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 The profiler was written after only programming in Python for 3 weeks. As a
36 result, it is probably clumsy code, but I don't know for sure yet 'cause I'm a
37 beginner :-). I did work hard to make the code run fast, so that profiling
38 would be a reasonable thing to do. I tried not to repeat code fragments, but
39 I'm sure I did some stuff in really awkward ways at times. Please send
40 suggestions for improvements to: jar@netscape.com. I won't promise *any*
41 support. ...but I'd appreciate the feedback.
44 .. _profiler-introduction:
46 Introduction to the profilers
47 =============================
50 single: deterministic profiling
51 single: profiling, deterministic
53 A :dfn:`profiler` is a program that describes the run time performance of a
54 program, providing a variety of statistics. This documentation describes the
55 profiler functionality provided in the modules :mod:`profile` and :mod:`pstats`.
56 This profiler provides :dfn:`deterministic profiling` of any Python programs.
57 It also provides a series of report generation tools to allow users to rapidly
58 examine the results of a profile operation.
60 The Python standard library provides three different profilers:
62 #. :mod:`profile`, a pure Python module, described in the sequel. Copyright ©
63 1994, by InfoSeek Corporation.
65 .. versionchanged:: 2.4
66 also reports the time spent in calls to built-in functions and methods.
68 #. :mod:`cProfile`, a module written in C, with a reasonable overhead that makes
69 it suitable for profiling long-running programs. Based on :mod:`lsprof`,
70 contributed by Brett Rosen and Ted Czotter.
74 #. :mod:`hotshot`, a C module focusing on minimizing the overhead while
75 profiling, at the expense of long data post-processing times.
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 interchangeables; :mod:`cProfile` has a much lower overhead but
83 is not so far as well-tested 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 to specialized
88 .. \section{How Is This Profiler Different From The Old Profiler?}
89 \nodename{Profiler Changes}
91 (This section is of historical importance only; the old profiler
92 discussed here was last seen in Python 1.1.)
94 The big changes from old profiling module are that you get more
95 information, and you pay less CPU time. It's not a trade-off, it's a
103 Local stack frame is no longer molested, execution time is now charged
104 to correct functions.
106 \item[Accuracy increased:]
107 Profiler execution time is no longer charged to user's code,
108 calibration for platform is supported, file reads are not done \emph{by}
109 profiler \emph{during} profiling (and charged to user's code!).
111 \item[Speed increased:]
112 Overhead CPU cost was reduced by more than a factor of two (perhaps a
113 factor of five), lightweight profiler module is all that must be
114 loaded, and the report generating module (\module{pstats}) is not needed
117 \item[Recursive functions support:]
118 Cumulative times in recursive functions are correctly calculated;
119 recursive entries are counted.
121 \item[Large growth in report generating UI:]
122 Distinct profiles runs can be added together forming a comprehensive
123 report; functions that import statistics take arbitrary lists of
124 files; sorting criteria is now based on keywords (instead of 4 integer
125 options); reports shows what functions were profiled as well as what
126 profile file was referenced; output format has been improved.
133 Instant User's Manual
134 =====================
136 This section is provided for users that "don't want to read the manual." It
137 provides a very brief overview, and allows a user to rapidly perform profiling
138 on an existing application.
140 To profile an application with a main entry point of :func:`foo`, you would add
141 the following to your module::
144 cProfile.run('foo()')
146 (Use :mod:`profile` instead of :mod:`cProfile` if the latter is not available on
149 The above action would cause :func:`foo` to be run, and a series of informative
150 lines (the profile) to be printed. The above approach is most useful when
151 working with the interpreter. If you would like to save the results of a
152 profile into a file for later examination, you can supply a file name as the
153 second argument to the :func:`run` function::
156 cProfile.run('foo()', 'fooprof')
158 The file :file:`cProfile.py` can also be invoked as a script to profile another
159 script. For example::
161 python -m cProfile myscript.py
163 :file:`cProfile.py` accepts two optional arguments on the command line::
165 cProfile.py [-o output_file] [-s sort_order]
167 :option:`-s` only applies to standard output (:option:`-o` is not supplied).
