Debian Codegraph @

Doc/library/timeit.rst

raw
1 :mod:`!timeit` --- Measure execution time of small code snippets
2 ================================================================
3
4 .. module:: timeit
5 :synopsis: Measure the execution time of small code snippets.
6
7 **Source code:** :source:`Lib/timeit.py`
8
9 .. index::
10 single: Benchmarking
11 single: Performance
12
13 --------------
14
15 This module provides a simple way to time small bits of Python code. It has both
16 a :ref:`timeit-command-line-interface` as well as a :ref:`callable <python-interface>`
17 one. It avoids a number of common traps for measuring execution times.
18 See also Tim Peters' introduction to the "Algorithms" chapter in the second
19 edition of *Python Cookbook*, published by O'Reilly.
20
21
22 Basic Examples
23 --------------
24
25 The following example shows how the :ref:`timeit-command-line-interface`
26 can be used to compare three different expressions:
27
28 .. code-block:: shell-session
29
30 $ python -m timeit "'-'.join(str(n) for n in range(100))"
31 10000 loops, best of 5: 30.2 usec per loop
32 $ python -m timeit "'-'.join([str(n) for n in range(100)])"
33 10000 loops, best of 5: 27.5 usec per loop
34 $ python -m timeit "'-'.join(map(str, range(100)))"
35 10000 loops, best of 5: 23.2 usec per loop
36
37 This can be achieved from the :ref:`python-interface` with::
38
39 >>> import timeit
40 >>> timeit.timeit('"-".join(str(n) for n in range(100))', number=10000)
41 0.3018611848820001
42 >>> timeit.timeit('"-".join([str(n) for n in range(100)])', number=10000)
43 0.2727368790656328
44 >>> timeit.timeit('"-".join(map(str, range(100)))', number=10000)
45 0.23702679807320237
46
47 A callable can also be passed from the :ref:`python-interface`::
48
49 >>> timeit.timeit(lambda: "-".join(map(str, range(100))), number=10000)
50 0.19665591977536678
51
52 Note however that :func:`.timeit` will automatically determine the number of
53 repetitions only when the command-line interface is used. In the
54 :ref:`timeit-examples` section you can find more advanced examples.
55
56
57 .. _python-interface:
58
59 Python Interface
60 ----------------
61
62 The module defines three convenience functions and a public class:
63
64
65 .. function:: timeit(stmt='pass', setup='pass', timer=<default timer>, number=1000000, globals=None)
66
67 Create a :class:`Timer` instance with the given statement, *setup* code and
68 *timer* function and run its :meth:`.timeit` method with *number* executions.
69 The optional *globals* argument specifies a namespace in which to execute the
70 code.
71
72 .. versionchanged:: 3.5
73 The optional *globals* parameter was added.
74
75
76 .. function:: repeat(stmt='pass', setup='pass', timer=<default timer>, repeat=5, number=1000000, globals=None)
77
78 Create a :class:`Timer` instance with the given statement, *setup* code and
79 *timer* function and run its :meth:`.repeat` method with the given *repeat*
80 count and *number* executions. The optional *globals* argument specifies a
81 namespace in which to execute the code.
82
83 .. versionchanged:: 3.5
84 The optional *globals* parameter was added.
85
86 .. versionchanged:: 3.7
87 Default value of *repeat* changed from 3 to 5.
88
89
90 .. function:: default_timer()
91
92 The default timer, which is always time.perf_counter(), returns float seconds.
93 An alternative, time.perf_counter_ns, returns integer nanoseconds.
94
95 .. versionchanged:: 3.3
96 :func:`time.perf_counter` is now the default timer.
97
98
99 .. class:: Timer(stmt='pass', setup='pass', timer=<timer function>, globals=None)
100
101 Class for timing execution speed of small code snippets.
102
103 The constructor takes a statement to be timed, an additional statement used
104 for setup, and a timer function. Both statements default to ``'pass'``;
105 the timer function is platform-dependent (see the module doc string).
106 *stmt* and *setup* may also contain multiple statements separated by ``;``
107 or newlines, as long as they don't contain multi-line string literals. The
108 statement will by default be executed within timeit's namespace; this behavior
109 can be controlled by passing a namespace to *globals*.
110
111 To measure the execution time of the first statement, use the :meth:`.timeit`
112 method. The :meth:`.repeat` and :meth:`.autorange` methods are convenience
113 methods to call :meth:`.timeit` multiple times.
114
115 The execution time of *setup* is excluded from the overall timed execution run.
116
117 The *stmt* and *setup* parameters can also take objects that are callable
118 without arguments. This will embed calls to them in a timer function that
119 will then be executed by :meth:`.timeit`. Note that the timing overhead is a
120 little larger in this case because of the extra function calls.
