Skip to content

Latest commit

 

History

History
79 lines (50 loc) · 2.94 KB

Python optimization tips.md

File metadata and controls

79 lines (50 loc) · 2.94 KB
Raw

Optimization Zen ☯️

  1. Does an existing implementation already do what you need? (credit Ariel Rokem, @arokem)
  2. Start with a reference implementation: slow, but clear
  3. Test reference implementation
  4. At any point, stop if performances are "good enough" (credit Boris Gorelik, @boris_gorelik)
  5. Benchmark the reference implementation for bottlenecks
  6. Focus on the "low-hanging fruits" first
  7. Test each iteration against the reference implementation
  8. Step-back: a different algorithm may yield better performances
  9. Do not sacrifice clarity for speed...
  10. ... although, if you must, do it as last step.

Extra rules for scientific computing 📜

  1. Start first with numpy vectorization
  2. Do not forget about numexpr
  3. Use Numba
  4. Switch to Cython for more control or if you need to distribute a smaller self-contained package
  5. To use multi-CPUs, release the GIL and use thread-based parallelism (joblib)

Numba vs Cython 🏁

Both are excellent solutions. Which one to choose?

Speed 🚀

  • Cython: excellent performances 🏃🏃🏃🏃
  • Numba: excellent performances (sometimes a bit faster than cython!) 🏃🏃🏃

🏆 Winner: Numba, but almost a tie (application dependent)

Ease of use 😃

  • Cython: fair
    • need to manully declare types and adding imports (more work than with Numba) 😐
    • Cython does not accept UTF-8 identifiers (α, β, γ, 😢)
    • Need duplicated code for a pure-python reference 😢
  • Numba: easy
    • just decorate with @numba.jit 😃
    • A pure-python reference can be the same function undecorated 😃
    • Default arguments needs to be passed 😢

🏆 Winner: Numba

Control optimization 🔧

  • Cython: good understanding of how is optimized:
    • HTML annotations: see "python-hits" in C code to guide optimization
    • line_profiler: see which line takes the most time
  • Numba: optimization is mostly a black-box. Either it works, or it doesn't. If things go wrong (fail to compile or poor performances) it is more of a guess work to understand why.

🏆 Winner: Cython

Distribution 🚚

  • Cython: you can create small pre-compiled packages with no run-time dependencies (wheels or conda packages)
  • Numba: has a run-time dependency on LLVM. LLVM is by default in Anaconda for all platforms, but still is a "big" dependency.

🏆 Winner: Cython

Multi-platform development 💻 🚴🏿 🎮

  • Cython: easy on Linux and macOS, compiler included in Anaconda. A pain on Windows: need to install MS Visual Studio 2015 (multi-GB) and let Cython find it. Alternatively, on Windows 10 you can try the "native" Linux subsystem: Using WSL and MobaXterm to Create a Linux Dev Environment on Windows.
  • Numba: easy, the LVMM dependency is easy to install on Linux, macOS and Windows

🏆 Winner: Numba