We Keep Coding

Benchmark Python programs

I'm new to Python. It is my first interpreted language. I used to study Java so far.
So, when Java program runs for the first time, it is execueted slower than for the next times.
The reasi is caching.
import time
def procedure():
time.sleep(2.5)
# measure process time
t0 = time.clock()
procedure()
print (time.clock() - t0), "seconds process time"
I tried this for several times. The result is always equal. So, am I right that no cashe interferes and that the benchmark is pretty reliable?

it's OK to do benchmarks like this, the accuracy is good enough for functions which run "long" and pretty constant, like in your example.
But there are some pitfalls: for "quick" functions (like the empty one), you run into precision limits. And for functions which vary in execution time (like net i/o, for example), you have to measure multiple times to find min/max/avg runtime.
And in addition to that, the clock best used differs on platforms: on windows, time.clock() is preferred, on *nix, time.time().
luckily, there is a module which takes care of all that: timeit:
>>> import time
>>> def procedure():
pass
>>> def time_this(f):
t0=time.clock()
f()
print((time.clock() - t0), "seconds process time")
>>> time_this(procedure)
1.9555558310457855e-06 seconds process time
>>> time_this(procedure)
1.9555557742023666e-06 seconds process time
>>> time_this(procedure)
1.9555557742023666e-06 seconds process time
>>> import timeit
>>> timeit.Timer(procedure).timeit()
0.09460783423588737
>>> timeit.Timer(procedure).repeat()
[0.09791419021132697, 0.09721947901198291, 0.09598943441130814]
you might want to look at it's source. Or just use it ;)
As for caching: python code compiles to bytecode when first used. This bytecode is cached by default - but this won't affect your benchmark as long as you don't do imports in your function.

Sleep for exact time in python

I need to wait for about 25ms in one of my functions. Sometimes this function is called when the processor is occupied with other things and other times it has the processor all to itself.
I've tried time.sleep(.25) but sometimes its actually 25ms and other times it takes much longer. Is there a way to sleep for an exact amount of time regardless of processor availability?

Because you're working with a preemptive operating system, there's no way you can guarantee that your process will be able to have control of the CPU in 25ms.
If you'd still like to try, it would be better to have a busy loop that polls until 25ms has passed. Something like this might work:
import time
target_time = time.clock() + 0.025
while time.clock() < target_time:
pass

0.25 seconds are 250 ms, not 25. Apart from this, there is no way to wait for exactly 25 ms on common operating systems – you would need some real-time operating system.

What system are you on? If you're on Windows you may want to do something like this for exact timing:
import ctypes
kernel32 = ctypes.windll.kernel32
# This sets the priority of the process to realtime--the same priority as the mouse pointer.
kernel32.SetThreadPriority(kernel32.GetCurrentThread(), 31)
# This creates a timer. This only needs to be done once.
timer = kernel32.CreateWaitableTimerA(ctypes.c_void_p(), True, ctypes.c_void_p())
# The kernel measures in 100 nanosecond intervals, so we must multiply .25 by 10000
delay = ctypes.c_longlong(.25 * 10000)
kernel32.SetWaitableTimer(timer, ctypes.byref(delay), 0, ctypes.c_void_p(), ctypes.c_void_p(), False)
kernel32.WaitForSingleObject(timer, 0xffffffff)
This code will pretty much guarentee your process will sleep .25 seconds. Watch out though- you may want to lower the priority to 2 or 3 unless it's absolutely critical that this sleeps for .25 seconds. Certainly don't change the priority too high for a user-end product.