121
122 .. versionchanged:: 3.5
123 The optional *globals* parameter was added.
124
125 .. method:: Timer.timeit(number=1000000)
126
127 Time *number* executions of the main statement. This executes the setup
128 statement once, and then returns the time it takes to execute the main
129 statement a number of times. The default timer returns seconds as a float.
130 The argument is the number of times through the loop, defaulting to one
131 million. The main statement, the setup statement and the timer function
132 to be used are passed to the constructor.
133
134 .. note::
135
136 By default, :meth:`.timeit` temporarily turns off :term:`garbage
137 collection` during the timing. The advantage of this approach is that
138 it makes independent timings more comparable. The disadvantage is
139 that GC may be an important component of the performance of the
140 function being measured. If so, GC can be re-enabled as the first
141 statement in the *setup* string. For example::
142
143 timeit.Timer('for i in range(10): oct(i)', 'gc.enable()').timeit()
144
145
146 .. method:: Timer.autorange(callback=None)
147
148 Automatically determine how many times to call :meth:`.timeit`.
149
150 This is a convenience function that calls :meth:`.timeit` repeatedly
151 so that the total time >= 0.2 second, returning the eventual
152 (number of loops, time taken for that number of loops). It calls
153 :meth:`.timeit` with increasing numbers from the sequence 1, 2, 5,
154 10, 20, 50, ... until the time taken is at least 0.2 seconds.
155
156 If *callback* is given and is not ``None``, it will be called after
157 each trial with two arguments: ``callback(number, time_taken)``.
158
159 .. versionadded:: 3.6
160
161
162 .. method:: Timer.repeat(repeat=5, number=1000000)
163
164 Call :meth:`.timeit` a few times.
165
166 This is a convenience function that calls the :meth:`.timeit` repeatedly,
167 returning a list of results. The first argument specifies how many times
168 to call :meth:`.timeit`. The second argument specifies the *number*
169 argument for :meth:`.timeit`.
170
171 .. note::
172
173 It's tempting to calculate mean and standard deviation from the result
174 vector and report these. However, this is not very useful.
175 In a typical case, the lowest value gives a lower bound for how fast
176 your machine can run the given code snippet; higher values in the
177 result vector are typically not caused by variability in Python's
178 speed, but by other processes interfering with your timing accuracy.
179 So the :func:`min` of the result is probably the only number you
180 should be interested in. After that, you should look at the entire
181 vector and apply common sense rather than statistics.
182
183 .. versionchanged:: 3.7
184 Default value of *repeat* changed from 3 to 5.
185
186
187 .. method:: Timer.print_exc(file=None)
188
189 Helper to print a traceback from the timed code.
190
191 Typical use::
192
193 t = Timer(...) # outside the try/except
194 try:
195 t.timeit(...) # or t.repeat(...)
196 except Exception:
197 t.print_exc()
198
199 The advantage over the standard traceback is that source lines in the
200 compiled template will be displayed. The optional *file* argument directs
201 where the traceback is sent; it defaults to :data:`sys.stderr`.
202
203
204 .. _timeit-command-line-interface:
205
206 Command-Line Interface
207 ----------------------
208
209 When called as a program from the command line, the following form is used::
210
211 python -m timeit [-n N] [-r N] [-u U] [-s S] [-p] [-v] [-h] [statement ...]
212
213 Where the following options are understood:
214
215 .. program:: timeit
216
217 .. option:: -n N, --number=N
218
219 how many times to execute 'statement'
220
221 .. option:: -r N, --repeat=N
222
223 how many times to repeat the timer (default 5)
224
225 .. option:: -s S, --setup=S
226
227 statement to be executed once initially (default ``pass``)
228
229 .. option:: -p, --process
230
231 measure process time, not wallclock time, using :func:`time.process_time`
232 instead of :func:`time.perf_counter`, which is the default
233
234 .. versionadded:: 3.3
235
236 .. option:: -u, --unit=U
237
238 specify a time unit for timer output; can select ``nsec``, ``usec``, ``msec``, or ``sec``
239
240 .. versionadded:: 3.5
241
242 .. option:: -v, --verbose
243
244 print raw timing results; repeat for more digits precision
245
246 .. option:: -h, --help
247
248 print a short usage message and exit
249
250 A multi-line statement may be given by specifying each line as a separate
251 statement argument; indented lines are possible by enclosing an argument in
252 quotes and using leading spaces. Multiple :option:`-s` options are treated
253 similarly.
254
255 If :option:`-n` is not given, a suitable number of loops is calculated by trying
256 increasing numbers from the sequence 1, 2, 5, 10, 20, 50, ... until the total
257 time is at least 0.2 seconds.
