python-peps/pep-0418.txt

PEP: 418
Title: Add monotonic time, performance counter and process time functions
Version: $Revision$
Last-Modified: $Date$
Author: Jim Jewett <jimjjewett@gmail.com>, Victor Stinner <victor.stinner@gmail.com>
Status: Draft
Type: Standards Track
Content-Type: text/x-rst
Created: 26-March-2012
Python-Version: 3.3


Abstract
========

Add time.get_clock_info(name), time.monotonic(), time.perf_counter() and
time.process_time() functions to Python 3.3.


Rationale
=========

If a program uses the system clock to schedule events or to implement a
timeout, it will not run events at the right moment or stop the timeout too
early or too late when the system clock is set manually or adjusted
automatically by NTP. A monotonic clock should be used instead to not be
affected by system clock updates.

To measure the performance of a function, time.clock() can be used but it
is very different on Windows and on Unix. On Windows, time.clock() includes
time elapsed during sleep, whereas it does not on Unix. time.clock() precision
is very good on Windows, but very bad on Unix. A new time.perf_counter() should
be used instead to always get the most precise performance counter with a
portable behaviour.

To measure CPU time, Python does not provide directly a portable function.
time.clock() can be used on Unix, but it has a bad precision.
resource.getrusage() can also be used on Unix, but it requires to get fields of
a structure and compute the sum of time spend in kernel space and user space.
A new time.process_time() function is portable counter, always measure CPU time
(don't include time elapsed during sleep), and has the best available
precision.

Each operating system implements clocks and performance counters differently,
and it is useful to know exactly which function is used and some properties of
the clock like its resolution and its precision. A new time.get_clock_info()
function gives access to all available information of each Python time
function.

New functions:

 * time.monotonic(): timeout and scheduling, not affected by system clock
   updates
 * time.perf_counter(): benchmarking, most precise clock for short period
 * time.process_time(): profiling, CPU time of the process

time.clock() is deprecated by this PEP because it is not portable:
time.perf_counter() or time.process_time() should be used instead, depending
on your requirements.


Python functions
================

New functions
-------------

time.get_clock_info(name)
^^^^^^^^^^^^^^^^^^^^^^^^^

Get information on the specified clock. Supported clocks:

 * "clock": time.clock()
 * "monotonic": time.monotonic()
 * "perf_counter": time.perf_counter()
 * "process_time": time.process_time()
 * "time": time.time()

Return a dictionary with the following keys:

 * Mandatory keys:

   * "function" (str): name of the underlying operating system function.
     Examples: "QueryPerformanceCounter()", "clock_gettime(CLOCK_REALTIME)".
   * "resolution" (float): resolution in seconds of the clock
   * "is_monotonic" (bool): True if the clock cannot go backward

 * Optional keys:

   * "precision" (float): precision in seconds of the clock
   * "is_adjusted" (bool): True if the clock can be adjusted (e.g. by a NTP
     daemon)


time.monotonic()
^^^^^^^^^^^^^^^^

Monotonic clock, cannot go backward. It is not affected by system clock
updates. The reference point of the returned value is undefined so only the
difference between consecutive calls is valid.

Availability: Windows, Mac OS X, Unix.

Pseudo-code [#pseudo]_::

    if os.name == 'nt':
        # GetTickCount64() requires Windows Vista, Server 2008 or later
        if hasattr(time, '_GetTickCount64'):
            def monotonic():
                return _time.GetTickCount64() * 1e-3
        else:
            def monotonic():
                ticks = _time.GetTickCount()
                if ticks < monotonic.last:
                    # Integer overflow detected
                    monotonic.delta += 2**32
                monotonic.last = ticks
                return (ticks + monotonic.delta) * 1e-3
            monotonic.last = 0
            monotonic.delta = 0

    elif os.name == 'mac':
        def monotonic():
            if monotonic.factor is None:
                factor = _time.mach_timebase_info()
                monotonic.factor = timebase[0] / timebase[1]
            return _time.mach_absolute_time() * monotonic.factor
        monotonic.factor = None

    elif hasattr(time, "clock_gettime"):
        def monotonic():
            if monotonic.use_clock_highres:
                try:
                    time.clock_gettime(time.CLOCK_HIGHRES)
                except OSError:
                    monotonic.use_clock_highres = False
            return time.clock_gettime(time.CLOCK_MONOTONIC)
        monotonic.use_clock_highres = (hasattr(time, 'clock_gettime')
                                       and hasattr(time, 'CLOCK_HIGHRES'))


On Windows, QueryPerformanceCounter() is not used even though it has a better
precision than GetTickCount().  It is not reliable and has too many issues.

.. note::

   time.monotonic() detects GetTickCount() integer overflow (32 bits, roll-over
   after 49.7 days): it increases a delta by 2\ :sup:`32` each time than an
   overflow is detected.  The delta is stored in the process-local state and so
   the value of time.monotonic() may be different in two Python processes
   running for more than 49 days.


time.perf_counter()
^^^^^^^^^^^^^^^^^^^

Performance counter used for benchmarking. It is monotonic, cannot go backward.
It does include time elapsed during sleep. The reference point of the returned
value is undefined so only the difference between consecutive calls is valid
and is number of seconds.

Pseudo-code::

    def perf_counter():
        if perf_counter.use_performance_counter:
            if perf_counter.perf_frequency is None:
                try:
                    perf_counter.perf_frequency = float(_time.QueryPerformanceFrequency())
                except OSError:
                    # QueryPerformanceFrequency() fails if the installed
                    # hardware does not support a high-resolution performance
                    # counter
                    perf_counter.use_performance_counter = False
                else:
                    return _time.QueryPerformanceCounter() / perf_counter.perf_frequency
            else:
                return _time.QueryPerformanceCounter() / perf_counter.perf_frequency
        if perf_counter.use_monotonic:
            # Monotonic clock is preferred over system clock
            try:
                return time.monotonic()
            except OSError:
                perf_counter.use_monotonic = False
        return time.time()
    perf_counter.use_performance_counter = (os.name == 'nt')
    if perf_counter.use_performance_counter:
        perf_counter.perf_frequency = None
    perf_counter.use_monotonic = hasattr(time, 'monotonic')


time.process_time()
^^^^^^^^^^^^^^^^^^^

Process time used for profiling: some of kernel and user-space CPU
time. It does not include time elapsed during sleep. The reference point of
the returned value is undefined so only the difference between consecutive
calls is valid.

