line_profiler and kernprof
NOTICE: This is the official line_profiler repository. The most recent version of line-profiler on pypi points to this repo. The the original line_profiler package by @rkern is currently unmaintained. This fork seeks to simply maintain the original code so it continues to work in new versions of Python.
line_profiler is a module for doing line-by-line profiling of functions. kernprof is a convenient script for running either line_profiler or the Python standard library's cProfile or profile modules, depending on what is available.
They are available under a BSD license.
Releases of line_profiler can be installed using pip:
$ pip install line_profiler
Source releases and any binaries can be downloaded from the PyPI link.
To check out the development sources, you can use Git:
$ git clone https://github.com/pyutils/line_profiler.git
You may also download source tarballs of any snapshot from that URL.
Source releases will require a C compiler in order to build line_profiler. In addition, git checkouts will also require Cython >= 0.10. Source releases on PyPI should contain the pregenerated C sources, so Cython should not be required in that case.
kernprof is a single-file pure Python script and does not require a compiler. If you wish to use it to run cProfile and not line-by-line profiling, you may copy it to a directory on your PATH manually and avoid trying to build any C extensions.
The current profiling tools supported in Python 2.7 and later only time function calls. This is a good first step for locating hotspots in one's program and is frequently all one needs to do to optimize the program. However, sometimes the cause of the hotspot is actually a single line in the function, and that line may not be obvious from just reading the source code. These cases are particularly frequent in scientific computing. Functions tend to be larger (sometimes because of legitimate algorithmic complexity, sometimes because the programmer is still trying to write FORTRAN code), and a single statement without function calls can trigger lots of computation when using libraries like numpy. cProfile only times explicit function calls, not special methods called because of syntax. Consequently, a relatively slow numpy operation on large arrays like this,
a[large_index_array] = some_other_large_array
is a hotspot that never gets broken out by cProfile because there is no explicit function call in that statement.
LineProfiler can be given functions to profile, and it will time the execution of each individual line inside those functions. In a typical workflow, one only cares about line timings of a few functions because wading through the results of timing every single line of code would be overwhelming. However, LineProfiler does need to be explicitly told what functions to profile. The easiest way to get started is to use the kernprof script.
$ kernprof -l script_to_profile.py
kernprof will create an instance of LineProfiler and insert it into the __builtins__ namespace with the name profile. It has been written to be used as a decorator, so in your script, you decorate the functions you want to profile with @profile.
@profile def slow_function(a, b, c): ...
The default behavior of kernprof is to put the results into a binary file script_to_profile.py.lprof . You can tell kernprof to immediately view the formatted results at the terminal with the [-v/--view] option. Otherwise, you can view the results later like so:
$ python -m line_profiler script_to_profile.py.lprof
For example, here are the results of profiling a single function from a decorated version of the pystone.py benchmark (the first two lines are output from pystone.py, not kernprof):
Pystone(1.1) time for 50000 passes = 2.48 This machine benchmarks at 20161.3 pystones/second Wrote profile results to pystone.py.lprof Timer unit: 1e-06 s File: pystone.py Function: Proc2 at line 149 Total time: 0.606656 s Line # Hits Time Per Hit % Time Line Contents ============================================================== 149 @profile 150 def Proc2(IntParIO): 151 50000 82003 1.6 13.5 IntLoc = IntParIO + 10 152 50000 63162 1.3 10.4 while 1: 153 50000 69065 1.4 11.4 if Char1Glob == 'A': 154 50000 66354 1.3 10.9 IntLoc = IntLoc - 1 155 50000 67263 1.3 11.1 IntParIO = IntLoc - IntGlob 156 50000 65494 1.3 10.8 EnumLoc = Ident1 157 50000 68001 1.4 11.2 if EnumLoc == Ident1: 158 50000 63739 1.3 10.5 break 159 50000 61575 1.2 10.1 return IntParIO
The source code of the function is printed with the timing information for each line. There are six columns of information.
