Python Performance: Past, Present and Future

Speaker

Victor Stinner

Material

• video
• slide

Note

Present

• PyPy
  • fully compatible with CPython
  • issues
    • slower when involving C extension
    • larger memory footprint
    • longer startup time
• multiprocessing
  • Work around the GIL limitation
  • Shared memory (Python 3.8) avoids memory copies between workers
• Cython
  • Easy way to write C extension
  • Support multiple Python versions
• Numba
  • JIT compiler tranlating subset of Python and NumPy
  • Simplified Threading (release GIL)
• Summary
  • PyPy doesn't require any code change
  • multiprocessing scales
  • Use Cython, don't use C API
  • Numba makes numpy faster

Future

• New PyHandle C API
  • CPython
    • on top of the current C API
  • PyPy
    • More efficient that the current C API
  • Cython
    • no need to change your code, Cython will handle it
• Reference counting
• Gilectomy
• Tracing CG
• Subinterpreters (WIP)
  • PEP 545
  • GIL -> on lock per interpreter
  • Expectation
    • lower memory footprint
    • faster locks
    • limitation: python object cannot be shared between interpreters

Share on: Twitter ❄ Facebook ❄ Email

---