Pseudo-code::

    if os.name == 'nt':
        def process_time():
            handle = win32process.GetCurrentProcess()
            process_times = win32process.GetProcessTimes(handle)
            return (process_times['UserTime'] + process_times['KernelTime']) * 1e-7
    elif (hasattr(time, 'clock_gettime')
          and hasattr(time, 'CLOCK_PROCESS_CPUTIME_ID')):
        def process_time():
            return time.clock_gettime(time.CLOCK_PROCESS_CPUTIME_ID)
    else:
        try:
            import resource
        except ImportError:
            def process_time():
                return _time.clock()
        else:
            def process_time():
                usage = resource.getrusage(resource.RUSAGE_SELF)
                return usage[0] + usage[1]


Existing functions
------------------

time.time()
^^^^^^^^^^^

The system time is the "wall clock". It can be set manually by the
system administrator or automatically by a NTP daemon.  It can jump
backward and forward.  It is not monotonic.

It is available on all platforms and cannot fail.

Pseudo-code [#pseudo]_::

    if os.name == "nt":
        def time():
            return _time.GetSystemTimeAsFileTime()
    else:
        def time():
            if hasattr(time, "clock_gettime"):
                try:
                    # resolution = 1 nanosecond
                    return time.clock_gettime(time.CLOCK_REALTIME)
                except OSError:
                    # CLOCK_REALTIME is not supported (unlikely)
                    pass
            if hasattr(_time, "gettimeofday"):
                try:
                    # resolution = 1 microsecond
                    return _time.gettimeofday()
                except OSError:
                    # gettimeofday() should not fail
                    pass
            if hasattr(_time, "ftime"):
                # resolution = 1 millisecond
                return _time.ftime()
            else:
                # resolution = 1 second
                return _time.time()


time.sleep()
^^^^^^^^^^^^

Suspend execution for the given number of seconds.  The actual suspension time
may be less than that requested because any caught signal will terminate the
time.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.

Pseudo-code [#pseudo]_::

    try:
        import select
    except ImportError:
        has_select = False
    else:
        has_select = hasattr(select, "select")

    if has_select:
        def sleep(seconds):
            return select.select([], [], [], seconds)

    elif hasattr(_time, "delay"):
        def sleep(seconds):
            milliseconds = int(seconds * 1000)
            _time.delay(milliseconds)

    elif os.name == "nt":
        def sleep(seconds):
            milliseconds = int(seconds * 1000)
            win32api.ResetEvent(hInterruptEvent);
            win32api.WaitForSingleObject(sleep.sigint_event, milliseconds)

        sleep.sigint_event = win32api.CreateEvent(NULL, TRUE, FALSE, FALSE)
        # SetEvent(sleep.sigint_event) will be called by the signal handler of SIGINT

    elif os.name == "os2":
        def sleep(seconds):
            milliseconds = int(seconds * 1000)
            DosSleep(milliseconds)

    else:
        def sleep(seconds):
            seconds = int(seconds)
            _time.sleep(seconds)

Deprecated functions
--------------------

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.

Pseudo-code [#pseudo]_::

    if os.name == 'nt':
        def clock():
            if clock.use_performance_counter:
                if clock.perf_frequency is None:
                    try:
                        clock.perf_frequency = float(_time.QueryPerformanceFrequency())
                    except OSError:
                        # QueryPerformanceFrequency() fails if the installed
                        # hardware does not support a high-resolution performance
                        # counter
                        clock.use_performance_counter = False
                    else:
                        return _time.QueryPerformanceCounter() / clock.perf_frequency
                else:
                    return _time.QueryPerformanceCounter() / clock.perf_frequency
            return _time.clock()
        clock.use_performance_counter = True
        clock.perf_frequency = None
    else:
        clock = _time.clock



Glossary
========

:Accuracy:
   Is the answer correct?  Any clock will eventually <drift>; if a clock is
   intended to match <Civil Time>, it will need to be <adjusted> back to the
   "true" time.

:Adjusted:
   Resetting a clock to the correct time.  This may be done either with a
   <Step> or by <Slewing>.

:Civil Time:
   Time of day; external to the system.  10:45:13am is a Civil time; 45 seconds
   is not.  Provided by existing function time.localtime() and time.gmtime().
   Not changed by this PEP.

:Clock:
   An instrument for measuring time.  Different clocks have different
   characteristics; for example, a clock with <nanonsecond> <precision> may
   start to <drift> after a few minutes, while a less precise clock remained
   accurate for days.  This PEP is primarily concerned with clocks which use a
   unit of seconds.

:Counter:
   A clock which increments each time a certain event occurs.  A counter is
   <strictly monotonic>, but not <clock_monotonic>.  It can be used to generate
   a unique (and ordered) timestamp, but these timestamps cannot be mapped to
   <civil time>; tick creation may well be bursty, with several advances in the
   same millisecond followed by several days without any advance.

:CPU Time:
   A measure of how much CPU effort has been spent on a certain task.  CPU
   seconds are often normalized (so that a variable number can occur in the
   same actual second).  CPU seconds can be important when profiling, but they
   do not map directly to user response time, nor are they directly comparable
   to (real time) seconds.

:Duration:
   Elapsed time.  The difference between the starting and ending times.  A
   defined <epoch> creates an implicit (and usually large) duration.  More
   precision can generally be provided for a relatively small <duration>.

:Drift:
   The accumulated error against "true" time, as defined externally to the
   system.

:Epoch:
   The reference point of a clock.  For clocks providing <civil time>, this is
   often midnight as the day (and year) rolled over to January 1, 1970.  For a
   <clock_monotonic> clock, the epoch may be undefined (represented as None).

:Latency:
   Delay.  By the time a clock call returns, the <real time> has advanced,
   possibly by more than the precision of the clock.

:Monotonic:
   The characteristics expected of a monotonic clock in practice.  Moving in at
   most one direction; for clocks, that direction is forward. The <clock>
   should also be <steady>, and should be convertible to a unit of seconds.
   The tradeoffs often include lack of a defined <epoch> or mapping to <Civil
   Time>, and being more expensive (in <latency>, power usage, or <duration>
   spent within calls to the clock itself) to use.  For example, the clock may
   represent (a constant multiplied by) ticks of a specific quartz timer on a
   specific CPU core, and calls would therefore require synchronization between
   cores.

:Precision:
   Significant Digits.  What is the smallest duration that the clock can
   distinguish?  This differs from <resolution> in that a difference greater
   than the minimum precision is actually meaningful.

:Process Time:
   Time elapsed since the process began.  It is typically measured in <CPU
   time> rather than <real time>, and typically does not advance while the
   process is suspended.

:Real Time:
   Time in the real world.  This differs from <Civil time> in that it is not
   <adjusted>, but they should otherwise advance in lockstep.  It is not
   related to the "real time" of "Real Time [Operating] Systems".  It is
   sometimes called "wall clock time" to avoid that ambiguity; unfortunately,
   that introduces different ambiguities.

