8 .. sectionauthor:: James Roskind
11 .. index:: single: InfoSeek Corporation
13 Copyright © 1994, by InfoSeek Corporation, all rights reserved.
15 Written by James Roskind. [#]_
17 Permission to use, copy, modify, and distribute this Python software and its
18 associated documentation for any purpose (subject to the restriction in the
19 following sentence) without fee is hereby granted, provided that the above
20 copyright notice appears in all copies, and that both that copyright notice and
21 this permission notice appear in supporting documentation, and that the name of
22 InfoSeek not be used in advertising or publicity pertaining to distribution of
23 the software without specific, written prior permission. This permission is
24 explicitly restricted to the copying and modification of the software to remain
25 in Python, compiled Python, or other languages (such as C) wherein the modified
26 or derived code is exclusively imported into a Python module.
28 INFOSEEK CORPORATION DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
29 INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO EVENT
30 SHALL INFOSEEK CORPORATION BE LIABLE FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL
31 DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS,
32 WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING
33 OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
35 .. _profiler-introduction:
37 Introduction to the profilers
38 =============================
41 single: deterministic profiling
42 single: profiling, deterministic
44 A :dfn:`profiler` is a program that describes the run time performance
45 of a program, providing a variety of statistics. This documentation
46 describes the profiler functionality provided in the modules
47 :mod:`cProfile`, :mod:`profile` and :mod:`pstats`. This profiler
48 provides :dfn:`deterministic profiling` of Python programs. It also
49 provides a series of report generation tools to allow users to rapidly
50 examine the results of a profile operation.
52 The Python standard library provides three different profilers:
54 #. :mod:`cProfile` is recommended for most users; it's a C extension
55 with reasonable overhead
56 that makes it suitable for profiling long-running programs.
57 Based on :mod:`lsprof`,
58 contributed by Brett Rosen and Ted Czotter.
62 #. :mod:`profile`, a pure Python module whose interface is imitated by
63 :mod:`cProfile`. Adds significant overhead to profiled programs.
64 If you're trying to extend
65 the profiler in some way, the task might be easier with this module.
66 Copyright © 1994, by InfoSeek Corporation.
68 .. versionchanged:: 2.4
69 Now also reports the time spent in calls to built-in functions and methods.
71 #. :mod:`hotshot` was an experimental C module that focused on minimizing
72 the overhead of profiling, at the expense of longer data
73 post-processing times. It is no longer maintained and may be
74 dropped in a future version of Python.
77 .. versionchanged:: 2.5
78 The results should be more meaningful than in the past: the timing core
79 contained a critical bug.
81 The :mod:`profile` and :mod:`cProfile` modules export the same interface, so
82 they are mostly interchangeable; :mod:`cProfile` has a much lower overhead but
83 is newer and might not be available on all systems.
84 :mod:`cProfile` is really a compatibility layer on top of the internal
85 :mod:`_lsprof` module. The :mod:`hotshot` module is reserved for specialized
94 This section is provided for users that "don't want to read the manual." It
95 provides a very brief overview, and allows a user to rapidly perform profiling
96 on an existing application.
98 To profile an application with a main entry point of :func:`foo`, you would add
99 the following to your module::
102 cProfile.run('foo()')
104 (Use :mod:`profile` instead of :mod:`cProfile` if the latter is not available on
107 The above action would cause :func:`foo` to be run, and a series of informative
108 lines (the profile) to be printed. The above approach is most useful when
109 working with the interpreter. If you would like to save the results of a
110 profile into a file for later examination, you can supply a file name as the
111 second argument to the :func:`run` function::
114 cProfile.run('foo()', 'fooprof')
116 The file :file:`cProfile.py` can also be invoked as a script to profile another
117 script. For example::
119 python -m cProfile myscript.py
121 :file:`cProfile.py` accepts two optional arguments on the command line::
123 cProfile.py [-o output_file] [-s sort_order]
125 :option:`-s` only applies to standard output (:option:`-o` is not supplied).