168 Look in the :class:`Stats` documentation for valid sort values.
170 When you wish to review the profile, you should use the methods in the
171 :mod:`pstats` module. Typically you would load the statistics data as follows::
174 p = pstats.Stats('fooprof')
176 The class :class:`Stats` (the above code just created an instance of this class)
177 has a variety of methods for manipulating and printing the data that was just
178 read into ``p``. When you ran :func:`cProfile.run` above, what was printed was
179 the result of three method calls::
181 p.strip_dirs().sort_stats(-1).print_stats()
183 The first method removed the extraneous path from all the module names. The
184 second method sorted all the entries according to the standard module/line/name
185 string that is printed. The third method printed out all the statistics. You
186 might try the following sort calls:
188 .. (this is to comply with the semantics of the old profiler).
195 The first call will actually sort the list by function name, and the second call
196 will print out the statistics. The following are some interesting calls to
199 p.sort_stats('cumulative').print_stats(10)
201 This sorts the profile by cumulative time in a function, and then only prints
202 the ten most significant lines. If you want to understand what algorithms are
203 taking time, the above line is what you would use.
205 If you were looking to see what functions were looping a lot, and taking a lot
206 of time, you would do::
208 p.sort_stats('time').print_stats(10)
210 to sort according to time spent within each function, and then print the
211 statistics for the top ten functions.
215 p.sort_stats('file').print_stats('__init__')
217 This will sort all the statistics by file name, and then print out statistics
218 for only the class init methods (since they are spelled with ``__init__`` in
219 them). As one final example, you could try::
221 p.sort_stats('time', 'cum').print_stats(.5, 'init')
223 This line sorts statistics with a primary key of time, and a secondary key of
224 cumulative time, and then prints out some of the statistics. To be specific, the
225 list is first culled down to 50% (re: ``.5``) of its original size, then only
226 lines containing ``init`` are maintained, and that sub-sub-list is printed.
228 If you wondered what functions called the above functions, you could now (``p``
229 is still sorted according to the last criteria) do::
231 p.print_callers(.5, 'init')
233 and you would get a list of callers for each of the listed functions.
235 If you want more functionality, you're going to have to read the manual, or
236 guess what the following functions do::
241 Invoked as a script, the :mod:`pstats` module is a statistics browser for
242 reading and examining profile dumps. It has a simple line-oriented interface
243 (implemented using :mod:`cmd`) and interactive help.
246 .. _deterministic-profiling:
248 What Is Deterministic Profiling?
249 ================================
251 :dfn:`Deterministic profiling` is meant to reflect the fact that all *function
252 call*, *function return*, and *exception* events are monitored, and precise
253 timings are made for the intervals between these events (during which time the
254 user's code is executing). In contrast, :dfn:`statistical profiling` (which is
255 not done by this module) randomly samples the effective instruction pointer, and
256 deduces where time is being spent. The latter technique traditionally involves
257 less overhead (as the code does not need to be instrumented), but provides only
258 relative indications of where time is being spent.
260 In Python, since there is an interpreter active during execution, the presence
261 of instrumented code is not required to do deterministic profiling. Python
262 automatically provides a :dfn:`hook` (optional callback) for each event. In
263 addition, the interpreted nature of Python tends to add so much overhead to
264 execution, that deterministic profiling tends to only add small processing
265 overhead in typical applications. The result is that deterministic profiling is
266 not that expensive, yet provides extensive run time statistics about the
267 execution of a Python program.
