[BBLISA] ntp question

John P. Rouillard rouilj at cs.umb.edu
Tue Mar 17 17:24:32 EDT 2009


In message <49C01132.1030203 at sasser.com>,
Dewey Sasser writes:
>Edward Ned Harvey wrote:
>> Instead, I use cron to schedule ntpdate once per hour on every machine.  It
>> may be less precise (because the clock is allowed to drift, and only gets
>> corrected once per hour) but it's a whole lot more reliable.  How far off
>> does the clock drift over the course of an hour?  Typically, I'd say a
>> second.  Maybe five seconds if you're really pushing it.  I don't care about
>> that for any purposes in my infrastructure. 

I have seen a 5 second drift between a build server and it's nfs
server create a broken build (that took two weeks and a very pissed
customer to figure out what was shipped was seriously broken) because
a file that was supposed to be rebuilt wasn't.

>>I care about the reliability more.

I monitor the ntp servers and alert. Have had ntpdate fail when people
changed firewall settings and killed access to all the ntp servers for
a subgroup. If you aren't monitoring you are just asking for trouble
regardless of using ntpdate or ntpd or ntpclient.

>(Note:  I'm the de facto admin for the network Eric is talking about.)
>
>That's a good point, and I've actually tried that methodology on a host
>of systems.  One problem I've encountered is that many of our "systems"
>are actually VMWare based virtual machines and VMWare seems to have a
>borderline ridiculously bad clock drift.  From what I've read on the
>'net (which, as we all know, is always true) this can somehow manifest
>as either lead as well as lag.

Yup. I have seen 30 seconds-4 minutes/hour. Fortunately make complains
at that level. We have ntpdate running every minute now to try to keep
them under a couple of seconds slew.

>I have observed some of our virtual machines which were synchronized
hourly with ntpdate to drift several minutes apart over the course of
>their hour.  This seems to have made e.g. NFS very unhappy.

Not to mention make and a bunch of other programs that expect to use
timestamps.

--
				-- rouilj
John Rouillard
===========================================================================
My employers don't acknowledge my existence much less my opinions.




More information about the bblisa mailing list