time.time_ns() is not returing nanoseconds correctly on macOS?

问题

Starting from Python 3.7 we have new time functions supporting nanosecond resolution. However, I am not sure how time.time_ns() is supposed to work.

Look at the following example:

>>> for n in range(10):
...     time.sleep(random.random())
...     print((time.time(), time.time_ns(), time.monotonic_ns()))
...
(1545306865.8667252, 1545306865866727000, 439497985080)
(1545306866.084973, 1545306866084974000, 439716229679)
(1545306866.2972622, 1545306866297264000, 439928562751)
(1545306866.635714, 1545306866635716000, 440267014751)
(1545306866.745001, 1545306866745003000, 440376301646)
(1545306867.212074, 1545306867212076000, 440843415181)
(1545306867.7111092, 1545306867711111000, 441342449470)
(1545306867.792372, 1545306867792374000, 441423713091)
(1545306867.821886, 1545306867821887000, 441453223973)
(1545306868.127483, 1545306868127485000, 441758824065)

As you can see, time.time_ns() does return time as an integer with nanosecond precision, but the last digits are always 000. Which should not be the case. Is it a bug or am I missing something?

回答1:

It's about precision. Each clock in python has a corresponding precision which explains the difference that you are experiencing. Let's see the clock details from a Macbook Pro 2018 with MacOS Mojave. Python3.7 is installed via brew:

In [41]: time.perf_counter_ns()
Out[41]: 10464788941125

In [42]: time.process_time_ns()
Out[42]: 22502272000

In [43]: time.time_ns()
Out[43]: 1545312118561931000

In [44]: time.monotonic_ns()
Out[44]: 10477720411470

In [45]: time.get_clock_info('perf_counter')
Out[45]: namespace(adjustable=False, implementation='mach_absolute_time()', monotonic=True, resolution=1e-09)

In [46]: time.get_clock_info('process_time')
Out[46]: namespace(adjustable=False, implementation='clock_gettime(CLOCK_PROCESS_CPUTIME_ID)', monotonic=True, resolution=1.0000000000000002e-06)

In [47]: time.get_clock_info('time')
Out[47]: namespace(adjustable=True, implementation='clock_gettime(CLOCK_REALTIME)', monotonic=False, resolution=1.0000000000000002e-06)

In [48]: time.get_clock_info('monotonic')
Out[48]: namespace(adjustable=False, implementation='mach_absolute_time()', monotonic=True, resolution=1e-09)

Please pay attention to implementation and resolution. Here are the same info but from a VM running on an Ubuntu server:

>>> time.perf_counter_ns()
4094438601446186

>>> time.process_time_ns()
35344006

>>> time.time_ns()
1545312252720125938

>>> time.monotonic_ns()
4094449881239590

>>> time.get_clock_info('perf_counter')
namespace(adjustable=False, implementation='clock_gettime(CLOCK_MONOTONIC)', monotonic=True, resolution=1e-09)

>>> time.get_clock_info('time')
namespace(adjustable=True, implementation='clock_gettime(CLOCK_REALTIME)', monotonic=False, resolution=1e-09)

>>> time.get_clock_info('process_time')
namespace(adjustable=False, implementation='clock_gettime(CLOCK_PROCESS_CPUTIME_ID)', monotonic=True, resolution=1e-09)

>>> time.get_clock_info('monotonic')
namespace(adjustable=False, implementation='clock_gettime(CLOCK_MONOTONIC)', monotonic=True, resolution=1e-09)

As you can see each clock has a different implementation and precision which depends on platform. On MacOS the clock precision for process_time and time clocks is set to 1e-06 which is microseconds. That explains the difference.

You can also use time.clock_getres to get the precision:

In [51]: time.clock_getres(time.CLOCK_REALTIME)
Out[51]: 1.0000000000000002e-06