269 Call count statistics can be used to identify bugs in code (surprising counts),
270 and to identify possible inline-expansion points (high call counts). Internal
271 time statistics can be used to identify "hot loops" that should be carefully
272 optimized. Cumulative time statistics should be used to identify high level
273 errors in the selection of algorithms. Note that the unusual handling of
274 cumulative times in this profiler allows statistics for recursive
275 implementations of algorithms to be directly compared to iterative
279 Reference Manual -- :mod:`profile` and :mod:`cProfile`
280 ======================================================
283 :synopsis: Python profiler
286 The primary entry point for the profiler is the global function
287 :func:`profile.run` (resp. :func:`cProfile.run`). It is typically used to create
288 any profile information. The reports are formatted and printed using methods of
289 the class :class:`pstats.Stats`. The following is a description of all of these
290 standard entry points and functions. For a more in-depth view of some of the
291 code, consider reading the later section on Profiler Extensions, which includes
292 discussion of how to derive "better" profilers from the classes presented, or
293 reading the source code for these modules.
296 .. function:: run(command[, filename])
298 This function takes a single argument that can be passed to the
299 :keyword:`exec` statement, and an optional file name. In all cases this
300 routine attempts to :keyword:`exec` its first argument, and gather profiling
301 statistics from the execution. If no file name is present, then this function
302 automatically prints a simple profiling report, sorted by the standard name
303 string (file/line/function-name) that is presented in each line. The
304 following is a typical output from such a call::
306 2706 function calls (2004 primitive calls) in 4.504 CPU seconds
308 Ordered by: standard name
310 ncalls tottime percall cumtime percall filename:lineno(function)
311 2 0.006 0.003 0.953 0.477 pobject.py:75(save_objects)
312 43/3 0.533 0.012 0.749 0.250 pobject.py:99(evaluate)
315 The first line indicates that 2706 calls were monitored. Of those calls, 2004
316 were :dfn:`primitive`. We define :dfn:`primitive` to mean that the call was not
317 induced via recursion. The next line: ``Ordered by: standard name``, indicates
318 that the text string in the far right column was used to sort the output. The
319 column headings include:
322 for the number of calls,
325 for the total time spent in the given function (and excluding time made in calls
329 is the quotient of ``tottime`` divided by ``ncalls``
332 is the total time spent in this and all subfunctions (from invocation till
333 exit). This figure is accurate *even* for recursive functions.
336 is the quotient of ``cumtime`` divided by primitive calls
338 filename:lineno(function)
339 provides the respective data of each function
341 When there are two numbers in the first column (for example, ``43/3``), then the
342 latter is the number of primitive calls, and the former is the actual number of
343 calls. Note that when the function does not recurse, these two values are the
344 same, and only the single figure is printed.
347 .. function:: runctx(command, globals, locals[, filename])
349 This function is similar to :func:`run`, with added arguments to supply the
350 globals and locals dictionaries for the *command* string.
352 Analysis of the profiler data is done using the :class:`Stats` class.
356 The :class:`Stats` class is defined in the :mod:`pstats` module.
360 :synopsis: Statistics object for use with the profiler.
363 .. class:: Stats(filename[, stream=sys.stdout[, ...]])
365 This class constructor creates an instance of a "statistics object" from a
366 *filename* (or set of filenames). :class:`Stats` objects are manipulated by
367 methods, in order to print useful reports. You may specify an alternate output
368 stream by giving the keyword argument, ``stream``.
370 The file selected by the above constructor must have been created by the
371 corresponding version of :mod:`profile` or :mod:`cProfile`. To be specific,
372 there is *no* file compatibility guaranteed with future versions of this
373 profiler, and there is no compatibility with files produced by other profilers.
374 If several files are provided, all the statistics for identical functions will
375 be coalesced, so that an overall view of several processes can be considered in
376 a single report. If additional files need to be combined with data in an
377 existing :class:`Stats` object, the :meth:`add` method can be used.
379 .. (such as the old system profiler).
