by Maciej Fijalkowski for EuroPython 2012
When writing code in C, the mapping between written code and compiled assembler is generally pretty well understood. Compilers employ tricks, but barring few extreme examples they’re, typically within 20% performance difference from “naive” compilation. Tools used to asses performance are usually per-function profilers, like valgrind, oprofile or gnuprof. They all can display time spent per function as well as cumultative times and call chains. This works very well for small-to-medium programs, however large, already profiled programs tend to have flat performance profile with unclear message from profilers.
To make matters even more complicated, in JITted VMs, like PyPy, mapping between assembler and high level language, like python, is very unclear and not well known. Writing “good” vs “bad” python (from the JIT perspective), can make a 20x performance difference.
This talk will cover current profilers available for Pyhon and PyPy, as well as other tools that can be used to asses performance. I’ll also present in which cases using current tools does not give necessary information and what kind of tools can address this problem in the future.
