Python support for the perf map compatible profilers

作者:

Pablo Galindo

The Linux perf profiler and samply are powerful tools that allow you to profile and obtain information about the performance of your application. Both tools have vibrant ecosystems that aid with the analysis of the data they produce.

The main problem with using these profilers with Python applications is that they only get information about native symbols, that is, the names of functions and procedures written in C. This means that the names and file names of Python functions in your code will not appear in the profiler output.

Since Python 3.12, the interpreter can run in a special mode that allows Python functions to appear in the output of compatible profilers. When this mode is enabled, the interpreter will interpose a small piece of code compiled on the fly before the execution of every Python function and it will teach the profiler the relationship between this piece of code and the associated Python function using perf map files.

備註

Support for profiling is available on Linux and macOS on select architectures. Perf is available on Linux, while samply can be used on both Linux and macOS. samply support on macOS is available starting from Python 3.15. Check the output of the configure build step or check the output of python -m sysconfig | grep HAVE_PERF_TRAMPOLINE to see if your system is supported.

例如,參考以下腳本:

def foo(n):
    result = 0
    for _ in range(n):
        result += 1
    return result

def bar(n):
    foo(n)

def baz(n):
    bar(n)

if __name__ == "__main__":
    baz(1000000)

我們可以執行 perf 以 9999 赫茲取樣 CPU 堆疊追蹤 (stack trace):

$ perf record -F 9999 -g -o perf.data python my_script.py

然後我們可以使用 perf report 來分析資料:

$ perf report --stdio -n -g

# Children      Self       Samples  Command     Shared Object       Symbol
# ........  ........  ............  ..........  ..................  ..........................................
#
    91.08%     0.00%             0  python.exe  python.exe          [.] _start
            |
            ---_start
            |
                --90.71%--__libc_start_main
                        Py_BytesMain
                        |
                        |--56.88%--pymain_run_python.constprop.0
                        |          |
                        |          |--56.13%--_PyRun_AnyFileObject
                        |          |          _PyRun_SimpleFileObject
                        |          |          |
                        |          |          |--55.02%--run_mod
                        |          |          |          |
                        |          |          |           --54.65%--PyEval_EvalCode
                        |          |          |                     _PyEval_EvalFrameDefault
                        |          |          |                     PyObject_Vectorcall
                        |          |          |                     _PyEval_Vector
                        |          |          |                     _PyEval_EvalFrameDefault
                        |          |          |                     PyObject_Vectorcall
                        |          |          |                     _PyEval_Vector
                        |          |          |                     _PyEval_EvalFrameDefault
                        |          |          |                     PyObject_Vectorcall
                        |          |          |                     _PyEval_Vector
                        |          |          |                     |
                        |          |          |                     |--51.67%--_PyEval_EvalFrameDefault
                        |          |          |                     |          |
                        |          |          |                     |          |--11.52%--_PyCompactLong_Add
                        |          |          |                     |          |          |
                        |          |          |                     |          |          |--2.97%--_PyObject_Malloc
...

如你所見,Python 函式未顯示在輸出中,僅顯示 _Py_Eval_EvalFrameDefault (為 Python 位元組碼 (bytecode) 求值的函式)。不幸的是,這不是很有用,因為所有 Python 函式都使用相同的 C 函式來替位元組碼求值,因此我們無法知道哪個 Python 函式是對應於哪個位元組碼計算函式。

作為替代,如果我們在啟用 perf 支援的情況下執行相同的實驗,我們會得到:

$ perf report --stdio -n -g

# Children      Self       Samples  Command     Shared Object       Symbol
# ........  ........  ............  ..........  ..................  .....................................................................
#
    90.58%     0.36%             1  python.exe  python.exe          [.] _start
            |
            ---_start
            |
                --89.86%--__libc_start_main
                        Py_BytesMain
                        |
                        |--55.43%--pymain_run_python.constprop.0
                        |          |
                        |          |--54.71%--_PyRun_AnyFileObject
                        |          |          _PyRun_SimpleFileObject
                        |          |          |
                        |          |          |--53.62%--run_mod
                        |          |          |          |
                        |          |          |           --53.26%--PyEval_EvalCode
                        |          |          |                     py::<module>:/src/script.py
                        |          |          |                     _PyEval_EvalFrameDefault
                        |          |          |                     PyObject_Vectorcall
                        |          |          |                     _PyEval_Vector
                        |          |          |                     py::baz:/src/script.py
                        |          |          |                     _PyEval_EvalFrameDefault
                        |          |          |                     PyObject_Vectorcall
                        |          |          |                     _PyEval_Vector
                        |          |          |                     py::bar:/src/script.py
                        |          |          |                     _PyEval_EvalFrameDefault
                        |          |          |                     PyObject_Vectorcall
                        |          |          |                     _PyEval_Vector
                        |          |          |                     py::foo:/src/script.py
                        |          |          |                     |
                        |          |          |                     |--51.81%--_PyEval_EvalFrameDefault
                        |          |          |                     |          |
                        |          |          |                     |          |--13.77%--_PyCompactLong_Add
                        |          |          |                     |          |          |
                        |          |          |                     |          |          |--3.26%--_PyObject_Malloc

Using the samply profiler

samply is a modern profiler that can be used as an alternative to perf. It uses the same perf map files that Python generates, making it compatible with Python's profiling support. samply is particularly useful on macOS where perf is not available.

