On Sat, May 12, 2012 at 04:00:32PM -0700, Ask Bjørn Hansen wrote:
> 
> On May 12, 2012, at 6:23, Kurt Roeckx wrote:
> 
> >> Can you elaborate? The monitoring is supposed to take out crappy
> >> servers, so that you shouldn't be able to see them.
> > 
> > The monitoring isn't good enough.  I think you need to be off by
> > more than 100 ms before it reacts.  And I see servers going from
> > +100 to -100 and back all the time.
> 
> I'm surprised you see this often.  Can you send me some examples?  (Maybe 
> off-list).

All the examples I had currently don't seem to be in the pool
anymore, and don't even reply anymore for a while.

I think a lot of the problematic servers run openntpd.

> 100ms is fine for virtually all users of the pool.   If using ntpd, it will 
> likely pick a better clock.

ntpd ussually doesn't want to lock on anything if it can't find
servers that agree on the time.

> My concern with making the limit smaller is that any problems with the 
> monitoring system are more likely to affect things then; and from when I've 
> looked at the numbers usually if a server goes bad, it gets worse than 100ms 
> quickly anyway.

I can understand that.  For some of the server I monitor I see the
delay go up more than 100 ms during some parts of the day, and
with that also the error.

> >> The iburst feature is meant to speed this up (down to about 10s);
> >> somehow, this fails to work in this bug report.
> > 
> > Yes, we have iburst in the config file in Debian, and I've asked
> > this list if they think that that's a good idea or not, since
> > it might increase traffic to the servers, but never got a reply to
> > it.
> 
> I'd say it's a bad idea to have as a distributed default.  If any individual 
> users need it, fine - but distributed as the default is not a good idea.

Why?  I think most people want that.


Kurt

_______________________________________________
pool mailing list
[email protected]
http://lists.ntp.org/listinfo/pool

Reply via email to