126 Look in the :class:`Stats` documentation for valid sort values.
128 When you wish to review the profile, you should use the methods in the
129 :mod:`pstats` module. Typically you would load the statistics data as follows::
132 p = pstats.Stats('fooprof')
134 The class :class:`Stats` (the above code just created an instance of this class)
135 has a variety of methods for manipulating and printing the data that was just
136 read into ``p``. When you ran :func:`cProfile.run` above, what was printed was
137 the result of three method calls::
139 p.strip_dirs().sort_stats(-1).print_stats()
141 The first method removed the extraneous path from all the module names. The
142 second method sorted all the entries according to the standard module/line/name
143 string that is printed. The third method printed out all the statistics. You
144 might try the following sort calls:
146 .. (this is to comply with the semantics of the old profiler).
153 The first call will actually sort the list by function name, and the second call
154 will print out the statistics. The following are some interesting calls to
157 p.sort_stats('cumulative').print_stats(10)
159 This sorts the profile by cumulative time in a function, and then only prints
160 the ten most significant lines. If you want to understand what algorithms are
161 taking time, the above line is what you would use.
163 If you were looking to see what functions were looping a lot, and taking a lot
164 of time, you would do::
166 p.sort_stats('time').print_stats(10)
168 to sort according to time spent within each function, and then print the
169 statistics for the top ten functions.
173 p.sort_stats('file').print_stats('__init__')
175 This will sort all the statistics by file name, and then print out statistics
176 for only the class init methods (since they are spelled with ``__init__`` in
177 them). As one final example, you could try::
179 p.sort_stats('time', 'cum').print_stats(.5, 'init')
181 This line sorts statistics with a primary key of time, and a secondary key of
182 cumulative time, and then prints out some of the statistics. To be specific, the
183 list is first culled down to 50% (re: ``.5``) of its original size, then only
184 lines containing ``init`` are maintained, and that sub-sub-list is printed.
186 If you wondered what functions called the above functions, you could now (``p``
187 is still sorted according to the last criteria) do::
189 p.print_callers(.5, 'init')
191 and you would get a list of callers for each of the listed functions.
193 If you want more functionality, you're going to have to read the manual, or
194 guess what the following functions do::
199 Invoked as a script, the :mod:`pstats` module is a statistics browser for
200 reading and examining profile dumps. It has a simple line-oriented interface
201 (implemented using :mod:`cmd`) and interactive help.
204 .. _deterministic-profiling:
206 What Is Deterministic Profiling?
207 ================================
209 :dfn:`Deterministic profiling` is meant to reflect the fact that all *function
210 call*, *function return*, and *exception* events are monitored, and precise
211 timings are made for the intervals between these events (during which time the
212 user's code is executing). In contrast, :dfn:`statistical profiling` (which is
213 not done by this module) randomly samples the effective instruction pointer, and
214 deduces where time is being spent. The latter technique traditionally involves
215 less overhead (as the code does not need to be instrumented), but provides only
216 relative indications of where time is being spent.
218 In Python, since there is an interpreter active during execution, the presence
219 of instrumented code is not required to do deterministic profiling. Python
220 automatically provides a :dfn:`hook` (optional callback) for each event. In
221 addition, the interpreted nature of Python tends to add so much overhead to
222 execution, that deterministic profiling tends to only add small processing
223 overhead in typical applications. The result is that deterministic profiling is
224 not that expensive, yet provides extensive run time statistics about the
225 execution of a Python program.