:Resolution:
   Represented Digits.  Note that many clocks will have a resolution greater
   than their actual <precision>.

:Slew:
   A slight change to a clock's speed, usually intended to correct <drift> with
   respect to an external authority.

:Stability:
   Persistence of accuracy.  A measure of expected <drift>.

:Steady:
   A clock with high <stability> and relatively high <accuracy> and
   <precision>.  In practice, it is often used to indicate a <clock_monotonic>
   clock, but places greater emphasis on the consistency of the duration
   between subsequent ticks.

:Step:
   An instantaneous change in the represented time.  Instead of speeding or
   slowing the clock (<slew>), a single offset is permanently added.

:System Time:
   Time as represented by the Operating System.

:Thread Time:
   Time elapsed since the thread began.  It is typically measured in <CPU time>
   rather than <real time>, and typically does not advance while the thread is
   idle.

:Wallclock:
   What the clock on the wall says.  This is typically used as a synonym for
   <real time>; unfortunately, wall time is itself ambiguous.


Hardware clocks
===============

List of hardware clocks
-----------------------

* HPET: An High Precision Event Timer (HPET) chip consists of a 64-bit up-counter (main counter)
  counting at least at 10 MHz and a set of up to 256 comparators (at
  least 3).  Each HPET can have up to 32 timers. HPET can cause around 3 seconds of drift per day.
* TSC (Time Stamp Counter): Historically, the TSC increased with every internal
  processor clock cycle, but now the rate is usually constant (even if the
  processor changes frequency) and usually equals the maximum processor
  frequency. Multiple cores having different TSC values. Hibernation of system
  will reset TSC value. The RDTSC instruction can be used to read this counter.
  CPU frequency scaling for power saving.
* ACPI Power Management Timer: ACPI 24-bit timer with a frequency
  of 3.5 MHz (3,579,545 Hz).
* Cyclone: The Cyclone timer uses a 32-bit counter on IBM Extended
  X-Architecture (EXA) chipsets which include computers that use the
  IBM "Summit" series chipsets (ex: x440).  This is available in IA32
  and IA64 architectures.
* PIT (programmable interrupt timer): Intel 8253/8254 chipsets with a
  configurable frequency in range 18.2 Hz - 1.2 MHz. It is a 16-bit counter.
* RTC (Real-time clock). Most RTCs use a crystal oscillator with a frequency of
  32,768 Hz


Linux clocksource
-----------------

There were 4 implementations of the time in the Linux kernel: UTIME
(1996), timer wheel (1997), HRT (2001) and hrtimers (2007).  The
latter is the result of the "high-res-timers" project started by
George Anzinger in 2001, with contributions by Thomas Gleixner and
Douglas Niehaus.  hrtimers implementation was merged into Linux
2.6.21, released in 2007.

hrtimers supports various clock sources.  It sets a priority to each
source to decide which one will be used. Linux supports the following
clock sources:

* tsc
* hpet
* pit
* pmtmr: ACPI Power Management Timer
* cyclone

High-resolution timers are not supported on all hardware
architectures.  They are at least provided on x86/x86_64, ARM and
PowerPC.

The /sys/devices/system/clocksource/clocksource0 directory contains two useful files:

 * available_clocksource: list of available clock sources
 * current_clocksource: clock source currently used. It is possible to change
   the current clocksource by writing the name of a clocksource into this
   file.

/proc/timer_list contains the list of all hardware timers.

Read also the `time(7) manual page
<http://www.kernel.org/doc/man-pages/online/pages/man7/time.7.html>`_:
"overview of time and timers".


FreeBSD timecounter
-------------------

kern.timecounter.choice list available hardward clocks with their priority.
The sysctl program can be used to change the timecounter. Example::

    # dmesg|grep Timecounter
    Timecounter "i8254" frequency 1193182 Hz quality 0
    Timecounter "ACPI-safe" frequency 3579545 Hz quality 850
    Timecounter "HPET" frequency 100000000 Hz quality 900
    Timecounter "TSC" frequency 3411154800 Hz quality 800
    Timecounters tick every 10.000 msec
    # sysctl kern.timecounter.choice
    kern.timecounter.choice: TSC(800) HPET(900) ACPI-safe(850) i8254(0) dummy(-1000000)
    # sysctl kern.timecounter.hardware="ACPI-fast"
    kern.timecounter.hardware: HPET -> ACPI-fast

Available clocks:

 * "TSC": Time Stamp Counter of the procesor
 * "HPET": High Precision Event Timer
 * "ACPI-fast": ACPI Power Management timer (fast mode)
 * "ACPI-safe": ACPI Power Management timer (safe mode)
 * "i8254": PIT with Intel 8254 chipset

The `commit 222222
<http://svnweb.freebsd.org/base?view=revision&revision=222222>`_ (May 2011)
decreased ACPI-fast timecounter quality to 900 and increased HPET timecounter
quality to 950: "HPET on modern platforms usually have better resolution and
lower latency than ACPI timer".

Read `Timecounters: Efficient and precise timekeeping in SMP kernels
<http://phk.freebsd.dk/pubs/timecounter.pdf>`_ by Poul-Henning Kamp (2002) for
the FreeBSD Project.


Performances
------------

Reading an hardware clock has a cost. The following table compares the
performance of different hardware clocks on Linux 3.3 with Intel Core i7-2600
at 3.40GHz (8 cores).

========================  ======  =======  ======
Function                  TSC     ACPI PM  HPET
========================  ======  =======  ======
time()                      2 ns     2 ns    2 ns
CLOCK_REALTIME_COARSE      10 ns    10 ns   10 ns
CLOCK_MONOTONIC_COARSE     12 ns    13 ns   12 ns
CLOCK_THREAD_CPUTIME_ID   134 ns   135 ns  135 ns
CLOCK_PROCESS_CPUTIME_ID  127 ns   129 ns  129 ns
clock()                   146 ns   146 ns  143 ns
gettimeofday()             23 ns   726 ns  637 ns
CLOCK_MONOTONIC_RAW        31 ns   716 ns  607 ns
CLOCK_REALTIME             27 ns   707 ns  629 ns
CLOCK_MONOTONIC            27 ns   723 ns  635 ns
========================  ======  =======  ======

Each function was called 10,000,000 times and CLOCK_MONOTONIC was used to get
the time before and after. The benchmark was run 5 times to keep the minimum
time.