258
259 :func:`default_timer` measurements can be affected by other programs running on
260 the same machine, so the best thing to do when accurate timing is necessary is
261 to repeat the timing a few times and use the best time. The :option:`-r`
262 option is good for this; the default of 5 repetitions is probably enough in
263 most cases. You can use :func:`time.process_time` to measure CPU time.
264
265 .. note::
266
267 There is a certain baseline overhead associated with executing a pass statement.
268 The code here doesn't try to hide it, but you should be aware of it. The
269 baseline overhead can be measured by invoking the program without arguments,
270 and it might differ between Python versions.
271
272
273 .. _timeit-examples:
274
275 Examples
276 --------
277
278 It is possible to provide a setup statement that is executed only once at the beginning:
279
280 .. code-block:: shell-session
281
282 $ python -m timeit -s "text = 'sample string'; char = 'g'" "char in text"
283 5000000 loops, best of 5: 0.0877 usec per loop
284 $ python -m timeit -s "text = 'sample string'; char = 'g'" "text.find(char)"
285 1000000 loops, best of 5: 0.342 usec per loop
286
287 In the output, there are three fields. The loop count, which tells you how many
288 times the statement body was run per timing loop repetition. The repetition
289 count ('best of 5') which tells you how many times the timing loop was
290 repeated, and finally the time the statement body took on average within the
291 best repetition of the timing loop. That is, the time the fastest repetition
292 took divided by the loop count.
293
294 ::
295
296 >>> import timeit
297 >>> timeit.timeit('char in text', setup='text = "sample string"; char = "g"')
298 0.41440500499993504
299 >>> timeit.timeit('text.find(char)', setup='text = "sample string"; char = "g"')
300 1.7246671520006203
301
302 The same can be done using the :class:`Timer` class and its methods::
303
304 >>> import timeit
305 >>> t = timeit.Timer('char in text', setup='text = "sample string"; char = "g"')
306 >>> t.timeit()
307 0.3955516149999312
308 >>> t.repeat()
309 [0.40183617287970225, 0.37027556854118704, 0.38344867356679524, 0.3712595970846668, 0.37866875250654886]
310
311
312 The following examples show how to time expressions that contain multiple lines.
313 Here we compare the cost of using :func:`hasattr` vs. :keyword:`try`/:keyword:`except`
314 to test for missing and present object attributes:
315
316 .. code-block:: shell-session
317
318 $ python -m timeit "try:" " str.__bool__" "except AttributeError:" " pass"
319 20000 loops, best of 5: 15.7 usec per loop
320 $ python -m timeit "if hasattr(str, '__bool__'): pass"
321 50000 loops, best of 5: 4.26 usec per loop
322
323 $ python -m timeit "try:" " int.__bool__" "except AttributeError:" " pass"
324 200000 loops, best of 5: 1.43 usec per loop
325 $ python -m timeit "if hasattr(int, '__bool__'): pass"
326 100000 loops, best of 5: 2.23 usec per loop
327
328 ::
329
330 >>> import timeit
331 >>> # attribute is missing
332 >>> s = """\
333 ... try:
334 ... str.__bool__
335 ... except AttributeError:
336 ... pass
337 ... """
338 >>> timeit.timeit(stmt=s, number=100000)
339 0.9138244460009446
340 >>> s = "if hasattr(str, '__bool__'): pass"
341 >>> timeit.timeit(stmt=s, number=100000)
342 0.5829014980008651
343 >>>
344 >>> # attribute is present
345 >>> s = """\
346 ... try:
347 ... int.__bool__
348 ... except AttributeError:
349 ... pass
350 ... """
351 >>> timeit.timeit(stmt=s, number=100000)
352 0.04215312199994514
353 >>> s = "if hasattr(int, '__bool__'): pass"
354 >>> timeit.timeit(stmt=s, number=100000)
355 0.08588060699912603
356
357
358 To give the :mod:`!timeit` module access to functions you define, you can pass a
359 *setup* parameter which contains an import statement::
360
361 def test():
362 """Stupid test function"""
363 L = [i for i in range(100)]
364
365 if __name__ == '__main__':
366 import timeit
367 print(timeit.timeit("test()", setup="from __main__ import test"))
368
369 Another option is to pass :func:`globals` to the *globals* parameter, which will cause the code
370 to be executed within your current global namespace. This can be more convenient
371 than individually specifying imports::
372
373 def f(x):
374 return x**2
375 def g(x):
376 return x**4
377 def h(x):
378 return x**8
379
380 import timeit
381 print(timeit.timeit('[func(42) for func in (f,g,h)]', globals=globals()))