227 Call count statistics can be used to identify bugs in code (surprising counts),
228 and to identify possible inline-expansion points (high call counts). Internal
229 time statistics can be used to identify "hot loops" that should be carefully
230 optimized. Cumulative time statistics should be used to identify high level
231 errors in the selection of algorithms. Note that the unusual handling of
232 cumulative times in this profiler allows statistics for recursive
233 implementations of algorithms to be directly compared to iterative
237 Reference Manual -- :mod:`profile` and :mod:`cProfile`
238 ======================================================
241 :synopsis: Python profiler
244 The primary entry point for the profiler is the global function
245 :func:`profile.run` (resp. :func:`cProfile.run`). It is typically used to create
246 any profile information. The reports are formatted and printed using methods of
247 the class :class:`pstats.Stats`. The following is a description of all of these
248 standard entry points and functions. For a more in-depth view of some of the
249 code, consider reading the later section on Profiler Extensions, which includes
250 discussion of how to derive "better" profilers from the classes presented, or
251 reading the source code for these modules.
254 .. function:: run(command[, filename])
256 This function takes a single argument that can be passed to the
257 :keyword:`exec` statement, and an optional file name. In all cases this
258 routine attempts to :keyword:`exec` its first argument, and gather profiling
259 statistics from the execution. If no file name is present, then this function
260 automatically prints a simple profiling report, sorted by the standard name
261 string (file/line/function-name) that is presented in each line. The
262 following is a typical output from such a call::
264 2706 function calls (2004 primitive calls) in 4.504 CPU seconds
266 Ordered by: standard name
268 ncalls tottime percall cumtime percall filename:lineno(function)
269 2 0.006 0.003 0.953 0.477 pobject.py:75(save_objects)
270 43/3 0.533 0.012 0.749 0.250 pobject.py:99(evaluate)
273 The first line indicates that 2706 calls were monitored. Of those calls, 2004
274 were :dfn:`primitive`. We define :dfn:`primitive` to mean that the call was not
275 induced via recursion. The next line: ``Ordered by: standard name``, indicates
276 that the text string in the far right column was used to sort the output. The
277 column headings include:
280 for the number of calls,
283 for the total time spent in the given function (and excluding time made in calls
287 is the quotient of ``tottime`` divided by ``ncalls``
290 is the total time spent in this and all subfunctions (from invocation till
291 exit). This figure is accurate *even* for recursive functions.
294 is the quotient of ``cumtime`` divided by primitive calls
296 filename:lineno(function)
297 provides the respective data of each function
299 When there are two numbers in the first column (for example, ``43/3``), then the
300 latter is the number of primitive calls, and the former is the actual number of
301 calls. Note that when the function does not recurse, these two values are the
302 same, and only the single figure is printed.
305 .. function:: runctx(command, globals, locals[, filename])
307 This function is similar to :func:`run`, with added arguments to supply the
308 globals and locals dictionaries for the *command* string.
310 Analysis of the profiler data is done using the :class:`Stats` class.
314 The :class:`Stats` class is defined in the :mod:`pstats` module.
318 :synopsis: Statistics object for use with the profiler.
321 .. class:: Stats(filename[, stream=sys.stdout[, ...]])
323 This class constructor creates an instance of a "statistics object" from a
324 *filename* (or set of filenames). :class:`Stats` objects are manipulated by
325 methods, in order to print useful reports. You may specify an alternate output
326 stream by giving the keyword argument, ``stream``.
328 The file selected by the above constructor must have been created by the
329 corresponding version of :mod:`profile` or :mod:`cProfile`. To be specific,
330 there is *no* file compatibility guaranteed with future versions of this
331 profiler, and there is no compatibility with files produced by other profilers.
332 If several files are provided, all the statistics for identical functions will
333 be coalesced, so that an overall view of several processes can be considered in
334 a single report. If additional files need to be combined with data in an
335 existing :class:`Stats` object, the :meth:`add` method can be used.
337 .. (such as the old system profiler).