Edit: in Windows 10 this nonsense seems unnecessary. Try it like so:
>>> from time import sleep
>>> import timeit
>>> '%.2f%% overhead' % (timeit.timeit('sleep(0.025)', number=100, globals=globals()) / 0.025 - 100)
'0.29% overhead'
.29%, or thereabout, is fairly low overhead, and usually more than accurate enough.
Previous Windows versions will by default have a sleep resolution of 55 msecs, which means your sleep call will take somewhere between 25 and 55 msecs. To get the sleep resolution down to 1 millisecond you need to set the resolution used by Windows by calling timeBeginPeriod:
import ctypes
winmm = ctypes.WinDLL('winmm')
winmm.timeBeginPeriod(1)

Another solution for accurate timings and delay is to use the perf_counter() function from module time. Especially useful in windows as time.sleep is not accurate in milliseconds. See below example where function accurate_delay creates a delay in millisecond.
import time
def accurate_delay(delay):
''' Function to provide accurate time delay in millisecond
'''
_ = time.perf_counter() + delay/1000
while time.perf_counter() < _:
pass
delay = 10
t_start = time.perf_counter()
print('Wait for {:.0f} ms. Start: {:.5f}'.format(delay, t_start))
accurate_delay(delay)
t_end = time.perf_counter()
print('End time: {:.5f}. Delay is {:.5f} ms'.
format(t_end, 1000*(t_end - t_start)))
sum = 0
ntests = 1000
for _ in range(ntests):
t_start = time.perf_counter()
accurate_delay(delay)
t_end = time.perf_counter()
print('Test completed: {:.2f}%'.format(_/ntests * 100), end='\r', flush=True)
sum = sum + 1000*(t_end - t_start) - delay
print('Average difference in time delay is {:.5f} ms.'.format(sum/ntests))

What you intend to do is a real time application. Python (and probably the OS you are using) is not intended to program this kind of applications, where time restriction is so strict.
In order for you to achieve what you are looking for you need a RTOS (Real Time Operating System) and develop your application using a suitable programming language (usually C) following RT best practises.

From the docs of the sleep method:
Suspend execution for the given number of seconds. The argument may be
a floating point number to indicate a more precise sleep time. The
actual suspension time may be less than that requested because any
caught signal will terminate the sleep() following execution of that
signal’s catching routine. Also, the suspension time may be longer
than requested by an arbitrary amount because of the scheduling of
other activity in the system.
The fact is that it depends on your underlying OS.

Measuring elapsed time in python

Is there a simple way / module to correctly measure the elapsed time in python? I know that I can simply call time.time() twice and take the difference, but that will yield wrong results if the system time is changed. Granted, that doesn't happen very often, but it does indicate that I'm measuring the wrong thing.
Using time.time() to measure durations is incredibly roundabout when you think about it. You take the difference of two absolute time measurements which are in turn constructed from duration measurements (performed by timers) and known absolute times (set manually or via ntp), that you aren't interested in at all.
So, is there a way to query this "timer time" directly? I'd imagine that it can be represented as a millisecond or microsecond value that has no meaningful absolute representation (and thus doesn't need to be adjusted with system time). Looking around a bit it seems that this is exactly what System.nanoTime() does in Java, but I did not find a corresponding Python function, even though it should (hardware-technically) be easier to provide than time.time().
Edit: To avoid confusion and address the answers below: This is not about DST changes, and I don't want CPU time either - I want elapsed physical time. It doesn't need to be very fine-grained, and not even particularly accurate. It just shouldn't give me negative durations, or durations which are off by several orders of magnitude (above the granularity), just because someone decided to set the system clock to a different value. Here's what the Python docs say about 'time.time()':
"While this function normally returns non-decreasing values, it can return a lower value than a previous call if the system clock has been set back between the two calls"
This is exactly what I want to avoid, since it can lead to strange things like negative values in time calculations. I can work around this at the moment, but I believe it is a good idea to learn using the proper solutions where feasible, since the kludges will come back to bite you one day.
Edit2: Some research shows that you can get a system time independent measurement like I want in Windows by using GetTickCount64(), under Linux you can get it in the return value of times(). However, I still can't find a module which provides this functionality in Python.

