A profiler is a program that describes the run time performance of a program, providing a variety of statistics. This documentation describes the profiler functionality provided in the modules cProfile, profile and pstats. This profiler provides deterministic profiling of Python programs. It also provides a series of report generation tools to allow users to rapidly examine the results of a profile operation.
The Python standard library provides three different profilers:
cProfile is recommended for most users; it’s a C extension with reasonable overhead that makes it suitable for profiling long-running programs. Based on lsprof, contributed by Brett Rosen and Ted Czotter.
New in version 2.5.
profile, a pure Python module whose interface is imitated by cProfile. Adds significant overhead to profiled programs. If you’re trying to extend the profiler in some way, the task might be easier with this module.
Changed in version 2.4: Now also reports the time spent in calls to built-in functions and methods.
hotshot was an experimental C module that focused on minimizing the overhead of profiling, at the expense of longer data post-processing times. It is no longer maintained and may be dropped in a future version of Python.
Changed in version 2.5: The results should be more meaningful than in the past: the timing core contained a critical bug.
The profile and cProfile modules export the same interface, so they are mostly interchangeable; cProfile has a much lower overhead but is newer and might not be available on all systems. cProfile is really a compatibility layer on top of the internal _lsprof module. The hotshot module is reserved for specialized usage.
This section is provided for users that “don’t want to read the manual.” It provides a very brief overview, and allows a user to rapidly perform profiling on an existing application.
To profile an application with a main entry point of foo(), you would add the following to your module:
import cProfile
cProfile.run('foo()')
(Use profile instead of cProfile if the latter is not available on your system.)
The above action would cause foo() to be run, and a series of informative lines (the profile) to be printed. The above approach is most useful when working with the interpreter. If you would like to save the results of a profile into a file for later examination, you can supply a file name as the second argument to the run() function:
import cProfile
cProfile.run('foo()', 'fooprof')
The file cProfile.py can also be invoked as a script to profile another script. For example:
python -m cProfile myscript.py
cProfile.py accepts two optional arguments on the command line:
cProfile.py [-o output_file] [-s sort_order]
-s only applies to standard output (-o is not supplied). Look in the Stats documentation for valid sort values.
When you wish to review the profile, you should use the methods in the pstats module. Typically you would load the statistics data as follows:
import pstats
p = pstats.Stats('fooprof')
The class Stats (the above code just created an instance of this class) has a variety of methods for manipulating and printing the data that was just read into p. When you ran cProfile.run() above, what was printed was the result of three method calls:
p.strip_dirs().sort_stats(-1).print_stats()
The first method removed the extraneous path from all the module names. The second method sorted all the entries according to the standard module/line/name string that is printed. The third method printed out all the statistics. You might try the following sort calls:
p.sort_stats('name')
p.print_stats()
The first call will actually sort the list by function name, and the second call will print out the statistics. The following are some interesting calls to experiment with:
p.sort_stats('cumulative').print_stats(10)
This sorts the profile by cumulative time in a function, and then only prints the ten most significant lines. If you want to understand what algorithms are taking time, the above line is what you would use.
If you were looking to see what functions were looping a lot, and taking a lot of time, you would do:
p.sort_stats('time').print_stats(10)
to sort according to time spent within each function, and then print the statistics for the top ten functions.
You might also try:
p.sort_stats('file').print_stats('__init__')
This will sort all the statistics by file name, and then print out statistics for only the class init methods (since they are spelled with __init__ in them). As one final example, you could try:
p.sort_stats('time', 'cum').print_stats(.5, 'init')
This line sorts statistics with a primary key of time, and a secondary key of cumulative time, and then prints out some of the statistics. To be specific, the list is first culled down to 50% (re: .5) of its original size, then only lines containing init are maintained, and that sub-sub-list is printed.
If you wondered what functions called the above functions, you could now (p is still sorted according to the last criteria) do:
p.print_callers(.5, 'init')
and you would get a list of callers for each of the listed functions.
If you want more functionality, you’re going to have to read the manual, or guess what the following functions do:
p.print_callees()
p.add('fooprof')
Invoked as a script, the pstats module is a statistics browser for reading and examining profile dumps. It has a simple line-oriented interface (implemented using cmd) and interactive help.
Deterministic profiling is meant to reflect the fact that all function call, function return, and exception events are monitored, and precise timings are made for the intervals between these events (during which time the user’s code is executing). In contrast, statistical profiling (which is not done by this module) randomly samples the effective instruction pointer, and deduces where time is being spent. The latter technique traditionally involves less overhead (as the code does not need to be instrumented), but provides only relative indications of where time is being spent.
In Python, since there is an interpreter active during execution, the presence of instrumented code is not required to do deterministic profiling. Python automatically provides a hook (optional callback) for each event. In addition, the interpreted nature of Python tends to add so much overhead to execution, that deterministic profiling tends to only add small processing overhead in typical applications. The result is that deterministic profiling is not that expensive, yet provides extensive run time statistics about the execution of a Python program.
