Function profiling should use time.time_ns instead of time.time

[michaelosthege]

Description of your problem or feature request

The function profiling in

aesara/aesara/compile/function/types.py

Line 827 in 2dc0af2

t0 = time.time()

(and more lines that file) uses time.time() which has varying precision on different platforms.

time.time_ns() or time.time_ns() / (10 ** 9) for float seconds should be used for high-precision performance measurement.

Further reading: https://stackoverflow.com/a/1938096

Versions and main components

• Aesara version: 2.8.7
• Operating system: mac-OS, Windows

[brandonwillard]

Isn't time.perf_counter[_ns] provided explicitly for these purposes?

[anirudhacharya]

@brandonwillard yes, since we are only using it to measure time difference and nowhere are we logging the actual time of an event, time.perf_counter() makes more sense than time.time().

Depending on how we want to do the profiling, the options are -

1. time.time() / time.time_ns()
2. time.perf_counter() - can be used to record the difference in clock time.
3. time.process_time() - can be used to record the difference in process time.

import time

def perf_count():
    tic = time.perf_counter()
    time.sleep(1)
    toc = time.perf_counter()
    return toc - tic

def proc_time():
    tic = time.process_time()
    time.sleep(1)
    toc = time.process_time()
    return toc - tic

print(perf_count())
print(proc_time())

Output 
-------
1.003998823
5.9000000000003494e-05

Should I go ahead and change it to perf_counter?

[brandonwillard]

Should I go ahead and change it to perf_counter?

Yes, that sounds good.