FreeBSD 8.0 in kvm with hardware virtualization:

========================  ======  =========  =======  =======
Function                  TSC     ACPI-Safe  HPET     i8254
========================  ======  =========  =======  =======
time()                    191 ns    188 ns    189 ns   188 ns
CLOCK_SECOND              187 ns    184 ns    187 ns   183 ns
CLOCK_REALTIME_FAST       189 ns    180 ns    187 ns   190 ns
CLOCK_UPTIME_FAST         191 ns    185 ns    186 ns   196 ns
CLOCK_MONOTONIC_FAST      188 ns    187 ns    188 ns   189 ns
CLOCK_THREAD_CPUTIME_ID   208 ns    206 ns    207 ns   220 ns
CLOCK_VIRTUAL             280 ns    279 ns    283 ns   296 ns
CLOCK_PROF                289 ns    280 ns    282 ns   286 ns
clock()                   342 ns    340 ns    337 ns   344 ns
CLOCK_UPTIME_PRECISE      197 ns  10380 ns   4402 ns  4097 ns
CLOCK_REALTIME            196 ns  10376 ns   4337 ns  4054 ns
CLOCK_MONOTONIC_PRECISE   198 ns  10493 ns   4413 ns  3958 ns
CLOCK_UPTIME              197 ns  10523 ns   4458 ns  4058 ns
gettimeofday()            202 ns  10524 ns   4186 ns  3962 ns
CLOCK_REALTIME_PRECISE    197 ns  10599 ns   4394 ns  4060 ns
CLOCK_MONOTONIC           201 ns  10766 ns   4498 ns  3943 ns
========================  ======  =========  =======  =======

Each function was called 100,000 times and CLOCK_MONOTONIC was used to get
the time before and after. The benchmark was run 5 times to keep the minimum
time.



NTP adjustment
==============

NTP has diffent methods to adjust a clock:

 * "slewing": change the clock frequency to be slightly faster or slower
   (which is done with adjtime()). Since the slew rate is limited to 0.5 ms/s,
   each second of adjustment requires an amortization interval of 2000 s. Thus,
   an adjustment of many seconds can take hours or days to amortize.
 * "stepping": jump by a large amount in a single discrete step (which is done
   with settimeofday())

By default, the time is slewed if the offset is less than 128 ms, or stepped
otherwise.

Slewing is generally desirable (i.e. we should use CLOCK_MONOTONIC, not
CLOCK_MONOTONIC_RAW) if one wishes to measure "real" time (and not a time-like
object like CPU cycles). This is because the clock on the other end of the NTP
connection from you is probably better at keeping time: hopefully that thirty
five thousand dollars of Cesium timekeeping goodness is doing something better
than your PC's $3 quartz crystal, after all.

Get more detail in the `documentation of the NTP daemon
<http://doc.ntp.org/4.1.2/ntpd.htm>`_.


Operating system clocks
=======================

Monotonic clocks
----------------

=========================  ==========  ===============  =============  ===============
Name                       Resolution  Adjusted         Include sleep  Include suspend
=========================  ==========  ===============  =============  ===============
gethrtime                  1 ns        No               Yes            Yes
CLOCK_HIGHRES              1 ns        No               Yes            ?
CLOCK_MONOTONIC            1 ns        Slewed on Linux  Yes            No
CLOCK_MONOTONIC_RAW        1 ns        No               Yes            No
mach_absolute_time()       1 ns        No               Yes            No
QueryPerformanceCounter()  \-          No               Yes            ?
GetTickCount[64]()         1 ms        No               Yes            Yes
timeGetTime()              1 ms        No               Yes            ?
=========================  ==========  ===============  =============  ===============

Examples of clock precision on x86_64:

=========================  ================  ===============
Name                       Operating system  Precision
=========================  ================  ===============
CLOCK_MONOTONIC_RAW        Linux 3.2                    1 ns
CLOCK_MONOTONIC            Linux 3.2                    1 ns
CLOCK_HIGHRES              SunOS 5.11                   2 ns
CLOCK_MONOTONIC            SunOS 5.11                   2 ns
QueryPerformanceCounter    Windows Seven               10 ns
CLOCK_MONOTONIC            FreeBSD 8.2                 11 ns
CLOCK_MONOTONIC            OpenBSD 5.0                 10 ms
GetTickCount               Windows Seven             15.6 ms
=========================  ================  ===============

For CLOCK_MONOTONIC and CLOCK_MONOTONIC_RAW, the precision of this table is the
result of clock_getres(). It looks like Linux does not implement
clock_getres() and always return 1 nanosecond.


mach_absolute_time
^^^^^^^^^^^^^^^^^^

Mac OS X provides a monotonic clock: mach_absolute_time().  It is
based on absolute elapsed time delta since system boot.  It is not
adjusted and cannot be set.

mach_timebase_info() gives a fraction to convert the clock value to a
number of nanoseconds.  According to the documentation (`Technical Q&A
QA1398 <https://developer.apple.com/library/mac/#qa/qa1398/>`_),
mach_timebase_info() is always equal to one and never fails, even
if the function may fail according to its prototype.

mach_absolute_time() stops during a sleep on a PowerPC CPU, but not on
an Intel CPU: `Different behaviour of mach_absolute_time() on i386/ppc
<http://lists.apple.com/archives/PerfOptimization-dev/2006/Jul/msg00024.html>`_.

mach_absolute_time() has a resolution of 1 nanosecond.

CLOCK_MONOTONIC, CLOCK_MONOTONIC_RAW
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

CLOCK_MONOTONIC and CLOCK_MONOTONIC_RAW represent monotonic time since
some unspecified starting point.  They cannot be set.

Documentation: refer to the manual page of your operating system.
Examples:

* `FreeBSD clock_gettime() manual page
  <http://www.freebsd.org/cgi/man.cgi?query=clock_gettime>`_
* `Linux clock_gettime() manual page
  <http://linux.die.net/man/3/clock_gettime>`_

CLOCK_MONOTONIC is available at least on the following operating systems:

* DragonFly BSD, FreeBSD >= 5.0, OpenBSD, NetBSD
* Linux
* Solaris

The following operating systems don't support CLOCK_MONOTONIC:

* GNU/Hurd (see `open issues/ clock_gettime
  <http://www.gnu.org/software/hurd/open_issues/clock_gettime.html>`_)
* Mac OS X
* Windows

On Linux, NTP may adjust the CLOCK_MONOTONIC rate (slewed), but it cannot
jump backward.

CLOCK_MONOTONIC_RAW is specific to Linux. It is similar to CLOCK_MONOTONIC, but
provides access to a raw hardware-based time that is not subject to NTP
adjustments. CLOCK_MONOTONIC_RAW requires Linux 2.6.28 or later.

