hwclock accuracy

[Jerry Vonau]

On Thu, 2011-01-20 at 02:15 -0800, Hal Murray wrote:
> > What is the accuracy of the hwclock on XOs? Can we assume that it will keep
> > good time over an XO's five-year lifespan?
> 
> The ballpark for the hardware is 1 second per day.
> 
> > We have some XO-1.5s that have their hwclocks off by 10-20 minutes. Would
> > this likely be due to a fault at the factory (e.g. not setting the time
> > before shipping them), or did they skew over time? 
> 
> How long ago did you last set the clock?
> 
> Timekeeping is more complicated than that.
> 
> /etc/init.d/halt calls /sbin/hwclock --systohc
> So each time you shut down cleanly, your hwclock gets updated with the 
> current system time.  That's probably a good idea if you are running ntpd 
> like most Linux boxes.
> 

The main problem is /etc/adjtime doesn't keep it's data used to track
whether the rtc is in localtime or utc time, the readonly filesystem
layout resets this file on reboot. 

https://dev.laptop.org.au/issues/299
http://dev.laptop.org/ticket/10605


> However, if you aren't running ntpd (or somehow keeping the system clock 
> reasonably sane), you are swapping the drift on the hwclock for the drift of 
> the system clock.
> 

With my observations I'd used the rtc as the source, as the system is
drifting over a 1 min per hour.

> Sometime in the past few years, the timekeeping corner of the kernel was 
> cleaned up.  Unfortunately, they introduced a bug in the TSC clock 
> calibration code.
> 
> The main problem for most Linux users is that it doesn't get the same answer 
> each time it boots.  It's close, but not very good if you are interested in 
> timekeeping.
> 
> If it were consistent, ntpd could easily correct for the error.  That's what 
> /var/lib/ntp/drift is for.  (Small variations are expected due to 
> temperature.)
> 

The would be great to use but a ntp client is not installed by default. 

> The secondary problem is that it's likely to be off by 10s of seconds per 
> day.  I'll dig out some numbers if anybody wants more details.
> 
> I don't have good data on how well clocks work when power-saving is enabled.
> 

You might be on to something there, on one XO-1.5 with uptime of 12
hours (inactive, I was sleeping), the system clock is now out over 3
hours in that time frame while the rtc is accurate. Well it's not out 3
hours... ;)

> -------------
> 
> If you actually want to know how accurate your hwclock is:
>   comment out the line in /etc/init.d/halt,
>   reboot,
>   check the time,
>   wait a week or 10 days,
>   check again.
> "ntpdate -d <handy-server>" may be the simplest way to check the time.  Grab 
> ntpdate from a handy system.
> Or run "xclock -analog -update 1" on a system running ntpd and use that to 
> calibrate your wrist watch and ...

Going down that route chasing this, 

Jerry