Call count statistics can be used to identify bugs in code (surprising counts), and to identify possible inline-expansion points (high call counts). Internal time statistics can be used to identify “hot loops” that should be carefully optimized. Cumulative time statistics should be used to identify high level errors in the selection of algorithms. Note that the unusual handling of cumulative times in this profiler allows statistics for recursive implementations of algorithms to be directly compared to iterative implementations.
The primary entry point for the profiler is the global function profile.run() (resp. cProfile.run()). It is typically used to create any profile information. The reports are formatted and printed using methods of the class pstats.Stats. The following is a description of all of these standard entry points and functions. For a more in-depth view of some of the code, consider reading the later section on Profiler Extensions, which includes discussion of how to derive “better” profilers from the classes presented, or reading the source code for these modules.
This function takes a single argument that can be passed to the exec statement, and an optional file name. In all cases this routine attempts to exec its first argument, and gather profiling statistics from the execution. If no file name is present, then this function automatically prints a simple profiling report, sorted by the standard name string (file/line/function-name) that is presented in each line. The following is a typical output from such a call:
2706 function calls (2004 primitive calls) in 4.504 CPU seconds
Ordered by: standard name
ncalls tottime percall cumtime percall filename:lineno(function)
2 0.006 0.003 0.953 0.477 pobject.py:75(save_objects)
43/3 0.533 0.012 0.749 0.250 pobject.py:99(evaluate)
...
The first line indicates that 2706 calls were monitored. Of those calls, 2004 were primitive. We define primitive to mean that the call was not induced via recursion. The next line: Ordered by: standard name, indicates that the text string in the far right column was used to sort the output. The column headings include:
When there are two numbers in the first column (for example, 43/3), then the latter is the number of primitive calls, and the former is the actual number of calls. Note that when the function does not recurse, these two values are the same, and only the single figure is printed.
Analysis of the profiler data is done using the Stats class.
Note
The Stats class is defined in the pstats module.
This class constructor creates an instance of a “statistics object” from a filename (or set of filenames). Stats objects are manipulated by methods, in order to print useful reports. You may specify an alternate output stream by giving the keyword argument, stream.
The file selected by the above constructor must have been created by the corresponding version of profile or cProfile. To be specific, there is no file compatibility guaranteed with future versions of this profiler, and there is no compatibility with files produced by other profilers. If several files are provided, all the statistics for identical functions will be coalesced, so that an overall view of several processes can be considered in a single report. If additional files need to be combined with data in an existing Stats object, the add() method can be used.
Changed in version 2.5: The stream parameter was added.
Stats objects have the following methods:
Save the data loaded into the Stats object to a file named filename. The file is created if it does not exist, and is overwritten if it already exists. This is equivalent to the method of the same name on the profile.Profile and cProfile.Profile classes.
New in version 2.3.
This method modifies the Stats object by sorting it according to the supplied criteria. The argument is typically a string identifying the basis of a sort (example: 'time' or 'name').
When more than one key is provided, then additional keys are used as secondary criteria when there is equality in all keys selected before them. For example, sort_stats('name', 'file') will sort all the entries according to their function name, and resolve all ties (identical function names) by sorting by file name.
Abbreviations can be used for any key names, as long as the abbreviation is unambiguous. The following are the keys currently defined:
| Valid Arg | Meaning |
|---|---|
| 'calls' | call count |
| 'cumulative' | cumulative time |
| 'file' | file name |
| 'module' | file name |
| 'pcalls' | primitive call count |
| 'line' | line number |
| 'name' | function name |
| 'nfl' | name/file/line |
| 'stdname' | standard name |
| 'time' | internal time |
Note that all sorts on statistics are in descending order (placing most time consuming items first), where as name, file, and line number searches are in ascending order (alphabetical). The subtle distinction between 'nfl' and 'stdname' is that the standard name is a sort of the name as printed, which means that the embedded line numbers get compared in an odd way. For example, lines 3, 20, and 40 would (if the file names were the same) appear in the string order 20, 3 and 40. In contrast, 'nfl' does a numeric compare of the line numbers. In fact, sort_stats('nfl') is the same as sort_stats('name', 'file', 'line').
For backward-compatibility reasons, the numeric arguments -1, 0, 1, and 2 are permitted. They are interpreted as 'stdname', 'calls', 'time', and 'cumulative' respectively. If this old style format (numeric) is used, only one sort key (the numeric key) will be used, and additional arguments will be silently ignored.
This method for the Stats class prints out a report as described in the profile.run() definition.
The order of the printing is based on the last sort_stats() operation done on the object (subject to caveats in add() and strip_dirs()).
The arguments provided (if any) can be used to limit the list down to the significant entries. Initially, the list is taken to be the complete set of profiled functions. Each restriction is either an integer (to select a count of lines), or a decimal fraction between 0.0 and 1.0 inclusive (to select a percentage of lines), or a regular expression (to pattern match the standard name that is printed; as of Python 1.5b1, this uses the Perl-style regular expression syntax defined by the re module). If several restrictions are provided, then they are applied sequentially. For example:
print_stats(.1, 'foo:')
would first limit the printing to first 10% of list, and then only print functions that were part of filename .*foo:. In contrast, the command:
print_stats('foo:', .1)
would limit the list to all functions having file names .*foo:, and then proceed to only print the first 10% of them.