For measuring elapsed CPU time, look at time.clock(). This is the equivalent of Linux's times() user time field.
For benchmarking, use timeit.
The datetime module, which is part of Python 2.3+, also has microsecond time if supported by the platform.
Example:
>>> import datetime as dt
>>> n1=dt.datetime.now()
>>> n2=dt.datetime.now()
>>> (n2-n1).microseconds
678521
>>> (n2.microsecond-n1.microsecond)/1e6
0.678521
ie, it took me .678521 seconds to type the second n2= line -- slow
>>> n1.resolution
datetime.timedelta(0, 0, 1)
1/1e6 resolution is claimed.
If you are concerned about system time changes (from DS -> ST) just check the object returned by datetime.Presumably, the system time could have a small adjustment from an NTP reference adjustment. This should be slewed, and corrections are applied gradually, but ntp sync beats can have an effect with very small (millisec or microsec) time references.
You can also reference Alex Martelli's C function if you want something of that resolution. I would not go too far to reinvent the wheel. Accurate time is basic and most modern OS's do a pretty good job.
Edit
Based on your clarifications, it sounds like you need a simple side check if the system's clock has changed. Just compare to a friendly, local ntp server:
import socket
import struct
import time
ntp="pool.ntp.org" # or whatever ntp server you have handy
client = socket.socket( socket.AF_INET, socket.SOCK_DGRAM )
data = '\x1b' + 47 * '\0'
client.sendto( data, ( ntp, 123 ))
data, address = client.recvfrom( 1024 )
if data:
print 'Response received from:', address
t = struct.unpack( '!12I', data )[10]
t -= 2208988800L #seconds since Epoch
print '\tTime=%s' % time.ctime(t)
NTP is accurate to milliseconds over the Internet and has representation resolution of resolution of 2−32 seconds (233 picoseconds). Should be good enough?
Be aware that the NTP 64 bit data structure will overflow in 2036 and every 136 years thereafter -- if you really want a robust solution, better check for overflow...

What you seem to be looking for is a monotonic timer. A monotonic time reference does not jump or go backwards.
There have been several attempts to implement a cross platform monotomic clock for Python based on the OS reference of it. (Windows, POSIX and BSD are quite different) See the discussions and some of the attempts at monotonic time in this SO post.
Mostly, you can just use os.times():
os.times()
Return a 5-tuple of floating point numbers indicating
accumulated (processor or other) times, in seconds. The items are:
user time, system time, children’s user time, children’s system time,
and elapsed real time since a fixed point in the past, in that order.
See the Unix manual page times(2) or the corresponding Windows
Platform API documentation. On Windows, only the first two items are
filled, the others are zero.
Availability: Unix, Windows
But that does not fill in the needed elapsed real time (the fifth tuple) on Windows.
If you need Windows support, consider ctypes and you can call GetTickCount64() directly, as has been done in this recipe.

Python 3.3 added a monotonic timer into the standard library, which does exactly what I was looking for. Thanks to Paddy3118 for pointing this out in "How do I get monotonic time durations in python?".

>>> import datetime
>>> t1=datetime.datetime.utcnow()
>>> t2=datetime.datetime.utcnow()
>>> t2-t1
datetime.timedelta(0, 8, 600000)
Using UTC avoids those embarassing periods when the clock shifts due to daylight saving time.
As for using an alternate method rather than subtracting two clocks, be aware that the OS does actually contain a clock which is initialized from a hardware clock in the PC. Modern OS implementations will also keep that clock synchronized with some official source so that it doesn't drift. This is much more accurate than any interval timer the PC might be running.

