time.perf_counter() 应该在 Python on Windows 中的进程中保持一致吗?

Question

UPDATE：已提交此错误的修复程序，并将在 Python 3.10 中首次亮相，预计将于 2021 年 10 月发布。有关详细信息，请参阅错误报告。

---

time.perf_counter()的文档表明它是系统范围的

时间。perf_counter()→浮点数

返回性能计数器的值(以秒为单位)，即 具有最高可用分辨率的时钟，用于测量短持续时间。 它确实包括睡眠期间经过的时间，并且是系统范围的。这 返回值的参考点是未定义的，因此只有 连续调用结果之间的差异是有效的。

我在解释系统范围以包括跨流程的一致性时是否正确？

如下图所示，它在Linux上似乎是一致的，但在Windows上却不是。此外，Python 3.6 的 Windows 行为与 3.7 明显不同。

如果有人能指出涵盖此行为的文档或错误报告，我将不胜感激。

测试用例

import concurrent.futures
import time
def worker():
return time.perf_counter()
if __name__ == '__main__':
pool = concurrent.futures.ProcessPoolExecutor()
futures = []
for i in range(3):
print('Submitting worker {:d} at time.perf_counter() == {:.3f}'.format(i, time.perf_counter()))
futures.append(pool.submit(worker))
time.sleep(1)
for i, f in enumerate(futures):
print('Worker {:d} started at time.perf_counter() == {:.3f}'.format(i, f.result()))

视窗 7 上的结果

C:...>Python36python.exe -VV
Python 3.6.8 (tags/v3.6.8:3c6b436a57, Dec 24 2018, 00:16:47) [MSC v.1916 64 bit (AMD64)]
C:...>Python36python.exe perf_counter_across_processes.py
Submitting worker 0 at time.perf_counter() == 0.000
Submitting worker 1 at time.perf_counter() == 1.169
Submitting worker 2 at time.perf_counter() == 2.170
Worker 0 started at time.perf_counter() == 0.000
Worker 1 started at time.perf_counter() == 0.533
Worker 2 started at time.perf_counter() == 0.000
C:...>Python37python.exe -VV
Python 3.7.3 (v3.7.3:ef4ec6ed12, Mar 25 2019, 22:22:05) [MSC v.1916 64 bit (AMD64)]
C:...>Python37python.exe perf_counter_across_processes.py
Submitting worker 0 at time.perf_counter() == 0.376
Submitting worker 1 at time.perf_counter() == 1.527
Submitting worker 2 at time.perf_counter() == 2.529
Worker 0 started at time.perf_counter() == 0.380
Worker 1 started at time.perf_counter() == 0.956
Worker 2 started at time.perf_counter() == 1.963

为了简洁起见，我在Windows上省略了进一步的结果，但在Windows 8.1上观察到了相同的行为。此外，Python 3.6.7 的行为与 3.6.8 相同，而 Python 3.7.1 的行为与 3.7.3 相同。

优麒麟 18.04.1 LTS 的结果

$ python3 -VV
Python 3.6.7 (default, Oct 22 2018, 11:32:17) 
[GCC 8.2.0]
$ python3 perf_counter_across_processes.py 
Submitting worker 0 at time.perf_counter() == 2075.896
Submitting worker 1 at time.perf_counter() == 2076.900
Submitting worker 2 at time.perf_counter() == 2077.903
Worker 0 started at time.perf_counter() == 2075.900
Worker 1 started at time.perf_counter() == 2076.902
Worker 2 started at time.perf_counter() == 2077.905
$ python3.7 -VV
Python 3.7.1 (default, Oct 22 2018, 11:21:55) 
[GCC 8.2.0]
$ python3.7 perf_counter_across_processes.py 
Submitting worker 0 at time.perf_counter() == 1692.514
Submitting worker 1 at time.perf_counter() == 1693.518
Submitting worker 2 at time.perf_counter() == 1694.520
Worker 0 started at time.perf_counter() == 1692.517
Worker 1 started at time.perf_counter() == 1693.519
Worker 2 started at time.perf_counter() == 1694.522

Answer 1

在Windows中，time.perf_counter基于WINAPIQueryPerformanceCounter。此计数器是系统范围的。有关详细信息，请参阅获取高分辨率时间戳。

也就是说，Windows 中的perf_counter返回一个相对于进程启动值的值。因此，它不是系统范围的值。它这样做是为了减少将整数值转换为只有15个十进制数字的精度的float时的精度损失。在大多数情况下，不需要使用相对值，这只需要微秒精度。应该有一个可选参数来查询真正的 QPC 计数器值，特别是对于 3.7+ 中的perf_counter_ns。

关于perf_counter在 3.6 与 3.7 中返回的不同初始值，随着时间的推移，实现发生了一些变化。在 3.6.8 中，perf_counter在 Modules/timemodule.c 中实现，因此初始值在首次导入和初始化time模块时存储，这就是为什么您看到的第一个结果为 0.000 秒。在最近的版本中，它是在Python的C API中单独实现的。例如，请参阅最新 3.8 beta 版本中的"Python/pytime.c"。在这种情况下，当 Python 代码调用time.perf_counter()时，计数器已经远远超过了启动值。

下面是基于 ctypes 的替代实现，它使用系统范围的 QPC 值而不是相对值。

import sys
if sys.platform != 'win32':
from time import perf_counter
try:
from time import perf_counter_ns
except ImportError:
def perf_counter_ns():
"""perf_counter_ns() -> int
Performance counter for benchmarking as nanoseconds.
"""
return int(perf_counter() * 10**9)
else:
import ctypes
from ctypes import wintypes
kernel32 = ctypes.WinDLL('kernel32', use_last_error=True)
kernel32.QueryPerformanceFrequency.argtypes = (
wintypes.PLARGE_INTEGER,) # lpFrequency
kernel32.QueryPerformanceCounter.argtypes = (
wintypes.PLARGE_INTEGER,) # lpPerformanceCount
_qpc_frequency = wintypes.LARGE_INTEGER()
if not kernel32.QueryPerformanceFrequency(ctypes.byref(_qpc_frequency)):
raise ctypes.WinError(ctypes.get_last_error())
_qpc_frequency = _qpc_frequency.value
def perf_counter_ns():
"""perf_counter_ns() -> int
Performance counter for benchmarking as nanoseconds.
"""
count = wintypes.LARGE_INTEGER()
if not kernel32.QueryPerformanceCounter(ctypes.byref(count)):
raise ctypes.WinError(ctypes.get_last_error())
return (count.value * 10**9) // _qpc_frequency
def perf_counter():
"""perf_counter() -> float
Performance counter for benchmarking.
"""
count = wintypes.LARGE_INTEGER()
if not kernel32.QueryPerformanceCounter(ctypes.byref(count)):
raise ctypes.WinError(ctypes.get_last_error())
return count.value / _qpc_frequency

QPC 通常具有 0.1 微秒的分辨率。CPython 中的float具有 15 个十进制数字的精度。因此，这种perf_counter的实施在QPC决议范围内，正常运行时间约为3年。