms … 108.1 ms 27 runs
Benchmark 4: python cowsay.venv.czex -t -t Moo!
Time (mean ± σ): 105.8 ms ± 0.7 ms [User: 77.2 ms, System: 36.0 ms]
Range (min … max): 104.4 ms … 107.0 ms 27 runs
Summary
python cowsay.venv.czex -t -t Moo! ran
1.00 ± 0.01 times faster than python cowsay.zipapp.czex -t -t Moo!
6.38 ± 0.07 times faster than python cowsay.zipapp.pex -t Moo!
7.62 ± 0.22 times faster than python cowsay.venv.pex -t Moo!
# Compare warm cache speed:
:; hyperfine -w2 \
'python cowsay.zipapp.pex -t Moo!' \
'python cowsay.venv.pex -t Moo!' \
'python cowsay.zipapp.czex -t -t Moo!' \
'python cowsay.venv.czex -t -t Moo!'
Benchmark 1: python cowsay.zipapp.pex -t Moo!
Time (mean ± σ): 315.1 ms ± 2.9 ms [User: 272.6 ms, System: 43.5 ms]
Range (min … max): 309.5 ms … 318.2 ms 10 runs
Benchmark 2: python cowsay.venv.pex -t Moo!
Time (mean ± σ): 118.5 ms ± 1.3 ms [User: 92.5 ms, System: 27.2 ms]
Range (min … max): 117.1 ms … 122.8 ms 24 runs
Benchmark 3: python cowsay.zipapp.czex -t -t Moo!
Time (mean ± σ): 72.3 ms ± 0.9 ms [User: 50.3 ms, System: 23.2 ms]
Range (min … max): 70.9 ms … 75.5 ms 41 runs
Benchmark 4: python cowsay.venv.czex -t -t Moo!
Time (mean ± σ): 73.4 ms ± 8.4 ms [User: 49.4 ms, System: 23.8 ms]
Range (min … max): 71.3 ms … 125.8 ms 41 runs
Summary
python cowsay.zipapp.czex -t -t Moo! ran
1.01 ± 0.12 times faster than python cowsay.venv.czex -t -t Moo!
1.64 ± 0.03 times faster than python cowsay.venv.pex -t Moo!
4.36 ± 0.07 times faster than python cowsay.zipapp.pex -t Moo!On the huge PEX side of the spectrum, some extra tricks come to the fore. Namely, CZEXes use zstd
compression for all files except __main__.py and PEX-INFO and zip extraction is further
parallelized across all available cores.
Using the torch case:
# Given a traditional zipapp torch PEX:
:; pex torch -o torch.pex
# Create a CZEX from it:
:; time zig-out/bin/native/pexcz inject torch.pex
Injected pexcz runtime for torch.pex in torch.czex
real 0m41.022s
user 0m38.733s
sys 0m2.208s
# That took a little bit! But a pretty big space savings is a result:
:; ls -1sh torch.pex torch.czex
2.3G torch.czex
2.9G torch.pex
# Cold cache perf is improved:
:; hyperfine -w1 -r3 \
-p 'rm -rf ~/.cache/pexcz' 'python3.13 torch.czex -c -c "import torch; print(torch.__file__)"' \
-p 'rm -rf ~/.cache/pex' 'python3.13 torch.pex -c "import torch; print(torch.__file__)"'
Benchmark 1: python3.13 torch.czex -c -c "import torch; print(torch.__file__)"
Time (mean ± σ): 4.604 s ± 0.114 s [User: 12.974 s, System: 3.992 s]
Range (min … max): 4.481 s … 4.705 s 3 runs
Benchmark 2: python3.13 torch.pex -c "import torch; print(torch.__file__)"
Time (mean ± σ): 25.812 s ± 0.441 s [User: 23.585 s, System: 1.983 s]
Range (min … max): 25.459 s … 26.306 s 3 runs
Summary
python3.13 torch.czex -c -c "import torch; print(torch.__file__)" ran
5.61 ± 0.17 times faster than python3.13 torch.pex -c "import torch; print(torch.__file__)"
# As is warm cache perf:
:; hyperfine -w1 -r3 \
'python3.13 torch.czex -c -c "import torch; print(torch.__file__)"' \
'python3.13 torch.pex -c "import torch; print(torch.__file__)"'
Benchmark 1: python3.13 torch.czex -c -c "import torch; print(torch.__file__)"
Time (mean ± σ): 1.146 s ± 0.003 s [User: 1.015 s, System: 0.131 s]
Range (min … max): 1.142 s … 1.148 s 3 runs
Benchmark 2: python3.13 torch.pex -c "import torch; print(torch.__file__)"
Time (mean ± σ): 2.150 s ± 0.006 s [User: 1.962 s, System: 0.189 s]
Range (min … max): 2.143 s … 2.156 s 3 runs
Summary
python3.13 torch.czex -c -c "import torch; print(torch.__file__)" ran
1.88 ± 0.01 times faster than python3.13 torch.pex -c "import torch; print(torch.__file__)"N.B,.: The ideas developed in this repo, once proved out, will likely move into the main Pex repo or at least used by the Pex CLI tool to replace the current pure-Python PEX bootstrap runtime.