[ntp:questions] Re: ntpd server on Linux/Debian with SPARC Sun Blade 100

David Woolley david at djwhome.demon.co.uk
Sat Oct 21 19:39:53 UTC 2006


In article <4539f3de$0$324$e4fe514c at news.xs4all.nl>,
Rob van der Putten <rob at sput.nl> wrote:
> Olivier Bornet wrote:

> > Thanks for the link. If setting the HZ don't resolve my problem, I will
> > try to adjust the drift.

> AFAIK NTP can't handle asymmetrical link speeds. So if you have a busy 

NTP cannot handle asymmetrical link delays, asymmetric speeds don't
necessarily produce these because they tend to be used in contexts where
the downlink traffic is much higher than the uplink traffic.  In fact they
are likely to produce errors in the opposite direction to what you might
expected if you judged by link speed alone.  Also, it is variability
in the delay difference, not the mean value, that matters. (Actually,
I'm not sure how any protocol could fully account for assymmetric delays.)

In any case, this issue doesn't seem to be relevant in this case, as I
believe he had unidirectional stepping, and if you have unidirectional
stepping, it normally means that you would have to slew faster
than 500ppm to catch up.

> I solved the problem with the '-x' option;

The option that is intended to deal with this case is the huff and puff
tinker option.  I believe, that in recent versions, -x sets the step
limit high, but not infinite, and that high step limits cause the
kernel PLL to be disabled.

Given that almost all consumer links and probably many service provide
links are assymmetrically loaded, I don't think it a good idea to
advise -x in all such cases.  A lot of people are continually trying
to advise people against using it when they think they have a problem
with databases.

> At boot the clock is synced with ntpdate, so the above doesn't cause any 
>  problems.

ntpdate is deprecated.  More interestingly, it seems that certain optimisations
to achieve fast initial convergence only work if the initial time is outside
the step limit!




More information about the questions mailing list