To use samply with Python, first install it following the instructions at https://github.com/mstange/samply, then run:

$ samply record PYTHONPERFSUPPORT=1 python my_script.py

This will open a web interface where you can analyze the profiling data interactively. The advantage of samply is that it provides a modern web-based interface for analyzing profiling data and works on both Linux and macOS.

On macOS, samply support requires Python 3.15 or later. Also on macOS, samply can't profile signed Python executables due to restrictions by macOS. You can profile with Python binaries that you've compiled yourself, or which are unsigned or locally-signed (such as anything installed by Homebrew). In order to attach to running processes on macOS, run samply setup once (and every time samply is updated) to self-sign the samply binary.

如何啟用 perf 分析支援

要啟用 perf 分析支援,可以在一開始就使用環境變數 PYTHONPERFSUPPORT 或使用 -X perf 選項,也可以使用 sys.activate_stack_trampoline()sys.deactivate_stack_trampoline() 來動態啟用。

sys 函式優先於 -X 選項、-X 選項優先於環境變數。

例如,使用環境變數:

$ PYTHONPERFSUPPORT=1 perf record -F 9999 -g -o perf.data python my_script.py
$ perf report -g -i perf.data

例如,使用 -X 選項:

$ perf record -F 9999 -g -o perf.data python -X perf my_script.py
$ perf report -g -i perf.data

例如,在 example.py 檔案中使用 sys API:

import sys

sys.activate_stack_trampoline("perf")
do_profiled_stuff()
sys.deactivate_stack_trampoline()

non_profiled_stuff()

...然後:

$ perf record -F 9999 -g -o perf.data python ./example.py
$ perf report -g -i perf.data

如何獲得最佳結果

為了獲得最佳結果,應使用 CFLAGS="-fno-omit-frame-pointer -mno-omit-leaf-frame-pointer" 來進行 Python 編譯,因為這能允許分析器僅使用 frame 指標而不是 DWARF 除錯資訊來解析 (unwind)。這是因為,由於插入以允許 perf 支援的程式碼是動態生成的,因此它沒有任何可用的 DWARF 除錯資訊。

你可以透過執行以下指令來檢查你的系統是否已使用此旗標進行編譯:

$ python -m sysconfig | grep 'no-omit-frame-pointer'

如果你沒有看到任何輸出,則表示你的直譯器尚未使用 frame 指標進行編譯,因此它可能無法在 perf 的輸出中顯示 Python 函式。

How to work without frame pointers

If you are working with a Python interpreter that has been compiled without frame pointers, you can still use the perf profiler, but the overhead will be a bit higher because Python needs to generate unwinding information for every Python function call on the fly. Additionally, perf will take more time to process the data because it will need to use the DWARF debugging information to unwind the stack and this is a slow process.

To enable this mode, you can use the environment variable PYTHON_PERF_JIT_SUPPORT or the -X perf_jit option, which will enable the JIT mode for the perf profiler.

備註

Due to a bug in the perf tool, only perf versions higher than v6.8 will work with the JIT mode. The fix was also backported to the v6.7.2 version of the tool.

Note that when checking the version of the perf tool (which can be done by running perf version) you must take into account that some distros add some custom version numbers including a - character. This means that perf 6.7-3 is not necessarily perf 6.7.3.

When using the perf JIT mode, you need an extra step before you can run perf report. You need to call the perf inject command to inject the JIT information into the perf.data file.:

$ perf record -F 9999 -g -k 1 --call-graph dwarf -o perf.data python -Xperf_jit my_script.py
$ perf inject -i perf.data --jit --output perf.jit.data
$ perf report -g -i perf.jit.data

或使用環境變數:

$ PYTHON_PERF_JIT_SUPPORT=1 perf record -F 9999 -g --call-graph dwarf -o perf.data python my_script.py
$ perf inject -i perf.data --jit --output perf.jit.data
$ perf report -g -i perf.jit.data

perf inject --jit command will read perf.data, automatically pick up the perf dump file that Python creates (in /tmp/perf-$PID.dump), and then create perf.jit.data which merges all the JIT information together. It should also create a lot of jitted-XXXX-N.so files in the current directory which are ELF images for all the JIT trampolines that were created by Python.

警告

When using --call-graph dwarf, the perf tool will take snapshots of the stack of the process being profiled and save the information in the perf.data file. By default, the size of the stack dump is 8192 bytes, but you can change the size by passing it after a comma like --call-graph dwarf,16384.

The size of the stack dump is important because if the size is too small perf will not be able to unwind the stack and the output will be incomplete. On the other hand, if the size is too big, then perf won't be able to sample the process as frequently as it would like as the overhead will be higher.

The stack size is particularly important when profiling Python code compiled with low optimization levels (like -O0), as these builds tend to have larger stack frames. If you are compiling Python with -O0 and not seeing Python functions in your profiling output, try increasing the stack dump size to 65528 bytes (the maximum):

$ perf record -F 9999 -g -k 1 --call-graph dwarf,65528 -o perf.data python -Xperf_jit my_script.py

Different compilation flags can significantly impact stack sizes:

  • Builds with -O0 typically have much larger stack frames than those with -O1 or higher

  • Adding optimizations (-O1, -O2, etc.) typically reduces stack size

  • Frame pointers (-fno-omit-frame-pointer) generally provide more reliable stack unwinding