CLOCK_MONOTONIC stops while the machine is suspended.

clock_gettime() fails if the system does not support the specified
clock, even if the standard C library supports it.  For example,
CLOCK_MONOTONIC_RAW requires a kernel version 2.6.28 or later.

clock_getres() gives the clock resolution.  It is 1 nanosecond on
Linux.

.. note::

   clock_gettime() requires to link the program against the rt
   (real-time) library.

.. note::

   Linux provides also CLOCK_MONOTONIC_COARSE since Linux 2.6.32 which has less
   accurate than CLOCK_MONOTONIC but is faster.



Windows: QueryPerformanceCounter
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

High-resolution performance counter.  It is monotonic.
QueryPerformanceFrequency() gives its frequency.

It has a much higher resolution, but has lower long term precision than
GetTickCount() and timeGetTime() clocks.  For example, it will drift
compared to the low precision clocks.

Documentation:

* `MSDN: QueryPerformanceCounter() documentation
  <http://msdn.microsoft.com/en-us/library/windows/desktop/ms644904%28v=vs.85%29.aspx>`_
* `MSDN: QueryPerformanceFrequency() documentation
  <http://msdn.microsoft.com/en-us/library/windows/desktop/ms644905%28v=vs.85%29.aspx>`_

Hardware clocks used by QueryPerformanceCounter:

* Windows XP: RDTSC instruction of Intel processors, the clock
  frequency is the frequency of the processor (between 200 MHz and 3
  GHz, usually greater than 1 GHz nowadays).
* Windows 2000: ACPI power management timer, frequency = 3,549,545 Hz.
  It can be forced through the "/usepmtimer" flag in boot.ini.

.. * Windows 95/98: 8245 PIT chipset, frequency = 1,193,181 Hz

QueryPerformanceFrequency() should only be called once: the frequency
will not change while the system is running.  It fails if the
installed hardware does not support a high-resolution performance
counter.

QueryPerformanceCounter() cannot be adjusted:
`SetSystemTimeAdjustment()
<http://msdn.microsoft.com/en-us/library/windows/desktop/ms724943(v=vs.85).aspx>`_
only adjusts the system time.

Bugs:

* The performance counter value may unexpectedly leap forward because
  of a hardware bug, see `KB274323`_.
* On VirtualBox, QueryPerformanceCounter() does not increment the high
  part every time the low part overflows, see `Monotonic timers
  <http://code-factor.blogspot.fr/2009/11/monotonic-timers.html>`_
  (2009).
* VirtualBox had a bug in its HPET virtualized device:
  QueryPerformanceCounter() did jump forward by approx. 42 seconds (`issue
  #8707 <https://www.virtualbox.org/ticket/8707>`_).
* Windows XP had a bug (see `KB896256`_): on a multiprocessor
  computer, QueryPerformanceCounter() returned a different value for
  each processor.  The bug was fixed in Windows XP SP2.
* Issues with processor with variable frequency: the frequency is changed
  depending on the workload to reduce memory consumption.
* Chromium don't use QueryPerformanceCounter() on Athlon X2 CPUs (model 15)
  because "QueryPerformanceCounter is unreliable" (see base/time_win.cc in
  Chromium source code)

.. _KB896256: http://support.microsoft.com/?id=896256
.. _KB274323: http://support.microsoft.com/?id=274323


Windows: GetTickCount(), GetTickCount64()
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

GetTickCount() and GetTickCount64() are monotonic, cannot fail and are
not adjusted by SetSystemTimeAdjustment().  MSDN documentation:
`GetTickCount()
<http://msdn.microsoft.com/en-us/library/windows/desktop/ms724408(v=vs.85).aspx>`_,
`GetTickCount64()
<http://msdn.microsoft.com/en-us/library/windows/desktop/ms724411(v=vs.85).aspx>`_.

The elapsed time retrieved by GetTickCount() or GetTickCount64()
includes time the system spends in sleep or hibernation.

GetTickCount64() was added to Windows Vista and Windows Server 2008.

The clock resolution is 1 millisecond.  Its precision is usually around
15 ms.  It is possible to improve the precision using the `undocumented
NtSetTimerResolution() function
<http://undocumented.ntinternals.net/UserMode/Undocumented%20Functions/Time/NtSetTimerResolution.html>`_.
There are applications using this undocumented function, example:
`Timer Resolution <http://www.lucashale.com/timer-resolution/>`_.

WaitForSingleObject() use the same timer than GetTickCount() with the same
resolution.

GetTickCount() has an precision of 55 ms on Windows 9x.


Windows: timeGetTime
^^^^^^^^^^^^^^^^^^^^

The timeGetTime function retrieves the system time, in milliseconds.
The system time is the time elapsed since Windows was started.  Read
the `timeGetTime() documentation
<http://msdn.microsoft.com/en-us/library/windows/desktop/dd757629(v=vs.85).aspx>`_.

The return type of timeGetTime() is a 32-bit unsigned integer.  As
GetTickCount(), timeGetTime() rolls over after 2^32 milliseconds (49.7
days).

The elapsed time retrieved by timeGetTime() includes time the system spends in
sleep.

The default precision of the timeGetTime function can be five
milliseconds or more, depending on the machine.

timeBeginPeriod() can be used to increase the precision of timeGetTime() up to
1 millisecond, but it negatively affects power consumption. Calling
timeBeginPeriod() also affects the granularity of some other timing calls, such
as CreateWaitableTimer(), WaitForSingleObject() and Sleep().

.. note::

   timeGetTime() and timeBeginPeriod() are part the Windows multimedia
   library and so require to link the program against winmm or to
   dynamically load the library.


Solaris: CLOCK_HIGHRES
^^^^^^^^^^^^^^^^^^^^^^

The Solaris OS has a CLOCK_HIGHRES timer that attempts to use an
optimal hardware source, and may give close to nanosecond resolution.
CLOCK_HIGHRES is the nonadjustable, high-resolution clock.  For timers
created with a clockid_t value of CLOCK_HIGHRES, the system will
attempt to use an optimal hardware source.

Solaris: gethrtime
^^^^^^^^^^^^^^^^^^

The gethrtime() function returns the current high-resolution real
time.  Time is expressed as nanoseconds since some arbitrary time in
the past; it is not correlated in any way to the time of day, and thus
is not subject to resetting or drifting by way of adjtime() or
settimeofday().  The hires timer is ideally suited to performance
measurement tasks, where cheap, accurate interval timing is required.

The linearity of gethrtime() is not preserved accross cpr
suspend-resume cycle (`Bug 4272663
<http://wesunsolve.net/bugid/id/4272663>`_).

Read the `gethrtime() manual page of Solaris 11
<http://docs.oracle.com/cd/E23824_01/html/821-1465/gethrtime-3c.html#scrolltoc>`_.