This method for the Stats class prints a list of all functions that called each function in the profiled database. The ordering is identical to that provided by print_stats(), and the definition of the restricting argument is also identical. Each caller is reported on its own line. The format differs slightly depending on the profiler that produced the stats:
One limitation has to do with accuracy of timing information. There is a fundamental problem with deterministic profilers involving accuracy. The most obvious restriction is that the underlying “clock” is only ticking at a rate (typically) of about .001 seconds. Hence no measurements will be more accurate than the underlying clock. If enough measurements are taken, then the “error” will tend to average out. Unfortunately, removing this first error induces a second source of error.
The second problem is that it “takes a while” from when an event is dispatched until the profiler’s call to get the time actually gets the state of the clock. Similarly, there is a certain lag when exiting the profiler event handler from the time that the clock’s value was obtained (and then squirreled away), until the user’s code is once again executing. As a result, functions that are called many times, or call many functions, will typically accumulate this error. The error that accumulates in this fashion is typically less than the accuracy of the clock (less than one clock tick), but it can accumulate and become very significant.
The problem is more important with profile than with the lower-overhead cProfile. For this reason, profile provides a means of calibrating itself for a given platform so that this error can be probabilistically (on the average) removed. After the profiler is calibrated, it will be more accurate (in a least square sense), but it will sometimes produce negative numbers (when call counts are exceptionally low, and the gods of probability work against you :-). ) Do not be alarmed by negative numbers in the profile. They should only appear if you have calibrated your profiler, and the results are actually better than without calibration.
The profiler of the profile module subtracts a constant from each event handling time to compensate for the overhead of calling the time function, and socking away the results. By default, the constant is 0. The following procedure can be used to obtain a better constant for a given platform (see discussion in section Limitations above).
import profile
pr = profile.Profile()
for i in range(5):
print pr.calibrate(10000)
The method executes the number of Python calls given by the argument, directly and again under the profiler, measuring the time for both. It then computes the hidden overhead per profiler event, and returns that as a float. For example, on an 800 MHz Pentium running Windows 2000, and using Python’s time.clock() as the timer, the magical number is about 12.5e-6.
The object of this exercise is to get a fairly consistent result. If your computer is very fast, or your timer function has poor resolution, you might have to pass 100000, or even 1000000, to get consistent results.
When you have a consistent answer, there are three ways you can use it: [1]
import profile
# 1. Apply computed bias to all Profile instances created hereafter.
profile.Profile.bias = your_computed_bias
# 2. Apply computed bias to a specific Profile instance.
pr = profile.Profile()
pr.bias = your_computed_bias
# 3. Specify computed bias in instance constructor.
pr = profile.Profile(bias=your_computed_bias)
If you have a choice, you are better off choosing a smaller constant, and then your results will “less often” show up as negative in profile statistics.
The Profile class of both modules, profile and cProfile, were written so that derived classes could be developed to extend the profiler. The details are not described here, as doing this successfully requires an expert understanding of how the Profile class works internally. Study the source code of the module carefully if you want to pursue this.
If all you want to do is change how current time is determined (for example, to force use of wall-clock time or elapsed process time), pass the timing function you want to the Profile class constructor:
pr = profile.Profile(your_time_func)
The resulting profiler will then call your_time_func().
your_time_func() should return a single number, or a list of numbers whose sum is the current time (like what os.times() returns). If the function returns a single time number, or the list of returned numbers has length 2, then you will get an especially fast version of the dispatch routine.
Be warned that you should calibrate the profiler class for the timer function that you choose. For most machines, a timer that returns a lone integer value will provide the best results in terms of low overhead during profiling. (os.times() is pretty bad, as it returns a tuple of floating point values). If you want to substitute a better timer in the cleanest fashion, derive a class and hardwire a replacement dispatch method that best handles your timer call, along with the appropriate calibration constant.
your_time_func() should return a single number. If it returns plain integers, you can also invoke the class constructor with a second argument specifying the real duration of one unit of time. For example, if your_integer_time_func() returns times measured in thousands of seconds, you would constuct the Profile instance as follows:
pr = profile.Profile(your_integer_time_func, 0.001)
As the cProfile.Profile class cannot be calibrated, custom timer functions should be used with care and should be as fast as possible. For the best results with a custom timer, it might be necessary to hard-code it in the C source of the internal _lsprof module.
Footnotes
| [1] | Updated and converted to LaTeX by Guido van Rossum. Further updated by Armin Rigo to integrate the documentation for the new cProfile module of Python 2.5. |
| [2] | Prior to Python 2.2, it was necessary to edit the profiler source code to embed the bias as a literal number. You still can, but that method is no longer described, because no longer needed. |