Re: [chrony-dev] PPSAPI: kernel consumer

[Håkan Johansson <f96hajo@xxxxxxxxxxx>]

On Mon, 8 Feb 2010, Miroslav Lichvar wrote:

> On Mon, Feb 08, 2010 at 12:35:59PM +0100, Håkan Johansson wrote:
> > I've recently (for slightly other reasons) measured the in-kernel
> > time consumed by doing getnstimeofday() (several in a row), and for
> > most machines (2 GHz+) it took a whopping ~ 500 ns.  I.e. one cannot
> > know to better than 0.5 us what the time was, if one e.g. try to
> > sandwich something else with two such calls.  The best machine was
> > actually a slower one (~1.05 GHz) that had about 60 ns between
> > return values for two back-to-back getnstimeofday().  I suppose we
> > are down to PCI bus latencies and so on.
>
> Which clocksource was used? The best results should provide tsc,
> usually few tens of nanoseconds. From some tests I did a while ago it
> seems to worsen with more CPU cores.

The measurements:

AMD 1050 MHz,       back-to-back getnstimeofday takes 62 ns
Pentium M 1600 Mhz, back-to-back getnstimeofday takes 560 ns
Athlon 2000 MHz,    back-to-back getnstimeofday takes 560 ns

All systems single-core.  Where I had available, I changed the timesource 
between acpi_pm and tsc with 
/sys/devices/system/clocksource/clocksource0/current_clocksource.
No major effect.

It was however crucial in these things to do one or a few calls to 
getnstimeofday before the two ones to do the timing to heat the processor 
caches.

Cheers,
Håkan