On Solaris, gethrtime() is the same as clock_gettime(CLOCK_MONOTONIC).


System time clocks
------------------

========================= ===============  =============
Name                      Resolution       Include sleep
========================= ===============  =============
CLOCK_REALTIME            1 ns             Yes
GetSystemTimeAsFileTime   100 ns           Yes
gettimeofday()            1 µs             Yes
ftime()                   1 ms             Yes
time()                    1 sec            Yes
========================= ===============  =============

Examples of clock precision on x86_64:

=========================  ================  ===============
Name                       Operating system  Precision
=========================  ================  ===============
CLOCK_REALTIME             Linux 3.2         1 ns
CLOCK_REALTIME             FreeBSD 8.2       11 ns
CLOCK_REALTIME             SunOS 5.11        10 ms
CLOCK_REALTIME             OpenBSD 5.0       10 ms
GetSystemTimeAsFileTime    Windows Seven     15.6 ms
=========================  ================  ===============

For CLOCK_REALTIME, the precision of this table is the result of clock_getres().
It looks like Linux does not implement clock_getres() and always return 1
nanosecond.

.. note::

   Linux provides also CLOCK_REALTIME_COARSE since Linux 2.6.32 which has less
   accurate than CLOCK_REALTIME but is faster.


Windows: GetSystemTimeAsFileTime
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

The system time can be read using GetSystemTimeAsFileTime(), ftime()
and time().

The system time resolution can be read using
GetSystemTimeAdjustment().  The precision is usually between 1
millisecond and 15 milliseconds.  Resolution:

* GetSystemTimeAsFileTime(): 100 nanoseconds
* ftime(): 1 millisecond
* time(): 1 second

Read the `GetSystemTimeAsFileTime() documentation
<http://msdn.microsoft.com/en-us/library/windows/desktop/ms724397(v=vs.85).aspx>`_.

The system time can be set using SetSystemTime().

System time on UNIX
^^^^^^^^^^^^^^^^^^^

gettimeofday(), ftime(), time() and clock_gettime(CLOCK_REALTIME)
return the system clock.

Resolution:

* clock_gettime(): clock_getres(CLOCK_REALTIME)
* gettimeofday(): 1 microsecond
* ftime(): 1 millisecond
* time(): 1 second

The system time can be set using settimeofday() or
clock_settime(CLOCK_REALTIME).


Process time
------------

The process time cannot be set. It is not monotonic: the clocks stop while the
process is idle.

=========================  ===============  =============
Name                       Resolution       Include sleep
=========================  ===============  =============
GetProcessTimes()          100 ns           No
CLOCK_PROCESS_CPUTIME_ID   1 ns             No
getrusage()                1 µs             No
clock()                    \-               No
=========================  ===============  =============

Examples of clock precision on x86_64:

=========================  ================  ===============
Name                       Operating system  Precision
=========================  ================  ===============
CLOCK_PROCESS_CPUTIME_ID   Linux 3.2                    1 ns
clock()                    SunOS 5.11                   1 µs
getrusage()                Linux 3.0                    4 ms
clock()                    FreeBSD 8.2                7.8 ms
clock()                    Linux 3.2                   10 ms
clock()                    OpenBSD 5.0                 10 ms
GetProcessTimes()          Windows Seven             15.6 ms
=========================  ================  ===============

The precision of clock() in this table is the result of 1 / CLOCKS_PER_SEC.
For CLOCK_PROCESS_CPUTIME_ID, the precision of this table is the result of
clock_getres(). It looks like Linux does not implement clock_getres() and
always return 1 nanosecond. For GetProcessTimes(), the precision is read using
GetSystemTimeAdjustment().


Functions
^^^^^^^^^

* Windows: GetProcessTimes()
* clock_gettime(CLOCK_PROCESS_CPUTIME_ID): High-resolution per-process
  timer from the CPU.
* clock():

  * Windows: The elapsed wall-clock time since the start of the
    process (elapsed time in seconds times CLOCKS_PER_SEC).  It can
    fail.
  * UNIX: returns an approximation of processor time used by the
    program.

* times()
* getrusage(): ru_utime and ru_stime fields

Resolution:

* clock() rate is CLOCKS_PER_SEC.  It was called CLK_TCK in Microsoft
  C before 6.0.
* The clock resolution can be read using clock_getres().
* GetProcessTimes(): call GetSystemTimeAdjustment().

Python source code includes a portable library to get the process time (CPU
time): `Tools/pybench/systimes.py
<http://hg.python.org/cpython/file/tip/Tools/pybench/systimes.py>`_.


Thread time
-----------

The thread time cannot be set. It is not monotonic: the clocks stop while the
thread is idle.

=========================  ===============
Name                       Resolution
=========================  ===============
GetThreadTimes()           100 ns
CLOCK_THREAD_CPUTIME_ID    1 ns
=========================  ===============

Examples of clock precision on x86_64:

=========================  ================  ===============
Name                       Operating system  Precision
=========================  ================  ===============
CLOCK_THREAD_CPUTIME_ID    Linux 3.2         1 ns
CLOCK_THREAD_CPUTIME_ID    FreeBSD 8.2       1 µs
GetThreadTimes()           Windows Seven     15.6 ms
=========================  ================  ===============

For CLOCK_THREAD_CPUTIME_ID, the precision of this table is the result of
clock_getres(). It looks like Linux does not implement clock_getres() and
always return 1 nanosecond. For GetThreadTimes(), the precision is read using
GetSystemTimeAdjustment().

Functions
^^^^^^^^^

* Windows: GetThreadTimes()
* clock_gettime(CLOCK_THREAD_CPUTIME_ID): Thread-specific CPU-time
  clock.

Resolution:

* CLOCK_THREAD_CPUTIME_ID: call clock_getres().
* GetThreadTimes(): call GetSystemTimeAdjustment()

See also pthread_getcpuclockid().


Windows: QueryUnbiasedInterruptTime
-----------------------------------

Gets the current unbiased interrupt time from the biased interrupt
time and the current sleep bias amount.  This time is not affected by
power management sleep transitions.

The elapsed time retrieved by the QueryUnbiasedInterruptTime function
includes only time that the system spends in the working state.
QueryUnbiasedInterruptTime() is not monotonic.

QueryUnbiasedInterruptTime() was introduced in Windows 7.


Sleep, suspend and monotonic time
=================================

Linux
-----

sleep() is not affected by system clock update.

On Linux, CLOCK_MONOTONIC and CLOCK_MONOTONIC_RAW includes time the system
spends in sleep; but it does not include time spent in hibernation (ACPI S3
mode). If the system clock jumps backward, CLOCK_MONOTONIC and
CLOCK_MONOTONIC_RAW are not affected.