You can use perf_counter function of time module in Python Standard Library:
from datetime import timedelta
from time import perf_counter
startTime = perf_counter()
CallYourFunc()
finishedTime = perf_counter()
duration = timedelta(seconds=(finishedTime - startTime))

The example functions you state in your edit are two completely different things:
Linux times() returns process times in CPU milliseconds. Python's equivalent is time.clock() or os.times().
Windows GetTickCount64() returns system uptime.
Although two different functions, both (potentially) could be used to reveal a system clock that had a "burp" with these methods:
First:
You could take both a system time with time.time() and a CPU time with time.clock(). Since wall clock time will ALWAYS be greater than or equal to CPU time, discard any measurements where the interval between the two time.time() readings is less than the paired time.clock() check readings.
Example:
t1=time.time()
t1check=time.clock()
# your timed event...
t2=time.time()
t2check=time.clock()
if t2-t1 < t2check - t1check:
print "Things are rotten in Denmark"
# discard that sample
else:
# do what you do with t2 - t1...
Second:
Getting system uptime is also promising if you are concerned about the system's clock, since a user reset does not reset the uptime tick count in most cases. (that I am aware of...)
Now the harder question: getting system uptime in a platform independent way -- especially without spawning a new shell -- at the sub second accuracy. Hmmm...
Probably the best bet is psutil. Browsing the source, they use uptime = GetTickCount() / 1000.00f; for Windows and sysctl "kern.boottime" for BSD / OS X, etc. Unfortunately, these are all 1 second resolution.

from datetime import datetime
start = datetime.now()
print 'hello'
end = datetime.now()
delta = end-start
print type(delta)
<type 'datetime.timedelta'>
import datetime
help(datetime.timedelta)
...elapsed seconds and microseconds...

After using time.monotonic from Python 3.3+ I found that on Mac it works well, never going backwards. On Windows, where it uses GetTickCount64() it can very rarely go backwards by a substantial amount (for the purposes of my program that was in excess of 5.0) Adding a wrapper can prevent monotonic from going backwards:
with a_lock:
original_value = time.monotonic()
if original_value < last_value:
# You can post a metric here to monitor frequency and size of backward jumps
offset = last_value - original_value
last_value = original_value
return offset + original_value
How often did it go backwards? Perhaps a handful of times over many days across millions of machines and again, only on Windows. Sorry, I did not track which versions of Windows. I can update this later if people wish.

Why time.clock() returns such a large value on Windows Server 2008 X64

I ran following script on different machine and got quite different results. The elapsed time.clock() is so large.
Script:
#------------------------------------------------------------------------------------
import time
start_clock = time.clock()
time.sleep(60)
end_clock = time.clock()
print "Sleep Clock = ", str(end_clock - start_clock)
start_time = time.time()
time.sleep(60)
end_time = time.time()
print "Sleep Time = ", str(end_time - start_time)
#-------------------------------------------------------------------------------------
Output:
Instance (Windows Server 2008, X64):
Sleep Clock = 938.306471633
Sleep Time = 60.0119998455
Local Machine (Windows Vista, X86):
Sleep Clock = 59.9997987873
Sleep Time = 59.996999979
Following result really confused me:
Sleep Clock = 938.306471633
P.s:
I have not tested on other X64 OSs. This Windows Server 2008 is a running Amazon Instance.

Per the docs on time.clock
On Windows, this function returns
wall-clock seconds elapsed since the
first call to this function, as a floating point number, based on the Win32 function QueryPerformanceCounter().
so my (blind, i.e., I've never seen Amazon's code for Windows virtualization!-) guess would be that Amazon's virtualization doesn't go quite deep enough to trick QueryPerformanceCounter (which is a very low-level, low-overhead function). Tricking time.time (in a virtualizing hypervisor) is easier (and a more common need).
Do you know what happens e.g. on Microsoft's Azure, and with other non-Microsoft virtualizers such as Parallels or VMWare? I wouldn't be surprised to see different "depth" to the amount of "trickery" (deep virtualization) performed in each case. (I don't doubt that the explanation for this observation must have to do with virtualization, although the specific guess I make above could be flawed).
It would also be interesting to try (again, on various different virtualizers) a tiny C program doing just QueryPerformanceCounter, just to confirm that Python's runtime has nothing to do with the case (I believe so, by inspection of the runtime's source, but a confirmation could not hurt -- unfortunately I don't have access to the resources needed to try it myself).