381 .. versionchanged:: 2.5
382 The *stream* parameter was added.
387 The :class:`Stats` Class
388 ------------------------
390 :class:`Stats` objects have the following methods:
393 .. method:: Stats.strip_dirs()
395 This method for the :class:`Stats` class removes all leading path information
396 from file names. It is very useful in reducing the size of the printout to fit
397 within (close to) 80 columns. This method modifies the object, and the stripped
398 information is lost. After performing a strip operation, the object is
399 considered to have its entries in a "random" order, as it was just after object
400 initialization and loading. If :meth:`strip_dirs` causes two function names to
401 be indistinguishable (they are on the same line of the same filename, and have
402 the same function name), then the statistics for these two entries are
403 accumulated into a single entry.
406 .. method:: Stats.add(filename[, ...])
408 This method of the :class:`Stats` class accumulates additional profiling
409 information into the current profiling object. Its arguments should refer to
410 filenames created by the corresponding version of :func:`profile.run` or
411 :func:`cProfile.run`. Statistics for identically named (re: file, line, name)
412 functions are automatically accumulated into single function statistics.
415 .. method:: Stats.dump_stats(filename)
417 Save the data loaded into the :class:`Stats` object to a file named *filename*.
418 The file is created if it does not exist, and is overwritten if it already
419 exists. This is equivalent to the method of the same name on the
420 :class:`profile.Profile` and :class:`cProfile.Profile` classes.
422 .. versionadded:: 2.3
425 .. method:: Stats.sort_stats(key[, ...])
427 This method modifies the :class:`Stats` object by sorting it according to the
428 supplied criteria. The argument is typically a string identifying the basis of
429 a sort (example: ``'time'`` or ``'name'``).
431 When more than one key is provided, then additional keys are used as secondary
432 criteria when there is equality in all keys selected before them. For example,
433 ``sort_stats('name', 'file')`` will sort all the entries according to their
434 function name, and resolve all ties (identical function names) by sorting by
437 Abbreviations can be used for any key names, as long as the abbreviation is
438 unambiguous. The following are the keys currently defined:
440 +------------------+----------------------+
441 | Valid Arg | Meaning |
442 +==================+======================+
443 | ``'calls'`` | call count |
444 +------------------+----------------------+
445 | ``'cumulative'`` | cumulative time |
446 +------------------+----------------------+
447 | ``'file'`` | file name |
448 +------------------+----------------------+
449 | ``'module'`` | file name |
450 +------------------+----------------------+
451 | ``'pcalls'`` | primitive call count |
452 +------------------+----------------------+
453 | ``'line'`` | line number |
454 +------------------+----------------------+
455 | ``'name'`` | function name |
456 +------------------+----------------------+
457 | ``'nfl'`` | name/file/line |
458 +------------------+----------------------+
459 | ``'stdname'`` | standard name |
460 +------------------+----------------------+
461 | ``'time'`` | internal time |
462 +------------------+----------------------+
464 Note that all sorts on statistics are in descending order (placing most time
465 consuming items first), where as name, file, and line number searches are in
466 ascending order (alphabetical). The subtle distinction between ``'nfl'`` and
467 ``'stdname'`` is that the standard name is a sort of the name as printed, which
468 means that the embedded line numbers get compared in an odd way. For example,
469 lines 3, 20, and 40 would (if the file names were the same) appear in the string
470 order 20, 3 and 40. In contrast, ``'nfl'`` does a numeric compare of the line
471 numbers. In fact, ``sort_stats('nfl')`` is the same as ``sort_stats('name',
474 For backward-compatibility reasons, the numeric arguments ``-1``, ``0``, ``1``,
475 and ``2`` are permitted. They are interpreted as ``'stdname'``, ``'calls'``,
476 ``'time'``, and ``'cumulative'`` respectively. If this old style format
477 (numeric) is used, only one sort key (the numeric key) will be used, and
478 additional arguments will be silently ignored.
480 .. For compatibility with the old profiler,
483 .. method:: Stats.reverse_order()
485 This method for the :class:`Stats` class reverses the ordering of the basic list
486 within the object. Note that by default ascending vs descending order is
487 properly selected based on the sort key of choice.
