dummy-link

TimeIt

timeit macro for Julia

Readme

TimeIt

This module exports a @timeit macro that works similarly to the %timeit magic in IPython.

THIS PACKAGE IS DEPRECATED: It no longer works correctly on Julia v0.7+ due to scoping changes in Julia. Use @btime from BenchmarkTools.jl instead.

Installation

julia> Pkg.clone("git://github.com/kbarbary/TimeIt.jl.git")

Usage

julia> using TimeIt

julia> x = rand(10000); y = rand(10000);

julia> @timeit x .* y;
10000 loops, best of 3: 13.71 µs per loop

The time per loop in seconds is returned, so you can do t = @timeit x .* y to record the timing.

Caveats

TL;DR: For a more rigorous benchmarking experience, check out Benchmarks.jl.

  1. The macro will give inaccurate results for expressions that take less than about 1 µs.

@timeit wraps the supplied expression in a loop with an appropriate number of iterations and times the execution. For example, @timeit x .* y becomes something like

   # record start time
   for i in 1:10000
       val = x .* y
   end
   # record end time

As such, when you run @timeit x .* y in global scope (at the REPL), you're running a loop in global scope, which has a non-negligible overhead. Even a do-nothing expression takes a minimum of about 100 ns per iteration:

julia> @timeit nothing
   1000000 loops, best of 3: 98.05 ns per loop

For expressions that take more than approximately 1µs, this overhead becomes negligible.

  1. All typical Julia performance gotchas about global scope still apply.

Again, because the loop is in global scope, complex expressions will have poor performance. Consider an alternative way to compute x .* y:

   julia> @timeit [x[i] * y[i] for i=1:length(x)]
   100 loops, best of 3: 3.06 ms per loop

Placing this expression in a function results in a factor of 100 speed-up:

   julia> f(x, y) = [x[i] * y[i] for i=1:length(x)]
   f (generic function with 1 method)

   julia> @timeit f(x, y)
   10000 loops, best of 3: 32.44 µs per loop

First Commit

02/09/2014

Last Touched

10 months ago

Commits

9 commits