- Line #: The line number in the file.
- Hits: The number of times that line was executed.
- Time: The total amount of time spent executing the line in the timer's units. In the header information before the tables, you will see a line "Timer unit:" giving the conversion factor to seconds. It may be different on different systems.
- Per Hit: The average amount of time spent executing the line once in the timer's units.
- % Time: The percentage of time spent on that line relative to the total amount of recorded time spent in the function.
- Line Contents: The actual source code. Note that this is always read from disk when the formatted results are viewed, not when the code was executed. If you have edited the file in the meantime, the lines will not match up, and the formatter may not even be able to locate the function for display.
If you are using IPython, there is an implementation of an %lprun magic command which will let you specify functions to profile and a statement to execute. It will also add its LineProfiler instance into the __builtins__, but typically, you would not use it like that.
For IPython 0.11+, you can install it by editing the IPython configuration file ~/.ipython/profile_default/ipython_config.py to add the 'line_profiler' item to the extensions list:
c.TerminalIPythonApp.extensions = [ 'line_profiler', ]
To get usage help for %lprun, use the standard IPython help mechanism:
In : %lprun?
These two methods are expected to be the most frequent user-level ways of using LineProfiler and will usually be the easiest. However, if you are building other tools with LineProfiler, you will need to use the API. There are two ways to inform LineProfiler of functions to profile: you can pass them as arguments to the constructor or use the add_function(f) method after instantiation.
profile = LineProfiler(f, g) profile.add_function(h)
LineProfiler has the same run(), runctx(), and runcall() methods as cProfile.Profile as well as enable() and disable(). It should be noted, though, that enable() and disable() are not entirely safe when nested. Nesting is common when using LineProfiler as a decorator. In order to support nesting, use enable_by_count() and disable_by_count(). These functions will increment and decrement a counter and only actually enable or disable the profiler when the count transitions from or to 0.
After profiling, the dump_stats(filename) method will pickle the results out to the given file. print_stats([stream]) will print the formatted results to sys.stdout or whatever stream you specify. get_stats() will return LineStats object, which just holds two attributes: a dictionary containing the results and the timer unit.
kernprof also works with cProfile, its third-party incarnation lsprof, or the pure-Python profile module depending on what is available. It has a few main features:
- Encapsulation of profiling concerns. You do not have to modify your script in order to initiate profiling and save the results. Unless if you want to use the advanced __builtins__ features, of course.
- Robust script execution. Many scripts require things like __name__, __file__, and sys.path to be set relative to it. A naive approach at encapsulation would just use execfile(), but many scripts which rely on that information will fail. kernprof will set those variables correctly before executing the script.
- Easy executable location. If you are profiling an application installed on your PATH, you can just give the name of the executable. If kernprof does not find the given script in the current directory, it will search your PATH for it.
- Inserting the profiler into __builtins__. Sometimes, you just want to profile a small part of your code. With the [-b/--builtin] argument, the Profiler will be instantiated and inserted into your __builtins__ with the name "profile". Like LineProfiler, it may be used as a decorator, or enabled/disabled with enable_by_count() and disable_by_count(), or even as a context manager with the "with profile:" statement.
- Pre-profiling setup. With the [-s/--setup] option, you can provide a script which will be executed without profiling before executing the main script. This is typically useful for cases where imports of large libraries like wxPython or VTK are interfering with your results. If you can modify your source code, the __builtins__ approach may be easier.
The results of profile script_to_profile.py will be written to script_to_profile.py.prof by default. It will be a typical marshalled file that can be read with pstats.Stats(). They may be interactively viewed with the command:
$ python -m pstats script_to_profile.py.prof
Why the name "kernprof"?
I didn't manage to come up with a meaningful name, so I named it after myself.
Why not use hotshot instead of line_profile?
hotshot can do line-by-line timings, too. However, it is deprecated and may disappear from the standard library. Also, it can take a long time to process the results while I want quick turnaround in my workflows. hotshot pays this processing time in order to make itself minimally intrusive to the code it is profiling. Code that does network operations, for example, may even go down different code paths if profiling slows down execution too much. For my use cases, and I think those of many other people, their line-by-line profiling is not affected much by this concern.
Why not allow using hotshot from kernprof.py?
I don't use hotshot, myself. I will accept contributions in this vein, though.
The line-by-line timings don't add up when one profiled function calls another. What's up with that?
Let's say you have function F() calling function G(), and you are using LineProfiler on both. The total time reported for G() is less than the time reported on the line in F() that calls G(). The reason is that I'm being reasonably clever (and possibly too clever) in recording the times. Basically, I try to prevent recording the time spent inside LineProfiler doing all of the bookkeeping for each line. Each time Python's tracing facility issues a line event (which happens just before a line actually gets executed), LineProfiler will find two timestamps, one at the beginning before it does anything (t_begin) and one as close to the end as possible (t_end). Almost all of the overhead of LineProfiler's data structures happens in between these two times.
When a line event comes in, LineProfiler finds the function it belongs to. If it's the first line in the function, we record the line number and t_end associated with the function. The next time we see a line event belonging to that function, we take t_begin of the new event and subtract the old t_end from it to find the amount of time spent in the old line. Then we record the new t_end as the active line for this function. This way, we are removing most of LineProfiler's overhead from the results. Well almost. When one profiled function F calls another profiled function G, the line in F that calls G basically records the total time spent executing the line, which includes the time spent inside the profiler while inside G.
The first time this question was asked, the questioner had the G() function call as part of a larger expression, and he wanted to try to estimate how much time was being spent in the function as opposed to the rest of the expression. My response was that, even if I could remove the effect, it might still be misleading. G() might be called elsewhere, not just from the relevant line in F(). The workaround would be to modify the code to split it up into two lines, one which just assigns the result of G() to a temporary variable and the other with the rest of the expression.
I am open to suggestions on how to make this more robust. Or simple admonitions against trying to be clever.
Why do my list comprehensions have so many hits when I use the LineProfiler?
LineProfiler records the line with the list comprehension once for each iteration of the list comprehension.
Why is kernprof distributed with line_profiler? It works with just cProfile, right?
Partly because kernprof.py is essential to using line_profiler effectively, but mostly because I'm lazy and don't want to maintain the overhead of two projects for modules as small as these. However, kernprof.py is a standalone, pure Python script that can be used to do function profiling with just the Python standard library. You may grab it and install it by itself without line_profiler.
Do I need a C compiler to build line_profiler? kernprof.py?
You do need a C compiler for line_profiler. kernprof.py is a pure Python script and can be installed separately, though.
Do I need Cython to build line_profiler?
You should not have to if you are building from a released source tarball. It should contain the generated C sources already. If you are running into problems, that may be a bug; let me know. If you are building from a git checkout or snapshot, you will need Cython to generate the C sources. You will probably need version 0.10 or higher. There is a bug in some earlier versions in how it handles NULL PyObject* pointers.
As of version
3.0.0manylinux wheels containing the binaries are available on pypi. Work is still needed to publish osx and win32 wheels. (PRs for this would be helpful!)
What version of Python do I need?
Both line_profiler and kernprof have been tested with Python 3.5-3.9. Older versions of line_profiler support older versions of Python.
cProfile uses a neat "rotating trees" data structure to minimize the overhead of looking up and recording entries. LineProfiler uses Python dictionaries and extension objects thanks to Cython. This mostly started out as a prototype that I wanted to play with as quickly as possible, so I passed on stealing the rotating trees for now. As usual, I got it working, and it seems to have acceptable performance, so I am much less motivated to use a different strategy now. Maybe later. Contributions accepted!
Bugs and pull requested can be submitted on GitHub.