owned this note   changed 6 years ago  
Linked with GitHub

Develop Numerical Software - Yung-Yu Chen

歡迎來到 PyCon TW 2019 共筆

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

共筆入口：https://hackmd.io/@pycontw/2019
手機版請點選上方 按鈕展開議程列表。

PyCon Taiwan 2019 PyNight 活動規劃與 BoF 簽到/建立交流主題
https://hackmd.io/NaI9ymzRQ5urCvwOQ31WAA

tags: PyConTW2019

Talk link: https://tw.pycon.org/2019/en-us/events/talk/870171889119527188/

Slide deck: https://speakerdeck.com/yyc/develop-numerical-software-pycontw-2019

Open space: https://hackmd.io/JkVMAE6oR-qwUMeGeKmpjA

Starts here

What is numerical software?

  • Problem to Solve
    • Mathematics
    • Physics
    • Engineering
    • and many others…
  • Numerical software
    • Research code
    • Niche markets
  • It is not practical to ask everyone to have deep understanding inside math.
    • Engineering works here!
    • Create a numerical software for engineers easy to use.

Languages of Numerical Software

  • Numerical programmers speaks Fortran back then.
  • BLAS and LAPACK after Fortran.
  • NumPy!! yay!

NumPy

  • 5.24 sec (pure Python) / 58.3 ms (NumPy) = 89.87x faster!
  • Multi-dimentional arrays
    • Be careful of majoring! (row/column major)
  • memory management: zero-copy!
    • During cross-language programming, make one language to allocate memory, the other just access it.
    • Reduce memory use.
  • High performance code
    • number crunching code is in deep or long loops.
    • try to optimize the calculation kernel as possible.

How to optimize calculation kernel

  1. Timing for hotspots
    • Do not optimize without any time measurement!
    • time measurements
      1. CPU time
        • user time
        • system time
      2. Wall time: the time duration in real world (clock on the wall)
    • Linux time command
    • Python timeit module
      • By default if only calculates wall time
      • Alternative is to use perf_counter() or perf_counter_ns() from time module
  2. Memory management
    • Memory hierarchy
    ​​​​registers(0 cycles)
​​​​L1 cache (4 cycles)
​​​​L2 cache (10 cycles)
​​​​...
​​​​Memory
​​​​...
​​​​Disk     (100,000 cycles)
    • Locality: access memory consecutively.
    • Profiling
      • profile of memory is much harder than for runtime.
      • Need your own malloc, realloc, calloc, free.
    • Large memory
      • 10,000 x 10,000 matrix of double takes 800MB
    • Instance counting
    • Tips
      • When modularizing code, put a memory manager in each module.
      • Allow system to switch between different global memory managers. (glibc/jemalloc/tcmalloc/…)
      • Use runtime analyzers like valgrind
      • Learn PyMem
  3. Software Engineering
    • Keyword: Automation
    • Version control is foundation.
    • Build system
      • Makefile
      • CMake
      • Python distutil is still useful if you don't have too complicated system combined multiple languages.
    • Unit testing
      • Google-test
      • Python stdlib unittest and py.test
    • Code review and paper writing
      • Write down papers which is acceptable for version control.
        • TeX/LaTeX is worth to consider for mathematical formula.

System design patterns

  • Data Object
  • Nested loop
    • Except in high-level flow control, there 's no room for fancy things like list comprehension
  • Evolution design
    • Congrats! You should celebrate about it!
    • Doesn't go right in the first place.
    • Insert debug logs
  • Release and distribution
    • Packing and deliver is the first step to move from research code to product code!
Select a repo