Linux 2.6.39 and glibc 2.14 introduces a new clock: CLOCK_BOOTTIME.
CLOCK_BOOTTIME is idential to CLOCK_MONOTONIC, except it also includes any time
spent in suspend. Read also `Waking systems from suspend
<http://lwn.net/Articles/429925/>`_ (March, 2011).

Other operating systems
-----------------------

On Windows, GetTickCount() and GetTickCount64() include time the system
spends in sleep; sleep() is not affected by system clock update.

On FreeBSD 8, CLOCK_MONOTONIC include time the system spends in sleep; sleep()
is not affected by system clock update.

On OpenIndiana, CLOCK_MONOTONIC include time the system spends in sleep;
sleep() is not affected by system clock update.

On Mac OS X, mach_absolute_time() include time the system spends in sleep;
sleep() is not affected by system clock update. Sleep is interrupted during
suspend.


Sleeping
========

Sleep can be interrupted by a signal: the function fails with EINTR.

========================  ==========
Name                      Resolution
========================  ==========
nanosleep()               1 ns
clock_nanosleep()         1 ns
usleep()                  1 µs
delay()                   1 µs
sleep()                   1 sec
========================  ==========

Other functions:

========================  ==========
Name                      Resolution
========================  ==========
sigtimedwait()            1 ns
pthread_cond_timedwait()  1 ns
sem_timedwait()           1 ns
select()                  1 µs
epoll()                   1 ms
poll()                    1 ms
WaitForSingleObject()     1 ms
========================  ==========


Classic functions
-----------------

 * sleep(seconds)
 * usleep(microseconds)
 * nanosleep(nanoseconds, remaining):
   `Linux manpage of nanosleep()
   <http://www.kernel.org/doc/man-pages/online/pages/man2/nanosleep.2.html>`_
 * delay(milliseconds)


clock_nanosleep
---------------

clock_nanosleep(clock_id, flags, nanoseconds, remaining):
`Linux manpage of clock_nanosleep()
<http://www.kernel.org/doc/man-pages/online/pages/man2/clock_nanosleep.2.html>`_.

If flags is TIMER_ABSTIME, then request is interpreted as an absolute
time as measured by the clock, clock_id.  If request is less than or
equal to the current value of the clock, then clock_nanosleep() returns
immediately without suspending the calling thread.

POSIX.1 specifies that changing the value of the CLOCK_REALTIME clock via
clock_settime(2) shall have no effect on a thread that is blocked on a relative
clock_nanosleep().


select()
--------

select(nfds, readfds, writefds, exceptfs, timeout).

Since Linux 2.6.28, select() uses high-resolution timers to handle the timeout.
A process has a "slack" attribute to configure the precision of the timeout, the
default slack is 50 microseconds. Before Linux 2.6.28, timeouts for select()
were handled by the main timing subsystem at a jiffy-level resolution. Read
also `High- (but not too high-) resolution timeouts
<http://lwn.net/Articles/296578/>`_ and `Timer slack
<http://lwn.net/Articles/369549/>`_.


Other functions
---------------

 * poll(), epoll()
 * sigtimedwait(). POSIX: "If the Monotonic Clock option is supported, the
   CLOCK_MONOTONIC clock shall be used to measure the time interval specified
   by the timeout argument."
 * pthread_cond_timedwait(), pthread_condattr_setclock(). "The default value of
   the clock attribute shall refer to the system clock."
 * sem_timedwait(): "If the Timers option is supported, the timeout shall be
   based on the CLOCK_REALTIME clock. If the Timers option is not supported,
   the timeout shall be based on the system clock as returned by the time()
   function. The resolution of the timeout shall be the resolution of the clock
   on which it is based."
 * WaitForSingleObject(): use the same timer than GetTickCount() with the same
   resolution.


Alternatives: API design
========================

Other names for new functions
-----------------------------

time.monotonic():

* time.counter()
* time.seconds()
* time.steady()
* time.timeout_clock()
* time.wallclock(): it is not the system time aka the "wall clock", but
  a monotonic clock with an unspecified starting point

The name "time.try_monotonic()" was also proposed when time.monotonic() was
falling back to the system clock when no monotonic clock was available.

time.perf_counter():

 * time.hires()
 * time.highres()
 * time.timer()


Only expose operating system clocks
-----------------------------------

To not have to define high-level clocks, which is a difficult task, a simpler
approach is to only expose operating system clocks. time.clock_gettime() and
related clock identifiers were already added to Python 3.3 for example.


Fallback to system clock
------------------------

If no monotonic clock is available, time.monotonic() falls back to the system
clock.

Issues:

 * It is hard to define correctly such function in the documentation: is it
   monotonic? is it steady? is it adjusted?
 * Some user want to decide what to do when no monotonic clock is available:
   use another clock, display an error, or do something else


One function choosing the clock from a list of constraints
----------------------------------------------------------

``time.get_clock(*flags)`` with the following flags:

 * time.MONOTONIC: clock cannot go backard
 * time.STEADY: clock rate is steady and the clock is not adjusted
 * time.HIGHRES: clock with the highest resolutions

time.get_clock() returns None if the clock is found and so calls can be chained
using the or operator. Example::

 get_time = time.get_clock(time.MONOTONIC) or time.get_clock(time.STEADY) or time.time()
 t = get_time()

Example of flags of system clocks:

 * QueryPerformanceCounter: MONOTONIC | HIGHRES
 * GetTickCount: MONOTONIC | STEADY
 * CLOCK_MONOTONIC: MONOTONIC | STEADY (or only MONOTONIC on Linux)
 * CLOCK_MONOTONIC_RAW: MONOTONIC | STEADY
 * gettimeofday(): (no flag)


One function with a flag: time.monotonic(fallback=True)
-------------------------------------------------------

 * time.monotonic(fallback=True) falls back to the system clock if no monotonic
   clock is available or if the monotonic clock failed.
 * time.monotonic(fallback=False) raises OSError if monotonic clock fails and
   NotImplementedError if the system does not provide a monotonic clock

"A keyword argument that gets passed as a constant in the caller is usually
poor API."

Raising NotImplementedError for a function is something uncommon in Python and
should be avoided.


One function, no flag
---------------------

time.monotonic() returns (time: float, is_monotonic: bool).

An alternative is to use a function attribute:
time.monotonic.is_monotonic.  The attribute value would be None before
the first call to time.monotonic().


Working around operating system bugs?
=====================================

Should Python ensure manually that a monotonic clock is truly
monotonic by computing the maximum with the clock value and the
previous value?

Since it's relatively straightforward to cache the last value returned
using a static variable, it might be interesting to use this to make
sure that the values returned are indeed monotonic.