339 .. versionchanged:: 2.5
340 The *stream* parameter was added.
345 The :class:`Stats` Class
346 ------------------------
348 :class:`Stats` objects have the following methods:
351 .. method:: Stats.strip_dirs()
353 This method for the :class:`Stats` class removes all leading path information
354 from file names. It is very useful in reducing the size of the printout to fit
355 within (close to) 80 columns. This method modifies the object, and the stripped
356 information is lost. After performing a strip operation, the object is
357 considered to have its entries in a "random" order, as it was just after object
358 initialization and loading. If :meth:`strip_dirs` causes two function names to
359 be indistinguishable (they are on the same line of the same filename, and have
360 the same function name), then the statistics for these two entries are
361 accumulated into a single entry.
364 .. method:: Stats.add(filename[, ...])
366 This method of the :class:`Stats` class accumulates additional profiling
367 information into the current profiling object. Its arguments should refer to
368 filenames created by the corresponding version of :func:`profile.run` or
369 :func:`cProfile.run`. Statistics for identically named (re: file, line, name)
370 functions are automatically accumulated into single function statistics.
373 .. method:: Stats.dump_stats(filename)
375 Save the data loaded into the :class:`Stats` object to a file named *filename*.
376 The file is created if it does not exist, and is overwritten if it already
377 exists. This is equivalent to the method of the same name on the
378 :class:`profile.Profile` and :class:`cProfile.Profile` classes.
380 .. versionadded:: 2.3
383 .. method:: Stats.sort_stats(key[, ...])
385 This method modifies the :class:`Stats` object by sorting it according to the
386 supplied criteria. The argument is typically a string identifying the basis of
387 a sort (example: ``'time'`` or ``'name'``).
389 When more than one key is provided, then additional keys are used as secondary
390 criteria when there is equality in all keys selected before them. For example,
391 ``sort_stats('name', 'file')`` will sort all the entries according to their
392 function name, and resolve all ties (identical function names) by sorting by
395 Abbreviations can be used for any key names, as long as the abbreviation is
396 unambiguous. The following are the keys currently defined:
398 +------------------+----------------------+
399 | Valid Arg | Meaning |
400 +==================+======================+
401 | ``'calls'`` | call count |
402 +------------------+----------------------+
403 | ``'cumulative'`` | cumulative time |
404 +------------------+----------------------+
405 | ``'file'`` | file name |
406 +------------------+----------------------+
407 | ``'module'`` | file name |
408 +------------------+----------------------+
409 | ``'pcalls'`` | primitive call count |
410 +------------------+----------------------+
411 | ``'line'`` | line number |
412 +------------------+----------------------+
413 | ``'name'`` | function name |
414 +------------------+----------------------+
415 | ``'nfl'`` | name/file/line |
416 +------------------+----------------------+
417 | ``'stdname'`` | standard name |
418 +------------------+----------------------+
419 | ``'time'`` | internal time |
420 +------------------+----------------------+
422 Note that all sorts on statistics are in descending order (placing most time
423 consuming items first), where as name, file, and line number searches are in
424 ascending order (alphabetical). The subtle distinction between ``'nfl'`` and
425 ``'stdname'`` is that the standard name is a sort of the name as printed, which
426 means that the embedded line numbers get compared in an odd way. For example,
427 lines 3, 20, and 40 would (if the file names were the same) appear in the string
428 order 20, 3 and 40. In contrast, ``'nfl'`` does a numeric compare of the line
429 numbers. In fact, ``sort_stats('nfl')`` is the same as ``sort_stats('name',
432 For backward-compatibility reasons, the numeric arguments ``-1``, ``0``, ``1``,
433 and ``2`` are permitted. They are interpreted as ``'stdname'``, ``'calls'``,
434 ``'time'``, and ``'cumulative'`` respectively. If this old style format
435 (numeric) is used, only one sort key (the numeric key) will be used, and
436 additional arguments will be silently ignored.
438 .. For compatibility with the old profiler,
441 .. method:: Stats.reverse_order()
443 This method for the :class:`Stats` class reverses the ordering of the basic list
444 within the object. Note that by default ascending vs descending order is
445 properly selected based on the sort key of choice.
