[nycbug-talk] dealing with drift
Mon Feb 28 13:04:13 EST 2005
* George Georgalis <george at galis.org> [2005-02-28 12:33]:
> On Mon, Feb 28, 2005 at 12:17:52PM -0500, Isaac Levy wrote:
> >Hi All,
> >I was just wanting to post on some of the post-meeting conversations
> >I've repeatedly over-heard about managing time-drift across clusters of
> >machines- (you know who you people are...)
> >How do folks like to do this? Special tools beyond rockin' timed?
> >Anyone know any good urls on the topic (either practical or acidemic)?
> Am I missing something here? Isn't the obvious answer, for a cluster, to
> configure the headnode (whether is connected to internet and synced or
> not) to answer ntp queries from nodes, and configure the nodes ntp time
> server to use their headnode? /usr/ports/net/ntp
I would add diversity here. Having only one time master is a SPoF.
Diverse routing is helpful as well.
Just finished a similar exercise, though just for a collection
(not a cluster) of machines.
NTP on all BSDs but the OpenBSD box. OpenNTP on that.
I do miss the ntpdc utility on OpenBSD. Unless I'm missing something,
it's not easy to verify that you are, in fact, synched with another
Also- one other nit: OpenNTP comes preconfigured to synch to
a pool of servers from ntp.org. Why, I don't know. Simplicity
maybe- but how am I supposed to verify *those* servers?
More information about the talk