Python's time.clock() vs. time.time() accuracy?

Which is better to use for timing in Python? time.clock() or time.time()? Which one provides more accuracy?
for example:
start = time.clock()
... do something
elapsed = (time.clock() - start)
vs.
start = time.time()
... do something
elapsed = (time.time() - start)

As of 3.3, time.clock() is deprecated, and it's suggested to use time.process_time() or time.perf_counter() instead.
Previously in 2.7, according to the time module docs:
time.clock()
On Unix, return the current processor time as a floating point number
expressed in seconds. The precision, and in fact the very definition
of the meaning of “processor time”, depends on that of the C function
of the same name, but in any case, this is the function to use for
benchmarking Python or timing algorithms.
On Windows, this function returns wall-clock seconds elapsed since the
first call to this function, as a floating point number, based on the
Win32 function QueryPerformanceCounter(). The resolution is typically
better than one microsecond.
Additionally, there is the timeit module for benchmarking code snippets.

The short answer is: most of the time time.clock() will be better.
However, if you're timing some hardware (for example some algorithm you put in the GPU), then time.clock() will get rid of this time and time.time() is the only solution left.
Note: whatever the method used, the timing will depend on factors you cannot control (when will the process switch, how often, ...), this is worse with time.time() but exists also with time.clock(), so you should never run one timing test only, but always run a series of test and look at mean/variance of the times.

Others have answered re: time.time() vs. time.clock().
However, if you're timing the execution of a block of code for benchmarking/profiling purposes, you should take a look at the timeit module.

One thing to keep in mind:
Changing the system time affects time.time() but not time.clock().
I needed to control some automatic tests executions. If one step of the test case took more than a given amount of time, that TC was aborted to go on with the next one.
But sometimes a step needed to change the system time (to check the scheduler module of the application under test), so after setting the system time a few hours in the future, the TC timeout expired and the test case was aborted. I had to switch from time.time() to time.clock() to handle this properly.

clock() -> floating point number
Return the CPU time or real time since the start of the process or since
the first call to clock(). This has as much precision as the system
records.
time() -> floating point number
Return the current time in seconds since the Epoch.
Fractions of a second may be present if the system clock provides them.
Usually time() is more precise, because operating systems do not store the process running time with the precision they store the system time (ie, actual time)

Depends on what you care about. If you mean WALL TIME (as in, the time on the clock on your wall), time.clock() provides NO accuracy because it may manage CPU time.

time() has better precision than clock() on Linux. clock() only has precision less than 10 ms. While time() gives prefect precision.
My test is on CentOS 6.4， python 2.6
using time():
1 requests, response time: 14.1749382019 ms
2 requests, response time: 8.01301002502 ms
3 requests, response time: 8.01491737366 ms
4 requests, response time: 8.41021537781 ms
5 requests, response time: 8.38804244995 ms
using clock():
1 requests, response time: 10.0 ms
2 requests, response time: 0.0 ms
3 requests, response time: 0.0 ms
4 requests, response time: 10.0 ms
5 requests, response time: 0.0 ms
6 requests, response time: 0.0 ms
7 requests, response time: 0.0 ms
8 requests, response time: 0.0 ms

As others have noted time.clock() is deprecated in favour of time.perf_counter() or time.process_time(), but Python 3.7 introduces nanosecond resolution timing with time.perf_counter_ns(), time.process_time_ns(), and time.time_ns(), along with 3 other functions.
These 6 new nansecond resolution functions are detailed in PEP 564:
time.clock_gettime_ns(clock_id)
time.clock_settime_ns(clock_id, time:int)
time.monotonic_ns()
time.perf_counter_ns()
time.process_time_ns()
time.time_ns()
These functions are similar to the version without the _ns suffix, but
return a number of nanoseconds as a Python int.
As others have also noted, use the timeit module to time functions and small code snippets.