447 .. This method is provided primarily for compatibility with the old profiler.
450 .. method:: Stats.print_stats([restriction, ...])
452 This method for the :class:`Stats` class prints out a report as described in the
453 :func:`profile.run` definition.
455 The order of the printing is based on the last :meth:`sort_stats` operation done
456 on the object (subject to caveats in :meth:`add` and :meth:`strip_dirs`).
458 The arguments provided (if any) can be used to limit the list down to the
459 significant entries. Initially, the list is taken to be the complete set of
460 profiled functions. Each restriction is either an integer (to select a count of
461 lines), or a decimal fraction between 0.0 and 1.0 inclusive (to select a
462 percentage of lines), or a regular expression (to pattern match the standard
463 name that is printed; as of Python 1.5b1, this uses the Perl-style regular
464 expression syntax defined by the :mod:`re` module). If several restrictions are
465 provided, then they are applied sequentially. For example::
467 print_stats(.1, 'foo:')
469 would first limit the printing to first 10% of list, and then only print
470 functions that were part of filename :file:`.\*foo:`. In contrast, the
473 print_stats('foo:', .1)
475 would limit the list to all functions having file names :file:`.\*foo:`, and
476 then proceed to only print the first 10% of them.
479 .. method:: Stats.print_callers([restriction, ...])
481 This method for the :class:`Stats` class prints a list of all functions that
482 called each function in the profiled database. The ordering is identical to
483 that provided by :meth:`print_stats`, and the definition of the restricting
484 argument is also identical. Each caller is reported on its own line. The
485 format differs slightly depending on the profiler that produced the stats:
487 * With :mod:`profile`, a number is shown in parentheses after each caller to
488 show how many times this specific call was made. For convenience, a second
489 non-parenthesized number repeats the cumulative time spent in the function
492 * With :mod:`cProfile`, each caller is preceded by three numbers: the number of
493 times this specific call was made, and the total and cumulative times spent in
494 the current function while it was invoked by this specific caller.
497 .. method:: Stats.print_callees([restriction, ...])
499 This method for the :class:`Stats` class prints a list of all function that were
500 called by the indicated function. Aside from this reversal of direction of
501 calls (re: called vs was called by), the arguments and ordering are identical to
502 the :meth:`print_callers` method.
510 One limitation has to do with accuracy of timing information. There is a
511 fundamental problem with deterministic profilers involving accuracy. The most
512 obvious restriction is that the underlying "clock" is only ticking at a rate
513 (typically) of about .001 seconds. Hence no measurements will be more accurate
514 than the underlying clock. If enough measurements are taken, then the "error"
515 will tend to average out. Unfortunately, removing this first error induces a
516 second source of error.
518 The second problem is that it "takes a while" from when an event is dispatched
519 until the profiler's call to get the time actually *gets* the state of the
520 clock. Similarly, there is a certain lag when exiting the profiler event
521 handler from the time that the clock's value was obtained (and then squirreled
522 away), until the user's code is once again executing. As a result, functions
523 that are called many times, or call many functions, will typically accumulate
524 this error. The error that accumulates in this fashion is typically less than
525 the accuracy of the clock (less than one clock tick), but it *can* accumulate
526 and become very significant.
528 The problem is more important with :mod:`profile` than with the lower-overhead
529 :mod:`cProfile`. For this reason, :mod:`profile` provides a means of
530 calibrating itself for a given platform so that this error can be
531 probabilistically (on the average) removed. After the profiler is calibrated, it
532 will be more accurate (in a least square sense), but it will sometimes produce
533 negative numbers (when call counts are exceptionally low, and the gods of
534 probability work against you :-). ) Do *not* be alarmed by negative numbers in
535 the profile. They should *only* appear if you have calibrated your profiler,
536 and the results are actually better than without calibration.
