[time-nuts] Tracking NTP displacement and correlation betweentwo clients.
bownes
bownes at gmail.com
Fri Oct 5 22:57:03 UTC 2012
Comments inline.
On Oct 5, 2012, at 18:26, Hal Murray <hmurray at megapathdsl.net> wrote:
>
> bownes at gmail.com said:
>> The problem is that they start in sync and over the course of a day drift
>> that far apart despite having NTP running. We're not sure why NTP isn't
>> correcting it along the way. Though at this point, we are looking at a
>> firmware bug.
>
> I wouldn't think of it as two systems drifting apart, but rather at least one
> system with a broken clock.
>
Correct.
> Is it only one system that is broken?
>
Sort of. There are several systems consisting of a matched pair of nodes. In each case, one of the two wanders out into the weeds. But not every pair has one that goes south.
In this case, four systems, 8 nodes, all identical hw (sequential sn's even), identical iLOM/DRAC, same software the entire length of the stack.
Installing the latest firmware patch appears to have solved the problem. I'll know next week.
> How many systems do you have running the same firmware?
<redacted>
> Normally, if ntpd is off by more than 128 ms, it will step the clock. That
> puts a line in the log file. So it's more than a bit strange that the clocks
> get off by many seconds.
>
My thinking exactly. But it wasn't. I was hoping to use some tools to watch it drift off.
> I'd double check that ntpd really is still running.
It is.
> Are your drift-apart systems using only your 2 local stratum-2 servers? If
> so, that may be the problem. If those servers don't agree, which one do you
> believe? (There is endless discussion in the NTP community about how many
> servers you need. 3 lets you out-vote 1 bad guy. 4 lets you out-vote a bad
> guy if one of them is down. ...)
>
Two NTP servers agree. They even agree with my S1 at home. :)
Thanks for all the help folks. It looks like it was a firmware bug, even if I can't explain how the firmware was causing the NTP clock to be off.
More information about the time-nuts
mailing list