* Virtual machines provide less reliable clocks.
* QueryPerformanceCounter() has known bugs (only one is not fixed yet)

Python may only work around a specific known operating system bug:
`KB274323`_ contains a code example to workaround the bug (use
GetTickCount() to detect QueryPerformanceCounter() leap).

Issues of a hacked monotonic function:

 * if the clock is accidentally set forward by an hour and then back
   again, you wouldn't have a useful clock for an hour
 * the cache is not shared between processes so different processes wouldn't
   see the same clock value


Footnotes
=========

.. [#pseudo] "_time" is an hypothetical module only used for the example.
   The time module is implemented in C and so there is no need for
   such module.


Links
=====

Related Python issues:

* `Issue #12822: NewGIL should use CLOCK_MONOTONIC if possible.
  <http://bugs.python.org/issue12822>`_
* `Issue #14222: Use time.steady() to implement timeout
  <http://bugs.python.org/issue14222>`_
* `Issue #14309: Deprecate time.clock()
  <http://bugs.python.org/issue14309>`_
* `Issue #14397: Use GetTickCount/GetTickCount64 instead of
  QueryPerformanceCounter for monotonic clock
  <http://bugs.python.org/issue14397>`_
* `Issue #14428: Implementation of the PEP 418
  <http://bugs.python.org/issue14428>`_
* `Issue #14555: clock_gettime/settime/getres: Add more clock identifiers
  <http://bugs.python.org/issue14555>`_

Libraries exposing monotonic clocks:

* `Java: System.nanoTime
  <http://docs.oracle.com/javase/1.5.0/docs/api/java/lang/System.html#nanoTime()>`_
* `Qt library: QElapsedTimer
  <http://qt-project.org/doc/qt-4.8/qelapsedtimer.html>`_
* `glib library: g_get_monotonic_time ()
  <http://developer.gnome.org/glib/2.30/glib-Date-and-Time-Functions.html#g-get-monotonic-time>`_
  uses GetTickCount64()/GetTickCount() on Windows,
  clock_gettime(CLOCK_MONOTONIC) on UNIX or falls back to the system
  clock
* `python-monotonic-time
  <http://code.google.com/p/python-monotonic-time/>`_ (`github
  <https://github.com/gavinbeatty/python-monotonic-time>`_)
* `Monoclock.nano_count()
  <https://github.com/ludios/Monoclock>`_ uses clock_gettime(CLOCK_MONOTONIC)
  and returns a number of nanoseconds
* `monotonic_clock <https://github.com/ThomasHabets/monotonic_clock>`_
* `Perl: Time::HiRes <http://perldoc.perl.org/Time/HiRes.html>`_
  exposes clock_gettime(CLOCK_MONOTONIC)
* `Ruby: AbsoluteTime.now
  <https://github.com/bwbuchanan/absolute_time/>`_: use
  clock_gettime(CLOCK_MONOTONIC), mach_absolute_time() or
  gettimeofday().  "AbsoluteTime.monotonic?" method indicates if
  AbsoluteTime.now is monotonic or not.
* `libpthread
  <http://code.google.com/p/libpthread/>`_: POSIX thread library for Windows
  (`clock.c <http://code.google.com/p/libpthread/source/browse/src/clock.c>`_)
* `Boost.Chrono
  <http://www.boost.org/doc/libs/1_49_0/doc/html/chrono.html>`_ uses:

  * system_clock:

    * mac = gettimeofday()
    * posix = clock_gettime(CLOCK_REALTIME)
    * win = GetSystemTimeAsFileTime()

  * steady_clock:

    * mac = mach_absolute_time()
    * posix = clock_gettime(CLOCK_MONOTONIC)
    * win = QueryPerformanceCounter()

  * high_resolution_clock:

    * steady_clock, if available system_clock, otherwise

Time:

* `hrtimers - subsystem for high-resolution kernel timers
  <http://www.kernel.org/doc/Documentation/timers/hrtimers.txt>`_
* `C++ Timeout Specification
  <http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2010/n3128.html>`_
* `Windows: Game Timing and Multicore Processors
  <http://msdn.microsoft.com/en-us/library/ee417693.aspx>`_
* `Implement a Continuously Updating, High-Resolution Time Provider
  for Windows
  <http://msdn.microsoft.com/en-us/magazine/cc163996.aspx>`_
* `clockspeed <http://cr.yp.to/clockspeed.html>`_ uses a hardware tick
  counter to compensate for a persistently fast or slow system clock
* `Retrieving system time
  <http://en.wikipedia.org/wiki/System_time#Retrieving_system_time>`_
  lists hardware clocks and time functions with their resolution and
  epoch or range
* On Windows, the JavaScript runtime of Firefox interpolates
  GetSystemTimeAsFileTime() with QueryPerformanceCounter() to get an
  higher resolution. See the `Bug 363258 - bad millisecond resolution
  for (new Date).getTime() / Date.now() on Windows
  <https://bugzilla.mozilla.org/show_bug.cgi?id=363258>`_.
* `When microseconds matter
  <http://www.ibm.com/developerworks/library/i-seconds/>`_: How the
  IBM High Resolution Time Stamp Facility accurately measures itty
  bits of time
* `Win32 Performance Measurement Options
  <http://drdobbs.com/windows/184416651>`_ by Matthew Wilson, May 01, 2003
* `Counter Availability and Characteristics for Feed-forward Based Synchronization
  <http://www.cubinlab.ee.unimelb.edu.au/~jrid/Publications/ridoux_ispcs09.pdf>`_
  by Timothy Broomhead, Julien Ridoux, Darryl Veitch (2009)
* System Management Interrupt (SMI) issues:

  * `System Management Interrupt Free Hardware
    <http://linuxplumbersconf.org/2009/slides/Keith-Mannthey-SMI-plumers-2009.pdf>`_
    (Keith Mannthey, 2009)
  * `IBM Real-Time "SMI Free" mode driver
    <http://lwn.net/Articles/318725/>`_
  * `Fixing Realtime problems caused by SMI on Ubuntu
    <http://wiki.linuxcnc.org/cgi-bin/wiki.pl?FixingSMIIssues>`_
  * `[RFC] simple SMI detector
    <http://lwn.net/Articles/316622/>`_
  * `[PATCH 2.6.34-rc3] A nonintrusive SMI sniffer for x86.
    <http://marc.info/?l=linux-kernel&m=127058720921201&w=1>`_


Copyright
=========

This document has been placed in the public domain.



..
   Local Variables:
   mode: indented-text
   indent-tabs-mode: nil
   sentence-end-double-space: t
   fill-column: 70
   coding: utf-8
   End: