Re: [HACKERS] gettimeofday is at the end of its usefulness? - Mailing list pgsql-hackers

Haribabu Kommi <kommi.haribabu@gmail.com> writes:
> Attached a patch that replaces most of the getimeofday function calls,
> except timeofday(user callable) and GetCurrentTimestamp functions.
> Didn't add any configure checks in case if the clock_gettime function is
> not available, the fallback logic to gettimeofday function call.

Well, of course, configure is the hard part.

I got interested in this area again pursuant to a question from Joel
Jacobson, and looked around to see if things had changed any since
2014.  One pleasant surprise is that Apple got around to implementing
clock_gettime() as of the current macOS release (10.12 "Sierra").
That means that pretty much all interesting platforms now have
clock_gettime(), which removes one possible objection to starting to
use it.  However, it seems like there is not a lot of commonality
to the best "clockid" to use.

In theory, according to the POSIX spec, CLOCK_MONOTONIC would be what we
want to use for time interval measurement (EXPLAIN ANALYZE), since that
would be impervious to possible clock setting changes.  But some
implementations might only optimize the more common CLOCK_REALTIME,
and I found that there are a lot of platform-specific clock IDs that
we might want to consider.

On Linux (RHEL6, 2.4GHz x86_64), I find that gettimeofday(),
clock_gettime(CLOCK_MONOTONIC), and clock_gettime(CLOCK_REALTIME)
all take about 40ns.  Of course gettimeofday() only has 1us resolution,
but the other two have perhaps 10ns resolution (I get no duplicate
readings in a tight loop).  Other documented clockids include   CLOCK_REALTIME_COARSE: about 10ns to read, but only 1ms
resolution  CLOCK_MONOTONIC_COARSE: about 12ns to read, but only 1ms resolution   CLOCK_MONOTONIC_RAW: full resolution
butvery slow, ~145ns to read 
So CLOCK_MONOTONIC seems to be the thing to use here.  It won't buy
us anything speed-wise but the extra resolution will be nice.
However, we need to do more research to see if this holds true on
other popular distros.

On macOS (10.12.2, 2.7GHz x86_64), clock_gettime(CLOCK_REALTIME)
is actually a shade faster than gettimeofday: 40ns versus 46ns.
But it's only giving 1us resolution, no better than gettimeofday.
CLOCK_MONOTONIC is also 1us and it takes 75ns to read.  But there's
a CLOCK_MONOTONIC_RAW that takes 44ns to read and seems to offer
full precision -- no duplicate readings in a tight loop.  There's
also CLOCK_MONOTONIC_RAW_APPROX which can be read in 23ns but
the resolution is only around half an ms.

I also tried FreeBSD 11.0 on another Mac (2.3GHz x86_64),
and found that gettimeofday as well as basically all their
clock_gettime variants run in 27 to 28 ns; and clock_gettime
reliably delivers full precision, except for CLOCK_SECOND which
is intentionally truncated to 1s precision.  So there would be
no need to work with anything but CLOCK_MONOTONIC here.

However, it seems that these impressive results date back only to
June 2012, cf
https://github.com/freebsd/freebsd/commit/13a9f42818f6b89a72b3e40923be809b490400d8
and at least as of that commit, only x86 and x86_64 had the fast
clock_gettime code.  Older FreeBSD, or FreeBSD on another architecture,
is likely to be a lot worse.  But I lack an installation to try.

I also tried OpenBSD 6.0 on that same Mac, and got pretty horrid
results: gettimeofday, CLOCK_REALTIME, and CLOCK_MONOTONIC all
take about 613ns to read.  Ouch.  And so does time(NULL); double
ouch.  Evidently there's no optimization on this platform and
what we're seeing is the minimum cost for a kernel call.  Still,
we do get better precision from clock_gettime than gettimeofday,
so we might as well switch.

So it seems like the configure support we'd need is to detect
whether clock_gettime is available (note on Linux there's also
a library requirement, -lrt), and we would also need a way to
provide a platform-specific choice of clockid; we at least need
enough smarts to use CLOCK_MONOTONIC_RAW on macOS.
        regards, tom lane