489 .. This method is provided primarily for compatibility with the old profiler.
492 .. method:: Stats.print_stats([restriction, ...])
494 This method for the :class:`Stats` class prints out a report as described in the
495 :func:`profile.run` definition.
497 The order of the printing is based on the last :meth:`sort_stats` operation done
498 on the object (subject to caveats in :meth:`add` and :meth:`strip_dirs`).
500 The arguments provided (if any) can be used to limit the list down to the
501 significant entries. Initially, the list is taken to be the complete set of
502 profiled functions. Each restriction is either an integer (to select a count of
503 lines), or a decimal fraction between 0.0 and 1.0 inclusive (to select a
504 percentage of lines), or a regular expression (to pattern match the standard
505 name that is printed; as of Python 1.5b1, this uses the Perl-style regular
506 expression syntax defined by the :mod:`re` module). If several restrictions are
507 provided, then they are applied sequentially. For example::
509 print_stats(.1, 'foo:')
511 would first limit the printing to first 10% of list, and then only print
512 functions that were part of filename :file:`.\*foo:`. In contrast, the
515 print_stats('foo:', .1)
517 would limit the list to all functions having file names :file:`.\*foo:`, and
518 then proceed to only print the first 10% of them.
521 .. method:: Stats.print_callers([restriction, ...])
523 This method for the :class:`Stats` class prints a list of all functions that
524 called each function in the profiled database. The ordering is identical to
525 that provided by :meth:`print_stats`, and the definition of the restricting
526 argument is also identical. Each caller is reported on its own line. The
527 format differs slightly depending on the profiler that produced the stats:
529 * With :mod:`profile`, a number is shown in parentheses after each caller to
530 show how many times this specific call was made. For convenience, a second
531 non-parenthesized number repeats the cumulative time spent in the function
534 * With :mod:`cProfile`, each caller is preceeded by three numbers: the number of
535 times this specific call was made, and the total and cumulative times spent in
536 the current function while it was invoked by this specific caller.
539 .. method:: Stats.print_callees([restriction, ...])
541 This method for the :class:`Stats` class prints a list of all function that were
542 called by the indicated function. Aside from this reversal of direction of
543 calls (re: called vs was called by), the arguments and ordering are identical to
544 the :meth:`print_callers` method.
552 One limitation has to do with accuracy of timing information. There is a
553 fundamental problem with deterministic profilers involving accuracy. The most
554 obvious restriction is that the underlying "clock" is only ticking at a rate
555 (typically) of about .001 seconds. Hence no measurements will be more accurate
556 than the underlying clock. If enough measurements are taken, then the "error"
557 will tend to average out. Unfortunately, removing this first error induces a
558 second source of error.
560 The second problem is that it "takes a while" from when an event is dispatched
561 until the profiler's call to get the time actually *gets* the state of the
562 clock. Similarly, there is a certain lag when exiting the profiler event
563 handler from the time that the clock's value was obtained (and then squirreled
564 away), until the user's code is once again executing. As a result, functions
565 that are called many times, or call many functions, will typically accumulate
566 this error. The error that accumulates in this fashion is typically less than
567 the accuracy of the clock (less than one clock tick), but it *can* accumulate
568 and become very significant.
570 The problem is more important with :mod:`profile` than with the lower-overhead
571 :mod:`cProfile`. For this reason, :mod:`profile` provides a means of
572 calibrating itself for a given platform so that this error can be
573 probabilistically (on the average) removed. After the profiler is calibrated, it
574 will be more accurate (in a least square sense), but it will sometimes produce
575 negative numbers (when call counts are exceptionally low, and the gods of
576 probability work against you :-). ) Do *not* be alarmed by negative numbers in
577 the profile. They should *only* appear if you have calibrated your profiler,
578 and the results are actually better than without calibration.
