[ntp:questions] Win7: ntpd adjusting time backwards

Jeroen Mostert jmostert at xs4all.nl
Sun Dec 9 17:59:10 UTC 2012


On 2012-12-09 18:18, David Taylor wrote:
> On 09/12/2012 16:26, Jeroen Mostert wrote:
> []
>> As long as Microsoft still doesn't make it possible to upgrade the
>> kernel without upgrading the shell, that's not going to happen, because
>> the Windows 8 UI is a serious productivity killer for me.
>>
>> The broader context is that I'm working on an NTP setup for a "real"
>> network, where the only demand is that the servers stay within each
>> other's times to some reasonable degree (I'll take anything <= 10 ms),
>> with only very lax demands on synchronizing to absolute time. I've
>> already discovered that this is actually fairly easy as long as all
>> machines sync to a single local server with some frequency (setting
>> minpoll 4 maxpoll 4 is no problem on a LAN, of course, and keeps even
>> the most wayward machines in line). I'm still happy I ran into problems
>> locally, because upgrading to 4.2.7 significantly reduces jitter even in
>> this setup (as reported by your site as well), so I'll be sure to
>> slipstream that in.
>>
>> Of course, upgrading all machines to Windows 8/Windows Server 2012 just
>> to get better time synchronization isn't going to happen. :-)
>>
>> For now the local server is a Windows Server 2008 machine, but I've
>> already petitioned for a dedicated Linux machine (FreeBSD would be too
>> "out there" for IT, I'm afraid...) If possible, we can later upgrade
>> that to something with a GPS, but it's probably not necessary at the
>> moment.
> []
>> If 30-40 ms is realistic for syncing to internet time, I'll take it.
>> Though it's slightly disappointing, but I guess the limitation is in
>> Windows more than it is in NTP.
>
> Win-8 is almost the same as Win-7, just add a start-menu add-on such as
> ClassicShell. Quite tolerable, then (although DOS shell, run as Administrator
> takes a little more effort!).
>
Yeah, no. I categorically refuse to "upgrade" my OS (at a non-zero cost) and 
then "install some extra stuff" just to make it *acceptable* again. That's just 
insane. I guess if you need the new features in Windows 8, it might be worth 
considering the time and effort, but at present I know of no such features. The 
increased clock accuracy isn't one.

I'm waiting for Windows NT 6.3, or 7.0, whatever the marketing name ends up 
being. If that still has Metro, I'll see how I'll compensate. But I'm hoping it 
won't.

> What I would agree for your situation is to have a local server to which you
> tight-lock (low min/maxpoll) the rest. But as a result of my own recent tests, I
> would make that local server a Raspberry Pi or similar (it's very low cost)
> running Linux and synced to the Internet, and if it suits, make that a stratum-1
> server with a GPS/PPS for a score of dollar/pounds more.
>
I work for a Windows shop with too much money and too little time. It's going to 
be a spare server we had lying around; suggesting folks get enthusiastic with 
hardware that's not from Dell isn't going to fly. :-) So the Raspberry Pi is 
out, and upgrading to a stratum 1 only if there's a demonstrable need.

This is all basically a spare time effort from me, to enable more accurate 
performance analysis of cross-server events. It's not exactly mission critical, 
but nice to have. There was a significant period of time where we had no time 
syncing whatsoever, and we only started to run into problems when one machine 
was 19 seconds off from actual time. I then configured w32time on all machines, 
but w32time is not particularly robust. Microsoft itself admits it's just there 
to make sure Kerberos doesn't completely fail, and as long as it keeps time 
within 5 minutes from a domain controller they don't care. I'm now at a point 
where I think machines deviating from each other for more than 10 ms is 
annoying, so that won't do anymore.

> No need to upgrade all PCs, just one server. PCs Hydra and Narvik here are on a
> LAN connection.
>
> http://www.satsignal.eu/mrtg/performance_ntp.php
>
> and PCs Molde, Puffin, Torvik and Ystad are Wi-Fi connected (which doesn't help
> timekeeping). Note the difference between XP and Windows-7.
>
Statistics like those would be more than acceptable for my purposes. Your worst 
stats would be good enough for me.

> Your suggested arrangement of a Linux time server (which can be a very simple,
> low-powered PC) and tight coupling of the rest should give you excellent results.

It's actually going to be far heavier than is necessary, but I'm confident it'll 
get the job done.

-- 
J.



More information about the questions mailing list