| 1 |
.. highlight:: shell-session |
| 2 |
|
| 3 |
.. _perf_profiling: |
| 4 |
|
| 5 |
============================================== |
| 6 |
Python support for the Linux ``perf`` profiler |
| 7 |
============================================== |
| 8 |
|
| 9 |
:author: Pablo Galindo |
| 10 |
|
| 11 |
`The Linux perf profiler <https://perf.wiki.kernel.org>`_ |
| 12 |
is a very powerful tool that allows you to profile and obtain |
| 13 |
information about the performance of your application. |
| 14 |
``perf`` also has a very vibrant ecosystem of tools |
| 15 |
that aid with the analysis of the data that it produces. |
| 16 |
|
| 17 |
The main problem with using the ``perf`` profiler with Python applications is that |
| 18 |
``perf`` only gets information about native symbols, that is, the names of |
| 19 |
functions and procedures written in C. This means that the names and file names |
| 20 |
of Python functions in your code will not appear in the output of ``perf``. |
| 21 |
|
| 22 |
Since Python 3.12, the interpreter can run in a special mode that allows Python |
| 23 |
functions to appear in the output of the ``perf`` profiler. When this mode is |
| 24 |
enabled, the interpreter will interpose a small piece of code compiled on the |
| 25 |
fly before the execution of every Python function and it will teach ``perf`` the |
| 26 |
relationship between this piece of code and the associated Python function using |
| 27 |
:doc:`perf map files <../c-api/perfmaps>`. |
| 28 |
|
| 29 |
.. note:: |
| 30 |
|
| 31 |
Support for the ``perf`` profiler is currently only available for Linux on |
| 32 |
select architectures. Check the output of the ``configure`` build step or |
| 33 |
check the output of ``python -m sysconfig | grep HAVE_PERF_TRAMPOLINE`` |
| 34 |
to see if your system is supported. |
| 35 |
|
| 36 |
For example, consider the following script: |
| 37 |
|
| 38 |
.. code-block:: python |
| 39 |
|
| 40 |
def foo(n): |
| 41 |
result = 0 |
| 42 |
for _ in range(n): |
| 43 |
result += 1 |
| 44 |
return result |
| 45 |
|
| 46 |
def bar(n): |
| 47 |
foo(n) |
| 48 |
|
| 49 |
def baz(n): |
| 50 |
bar(n) |
| 51 |
|
| 52 |
if __name__ == "__main__": |
| 53 |
baz(1000000) |
| 54 |
|
| 55 |
We can run ``perf`` to sample CPU stack traces at 9999 hertz:: |
| 56 |
|
| 57 |
$ perf record -F 9999 -g -o perf.data python my_script.py |
| 58 |
|
| 59 |
Then we can use ``perf report`` to analyze the data: |
| 60 |
|
| 61 |
.. code-block:: shell-session |
| 62 |
|
| 63 |
$ perf report --stdio -n -g |
| 64 |
|
| 65 |
# Children Self Samples Command Shared Object Symbol |
| 66 |
# ........ ........ ............ .......... .................. .......................................... |
| 67 |
# |
| 68 |
91.08% 0.00% 0 python.exe python.exe [.] _start |
| 69 |
| |
| 70 |
---_start |
| 71 |
| |
| 72 |
--90.71%--__libc_start_main |
| 73 |
Py_BytesMain |
| 74 |
| |
| 75 |
|--56.88%--pymain_run_python.constprop.0 |
| 76 |
| | |
| 77 |
| |--56.13%--_PyRun_AnyFileObject |
| 78 |
| | _PyRun_SimpleFileObject |
| 79 |
| | | |
| 80 |
| | |--55.02%--run_mod |
| 81 |
| | | | |
| 82 |
| | | --54.65%--PyEval_EvalCode |
| 83 |
| | | _PyEval_EvalFrameDefault |
| 84 |
| | | PyObject_Vectorcall |
| 85 |
| | | _PyEval_Vector |
| 86 |
| | | _PyEval_EvalFrameDefault |
| 87 |
| | | PyObject_Vectorcall |
| 88 |
| | | _PyEval_Vector |
| 89 |
| | | _PyEval_EvalFrameDefault |
| 90 |
| | | PyObject_Vectorcall |
| 91 |
| | | _PyEval_Vector |
| 92 |
| | | | |
| 93 |
| | | |--51.67%--_PyEval_EvalFrameDefault |
| 94 |
| | | | | |
| 95 |
| | | | |--11.52%--_PyLong_Add |
| 96 |
| | | | | | |
| 97 |
| | | | | |--2.97%--_PyObject_Malloc |
| 98 |
... |
| 99 |
|
| 100 |
As you can see, the Python functions are not shown in the output, only ``_PyEval_EvalFrameDefault`` |
| 101 |
(the function that evaluates the Python bytecode) shows up. Unfortunately that's not very useful because all Python |
| 102 |
functions use the same C function to evaluate bytecode so we cannot know which Python function corresponds to which |
| 103 |
bytecode-evaluating function. |
| 104 |
|
| 105 |
Instead, if we run the same experiment with ``perf`` support enabled we get: |
| 106 |
|
| 107 |
.. code-block:: shell-session |
| 108 |
|
| 109 |
$ perf report --stdio -n -g |
| 110 |
|
| 111 |
# Children Self Samples Command Shared Object Symbol |
| 112 |
# ........ ........ ............ .......... .................. ..................................................................... |
| 113 |
# |
| 114 |
90.58% 0.36% 1 python.exe python.exe [.] _start |
| 115 |
| |
| 116 |
---_start |
| 117 |
| |
| 118 |
--89.86%--__libc_start_main |
| 119 |
Py_BytesMain |
| 120 |
| |
| 121 |
|--55.43%--pymain_run_python.constprop.0 |
| 122 |
| | |
| 123 |
| |--54.71%--_PyRun_AnyFileObject |
| 124 |
| | _PyRun_SimpleFileObject |
| 125 |
| | | |
| 126 |
| | |--53.62%--run_mod |
| 127 |
| | | | |
| 128 |
| | | --53.26%--PyEval_EvalCode |
| 129 |
| | | py::<module>:/src/script.py |
| 130 |
| | | _PyEval_EvalFrameDefault |
| 131 |
| | | PyObject_Vectorcall |
| 132 |
| | | _PyEval_Vector |
| 133 |
| | | py::baz:/src/script.py |
| 134 |
| | | _PyEval_EvalFrameDefault |
| 135 |
| | | PyObject_Vectorcall |
| 136 |
| | | _PyEval_Vector |
| 137 |
| | | py::bar:/src/script.py |
| 138 |
| | | _PyEval_EvalFrameDefault |
| 139 |
| | | PyObject_Vectorcall |
| 140 |
| | | _PyEval_Vector |
| 141 |
| | | py::foo:/src/script.py |
| 142 |
| | | | |
| 143 |
| | | |--51.81%--_PyEval_EvalFrameDefault |
| 144 |
| | | | | |
| 145 |
| | | | |--13.77%--_PyLong_Add |
| 146 |
| | | | | | |
| 147 |
| | | | | |--3.26%--_PyObject_Malloc |
| 148 |
|
| 149 |
|
| 150 |
|
| 151 |
How to enable ``perf`` profiling support |
| 152 |
---------------------------------------- |
| 153 |
|
| 154 |
``perf`` profiling support can be enabled either from the start using |
| 155 |
the environment variable :envvar:`PYTHONPERFSUPPORT` or the |
| 156 |
:option:`-X perf <-X>` option, |
| 157 |
or dynamically using :func:`sys.activate_stack_trampoline` and |
| 158 |
:func:`sys.deactivate_stack_trampoline`. |
| 159 |
|
| 160 |
The :mod:`!sys` functions take precedence over the :option:`!-X` option, |
| 161 |
the :option:`!-X` option takes precedence over the environment variable. |
| 162 |
|
| 163 |
Example, using the environment variable:: |
| 164 |
|
| 165 |
$ PYTHONPERFSUPPORT=1 perf record -F 9999 -g -o perf.data python my_script.py |
| 166 |
$ perf report -g -i perf.data |
| 167 |
|
| 168 |
Example, using the :option:`!-X` option:: |
| 169 |
|
| 170 |
$ perf record -F 9999 -g -o perf.data python -X perf my_script.py |
| 171 |
$ perf report -g -i perf.data |
| 172 |
|
| 173 |
Example, using the :mod:`sys` APIs in file :file:`example.py`: |
| 174 |
|
| 175 |
.. code-block:: python |
| 176 |
|
| 177 |
import sys |
| 178 |
|
| 179 |
sys.activate_stack_trampoline("perf") |
| 180 |
do_profiled_stuff() |
| 181 |
sys.deactivate_stack_trampoline() |
| 182 |
|
| 183 |
non_profiled_stuff() |
| 184 |
|
| 185 |
...then:: |
| 186 |
|
| 187 |
$ perf record -F 9999 -g -o perf.data python ./example.py |
| 188 |
$ perf report -g -i perf.data |
| 189 |
|
| 190 |
|
| 191 |
How to obtain the best results |
| 192 |
------------------------------ |
| 193 |
|
| 194 |
For best results, Python should be compiled with |
| 195 |
``CFLAGS="-fno-omit-frame-pointer -mno-omit-leaf-frame-pointer"`` as this allows |
| 196 |
profilers to unwind using only the frame pointer and not on DWARF debug |
| 197 |
information. This is because as the code that is interposed to allow ``perf`` |
| 198 |
support is dynamically generated it doesn't have any DWARF debugging information |
| 199 |
available. |
| 200 |
|
| 201 |
You can check if your system has been compiled with this flag by running:: |
| 202 |
|
| 203 |
$ python -m sysconfig | grep 'no-omit-frame-pointer' |
| 204 |
|
| 205 |
If you don't see any output it means that your interpreter has not been compiled with |
| 206 |
frame pointers and therefore it may not be able to show Python functions in the output |
| 207 |
of ``perf``. |
| 208 |
|
| 209 |
|
| 210 |
How to work without frame pointers |
| 211 |
---------------------------------- |
| 212 |
|
| 213 |
If you are working with a Python interpreter that has been compiled without |
| 214 |
frame pointers, you can still use the ``perf`` profiler, but the overhead will be |
| 215 |
a bit higher because Python needs to generate unwinding information for every |
| 216 |
Python function call on the fly. Additionally, ``perf`` will take more time to |
| 217 |
process the data because it will need to use the DWARF debugging information to |
| 218 |
unwind the stack and this is a slow process. |
| 219 |
|
| 220 |
To enable this mode, you can use the environment variable |
| 221 |
:envvar:`PYTHON_PERF_JIT_SUPPORT` or the :option:`-X perf_jit <-X>` option, |
| 222 |
which will enable the JIT mode for the ``perf`` profiler. |
| 223 |
|
| 224 |
.. note:: |
| 225 |
|
| 226 |
Due to a bug in the ``perf`` tool, only ``perf`` versions higher than v6.8 |
| 227 |
will work with the JIT mode. The fix was also backported to the v6.7.2 |
| 228 |
version of the tool. |
| 229 |
|
| 230 |
Note that when checking the version of the ``perf`` tool (which can be done |
| 231 |
by running ``perf version``) you must take into account that some distros |
| 232 |
add some custom version numbers including a ``-`` character. This means |
| 233 |
that ``perf 6.7-3`` is not necessarily ``perf 6.7.3``. |
| 234 |
|
| 235 |
When using the perf JIT mode, you need an extra step before you can run ``perf |
| 236 |
report``. You need to call the ``perf inject`` command to inject the JIT |
| 237 |
information into the ``perf.data`` file.:: |
| 238 |
|
| 239 |
$ perf record -F 9999 -g -k 1 --call-graph dwarf -o perf.data python -Xperf_jit my_script.py |
| 240 |
$ perf inject -i perf.data --jit --output perf.jit.data |
| 241 |
$ perf report -g -i perf.jit.data |
| 242 |
|
| 243 |
or using the environment variable:: |
| 244 |
|
| 245 |
$ PYTHON_PERF_JIT_SUPPORT=1 perf record -F 9999 -g --call-graph dwarf -o perf.data python my_script.py |
| 246 |
$ perf inject -i perf.data --jit --output perf.jit.data |
| 247 |
$ perf report -g -i perf.jit.data |
| 248 |
|
| 249 |
``perf inject --jit`` command will read ``perf.data``, |
| 250 |
automatically pick up the perf dump file that Python creates (in |
| 251 |
``/tmp/perf-$PID.dump``), and then create ``perf.jit.data`` which merges all the |
| 252 |
JIT information together. It should also create a lot of ``jitted-XXXX-N.so`` |
| 253 |
files in the current directory which are ELF images for all the JIT trampolines |
| 254 |
that were created by Python. |
| 255 |
|
| 256 |
.. warning:: |
| 257 |
When using ``--call-graph dwarf``, the ``perf`` tool will take |
| 258 |
snapshots of the stack of the process being profiled and save the |
| 259 |
information in the ``perf.data`` file. By default, the size of the stack dump |
| 260 |
is 8192 bytes, but you can change the size by passing it after |
| 261 |
a comma like ``--call-graph dwarf,16384``. |
| 262 |
|
| 263 |
The size of the stack dump is important because if the size is too small |
| 264 |
``perf`` will not be able to unwind the stack and the output will be |
| 265 |
incomplete. On the other hand, if the size is too big, then ``perf`` won't |
| 266 |
be able to sample the process as frequently as it would like as the overhead |
| 267 |
will be higher. |
| 268 |
|
| 269 |
The stack size is particularly important when profiling Python code compiled |
| 270 |
with low optimization levels (like ``-O0``), as these builds tend to have |
| 271 |
larger stack frames. If you are compiling Python with ``-O0`` and not seeing |
| 272 |
Python functions in your profiling output, try increasing the stack dump |
| 273 |
size to 65528 bytes (the maximum):: |
| 274 |
|
| 275 |
$ perf record -F 9999 -g -k 1 --call-graph dwarf,65528 -o perf.data python -Xperf_jit my_script.py |
| 276 |
|
| 277 |
Different compilation flags can significantly impact stack sizes: |
| 278 |
|
| 279 |
- Builds with ``-O0`` typically have much larger stack frames than those with ``-O1`` or higher |
| 280 |
- Adding optimizations (``-O1``, ``-O2``, etc.) typically reduces stack size |
| 281 |
- Frame pointers (``-fno-omit-frame-pointer``) generally provide more reliable stack unwinding |