539 .. _profile-calibration:
544 The profiler of the :mod:`profile` module subtracts a constant from each event
545 handling time to compensate for the overhead of calling the time function, and
546 socking away the results. By default, the constant is 0. The following
547 procedure can be used to obtain a better constant for a given platform (see
548 discussion in section Limitations above). ::
551 pr = profile.Profile()
553 print pr.calibrate(10000)
555 The method executes the number of Python calls given by the argument, directly
556 and again under the profiler, measuring the time for both. It then computes the
557 hidden overhead per profiler event, and returns that as a float. For example,
558 on an 800 MHz Pentium running Windows 2000, and using Python's time.clock() as
559 the timer, the magical number is about 12.5e-6.
561 The object of this exercise is to get a fairly consistent result. If your
562 computer is *very* fast, or your timer function has poor resolution, you might
563 have to pass 100000, or even 1000000, to get consistent results.
565 When you have a consistent answer, there are three ways you can use it: [#]_ ::
569 # 1. Apply computed bias to all Profile instances created hereafter.
570 profile.Profile.bias = your_computed_bias
572 # 2. Apply computed bias to a specific Profile instance.
573 pr = profile.Profile()
574 pr.bias = your_computed_bias
576 # 3. Specify computed bias in instance constructor.
577 pr = profile.Profile(bias=your_computed_bias)
579 If you have a choice, you are better off choosing a smaller constant, and then
580 your results will "less often" show up as negative in profile statistics.
583 .. _profiler-extensions:
585 Extensions --- Deriving Better Profilers
586 ========================================
588 The :class:`Profile` class of both modules, :mod:`profile` and :mod:`cProfile`,
589 were written so that derived classes could be developed to extend the profiler.
590 The details are not described here, as doing this successfully requires an
591 expert understanding of how the :class:`Profile` class works internally. Study
592 the source code of the module carefully if you want to pursue this.
594 If all you want to do is change how current time is determined (for example, to
595 force use of wall-clock time or elapsed process time), pass the timing function
596 you want to the :class:`Profile` class constructor::
598 pr = profile.Profile(your_time_func)
600 The resulting profiler will then call :func:`your_time_func`.
602 :class:`profile.Profile`
603 :func:`your_time_func` should return a single number, or a list of numbers whose
604 sum is the current time (like what :func:`os.times` returns). If the function
605 returns a single time number, or the list of returned numbers has length 2, then
606 you will get an especially fast version of the dispatch routine.
608 Be warned that you should calibrate the profiler class for the timer function
609 that you choose. For most machines, a timer that returns a lone integer value
610 will provide the best results in terms of low overhead during profiling.
611 (:func:`os.times` is *pretty* bad, as it returns a tuple of floating point
612 values). If you want to substitute a better timer in the cleanest fashion,
613 derive a class and hardwire a replacement dispatch method that best handles your
614 timer call, along with the appropriate calibration constant.
616 :class:`cProfile.Profile`
617 :func:`your_time_func` should return a single number. If it returns plain
618 integers, you can also invoke the class constructor with a second argument
619 specifying the real duration of one unit of time. For example, if
620 :func:`your_integer_time_func` returns times measured in thousands of seconds,
621 you would constuct the :class:`Profile` instance as follows::
623 pr = profile.Profile(your_integer_time_func, 0.001)
625 As the :mod:`cProfile.Profile` class cannot be calibrated, custom timer
626 functions should be used with care and should be as fast as possible. For the
627 best results with a custom timer, it might be necessary to hard-code it in the C
628 source of the internal :mod:`_lsprof` module.
630 .. rubric:: Footnotes
632 .. [#] Updated and converted to LaTeX by Guido van Rossum. Further updated by Armin
633 Rigo to integrate the documentation for the new :mod:`cProfile` module of Python
636 .. [#] Prior to Python 2.2, it was necessary to edit the profiler source code to embed
637 the bias as a literal number. You still can, but that method is no longer
638 described, because no longer needed.