The difference is very platform-specific.
clock() is very different on Windows than on Linux, for example.
For the sort of examples you describe, you probably want the "timeit" module instead.

I use this code to compare 2 methods .My OS is windows 8 , processor core i5 , RAM 4GB
import time
def t_time():
start=time.time()
time.sleep(0.1)
return (time.time()-start)
def t_clock():
start=time.clock()
time.sleep(0.1)
return (time.clock()-start)
counter_time=0
counter_clock=0
for i in range(1,100):
counter_time += t_time()
for i in range(1,100):
counter_clock += t_clock()
print "time() =",counter_time/100
print "clock() =",counter_clock/100
output:
time() = 0.0993799996376
clock() = 0.0993572257367

time.clock() was removed in Python 3.8 because it had platform-dependent behavior:
On Unix, return the current processor time as a floating point number expressed in seconds.
On Windows, this function returns wall-clock seconds elapsed since the first call to this function, as a floating point number
print(time.clock()); time.sleep(10); print(time.clock())
# Linux : 0.0382 0.0384 # see Processor Time
# Windows: 26.1224 36.1566 # see Wall-Clock Time
So which function to pick instead?
Processor Time: This is how long this specific process spends actively being executed on the CPU. Sleep, waiting for a web request, or time when only other processes are executed will not contribute to this.
Use time.process_time()
Wall-Clock Time: This refers to how much time has passed "on a clock hanging on the wall", i.e. outside real time.
Use time.perf_counter()
time.time() also measures wall-clock time but can be reset, so you could go back in time
time.monotonic() cannot be reset (monotonic = only goes forward) but has lower precision than time.perf_counter()

On Unix time.clock() measures the amount of CPU time that has been used by the current process, so it's no good for measuring elapsed time from some point in the past. On Windows it will measure wall-clock seconds elapsed since the first call to the function. On either system time.time() will return seconds passed since the epoch.
If you're writing code that's meant only for Windows, either will work (though you'll use the two differently - no subtraction is necessary for time.clock()). If this is going to run on a Unix system or you want code that is guaranteed to be portable, you will want to use time.time().

Short answer: use time.clock() for timing in Python.
On *nix systems, clock() returns the processor time as a floating point number, expressed in seconds. On Windows, it returns the seconds elapsed since the first call to this function, as a floating point number.
time() returns the the seconds since the epoch, in UTC, as a floating point number. There is no guarantee that you will get a better precision that 1 second (even though time() returns a floating point number). Also note that if the system clock has been set back between two calls to this function, the second function call will return a lower value.

To the best of my understanding, time.clock() has as much precision as your system will allow it.

Right answer : They're both the same length of a fraction.
But which faster if subject is time ?
A little test case :
import timeit
import time
clock_list = []
time_list = []
test1 = """
def test(v=time.clock()):
s = time.clock() - v
"""
test2 = """
def test(v=time.time()):
s = time.time() - v
"""
def test_it(Range) :
for i in range(Range) :
clk = timeit.timeit(test1, number=10000)
clock_list.append(clk)
tml = timeit.timeit(test2, number=10000)
time_list.append(tml)
test_it(100)
print "Clock Min: %f Max: %f Average: %f" %(min(clock_list), max(clock_list), sum(clock_list)/float(len(clock_list)))
print "Time Min: %f Max: %f Average: %f" %(min(time_list), max(time_list), sum(time_list)/float(len(time_list)))
I am not work an Swiss labs but I've tested..
Based of this question : time.clock() is better than time.time()
Edit : time.clock() is internal counter so can't use outside, got limitations max 32BIT FLOAT, can't continued counting if not store first/last values. Can't merge another one counter...

Comparing test result between Ubuntu Linux and Windows 7.
On Ubuntu
>>> start = time.time(); time.sleep(0.5); (time.time() - start)
0.5005500316619873
On Windows 7
>>> start = time.time(); time.sleep(0.5); (time.time() - start)
0.5