581 .. _profile-calibration:
586 The profiler of the :mod:`profile` module subtracts a constant from each event
587 handling time to compensate for the overhead of calling the time function, and
588 socking away the results. By default, the constant is 0. The following
589 procedure can be used to obtain a better constant for a given platform (see
590 discussion in section Limitations above). ::
593 pr = profile.Profile()
595 print pr.calibrate(10000)
597 The method executes the number of Python calls given by the argument, directly
598 and again under the profiler, measuring the time for both. It then computes the
599 hidden overhead per profiler event, and returns that as a float. For example,
600 on an 800 MHz Pentium running Windows 2000, and using Python's time.clock() as
601 the timer, the magical number is about 12.5e-6.
603 The object of this exercise is to get a fairly consistent result. If your
604 computer is *very* fast, or your timer function has poor resolution, you might
605 have to pass 100000, or even 1000000, to get consistent results.
607 When you have a consistent answer, there are three ways you can use it: [#]_ ::
611 # 1. Apply computed bias to all Profile instances created hereafter.
612 profile.Profile.bias = your_computed_bias
614 # 2. Apply computed bias to a specific Profile instance.
615 pr = profile.Profile()
616 pr.bias = your_computed_bias
618 # 3. Specify computed bias in instance constructor.
619 pr = profile.Profile(bias=your_computed_bias)
621 If you have a choice, you are better off choosing a smaller constant, and then
622 your results will "less often" show up as negative in profile statistics.
625 .. _profiler-extensions:
627 Extensions --- Deriving Better Profilers
628 ========================================
630 The :class:`Profile` class of both modules, :mod:`profile` and :mod:`cProfile`,
631 were written so that derived classes could be developed to extend the profiler.
632 The details are not described here, as doing this successfully requires an
633 expert understanding of how the :class:`Profile` class works internally. Study
634 the source code of the module carefully if you want to pursue this.
636 If all you want to do is change how current time is determined (for example, to
637 force use of wall-clock time or elapsed process time), pass the timing function
638 you want to the :class:`Profile` class constructor::
640 pr = profile.Profile(your_time_func)
642 The resulting profiler will then call :func:`your_time_func`.
644 :class:`profile.Profile`
645 :func:`your_time_func` should return a single number, or a list of numbers whose
646 sum is the current time (like what :func:`os.times` returns). If the function
647 returns a single time number, or the list of returned numbers has length 2, then
648 you will get an especially fast version of the dispatch routine.
650 Be warned that you should calibrate the profiler class for the timer function
651 that you choose. For most machines, a timer that returns a lone integer value
652 will provide the best results in terms of low overhead during profiling.
653 (:func:`os.times` is *pretty* bad, as it returns a tuple of floating point
654 values). If you want to substitute a better timer in the cleanest fashion,
655 derive a class and hardwire a replacement dispatch method that best handles your
656 timer call, along with the appropriate calibration constant.
658 :class:`cProfile.Profile`
659 :func:`your_time_func` should return a single number. If it returns plain
660 integers, you can also invoke the class constructor with a second argument
661 specifying the real duration of one unit of time. For example, if
662 :func:`your_integer_time_func` returns times measured in thousands of seconds,
663 you would constuct the :class:`Profile` instance as follows::
665 pr = profile.Profile(your_integer_time_func, 0.001)
667 As the :mod:`cProfile.Profile` class cannot be calibrated, custom timer
668 functions should be used with care and should be as fast as possible. For the
669 best results with a custom timer, it might be necessary to hard-code it in the C
670 source of the internal :mod:`_lsprof` module.
672 .. rubric:: Footnotes
674 .. [#] Updated and converted to LaTeX by Guido van Rossum. Further updated by Armin
675 Rigo to integrate the documentation for the new :mod:`cProfile` module of Python
678 .. [#] Prior to Python 2.2, it was necessary to edit the profiler source code to embed
679 the bias as a literal number. You still can, but that method is no longer
680 described, because no longer needed.