FrançaisGet started

Profiling a program

Measuring before changing anything: timing a snippet, choosing between an instrumenting and a sampling profiler, reading own time and cumulative time, a flame graph, and avoiding misleading measurements.

  • 22 cards
  • 5 topics
  • Free
Study this deck

The outline, topic by topic

Each topic is weighted the way the exam weighs it.

  1. Concepts
  2. Questions
  3. Sort and link
  4. Scenarios
  5. Understand

A few cards from the deck

  • Profiling: what is it?

    Show the answer

    Measuring, during a real run, the time or memory used by each part of a program.

    Intuition is often wrong about what is expensive: measure first, decide after. Without a measurement you optimise a line that took 1% of the time.

  • In this cProfile report, which function should be optimised first?

    ```text
    ncalls tottime percall cumtime percall filename:lineno(function)
    1 0.013 0.013 0.013 0.013 p.py:1(lent)
    1 0.000 0.000 0.000 0.000 {method 'disable' of '_lsprof.Profiler' objects}
    1 0.000 0.000 0.013 0.013 p.py:7(main)
    50 0.000 0.000 0.000 0.000 p.py:6(rapide)
    1 0.000 0.000 0.013 0.013 {built-in method builtins.exec}
    1 0.000 0.000 0.013 0.013 p.py:1(<module>)
    `

    • lent
    • main
    • rapide
    • <module>
    Show the answer

    lent

    lent has the biggest tottime: time is spent in its own body. main and <module> have the same cumtime, but only because they call lent. rapide is called 50 times and still costs almost nothing.

  • To time a snippet, repeat it several times and keep the … or the median of the times, since machine noise can only slow a measurement down.

    Show the answer

    To time a snippet, repeat it several times and keep the minimum or the median of the times, since machine noise can only slow a measurement down.

    The mean is pulled up by a run disturbed by another process. timeit.repeat returns the list of each series' times: keep their min.

  • Profiler: what is it?

    Show the answer

    A tool that runs a program while observing how much time (or memory) each function uses and how many times it is called.

    There are two families: instrumenting (every call is recorded, like cProfile, precise but slower) and sampling (the stack is looked at at regular intervals, light enough for production).

  • The first run of a function takes 900 ms, the following ones 40 ms. A colleague concludes the function takes 900 ms. What do you answer?

    • The 900 ms is the true value: later runs are skewed by the cache and do not count
    • Average the two measured values, which gives about 470 ms
    • Recompile the whole program before each of the measurements taken
    • Cold start: repeat and keep the median after warm-up
    Show the answer

    Cold start: repeat and keep the median after warm-up

    The first run loads code and data and fills the caches. Depending on the question, you measure cold start or steady state, never a mix of the two.

Study this deck