[ntp:questions] Re: NTP seems unsuitable for this application... what do you think?
David L. Mills
mills at udel.edu
Fri Dec 3 04:27:51 UTC 2004
The Solaris kernel works fine as long as the step threshold is not
tinkered above 0.5 s. The problem occurs when the kernel is NOT used
because somebody inists on slew corrections well above 0.5 s, in which
case the Solaris adjtime() syscall operation is incompatible with ntpd
and results in obnoxious overcorrections and generally unstable behavior.
Paul Croome wrote:
>>If you really care what time it is, leave one of the GPS systems powered
>>up to act as a server. When synched to GPS it should know to within a
>>few hundred microseconds what time it is and will share its knowledge
>>with anyone who boots up and asks. If you give an NTP client a stable
>>and accurate source of time, it will synch up and stay synchronized as
>>long as it's powered on, IFF you give it enough time to do so. It can
>>take twelve to twentyfour hours to achieve tight synchronization from a
>>cold start! By "tight synchronization" I mean within five milliseconds
>>or less; I've seen Solaris/SPARC systems synchronize with offsets in the
>>low microsecond range.
> IIRC, Dave Mills has remarked that the kernel PLL/FLL code in some versions
> of Solaris is broken, so that it does not converge optimally during startup
> (it's OK under steady-state conditions, though). My observations confirm that.
> The bottom line: Don't expect good NTP performance during the first few hours
> after coldstart. And BTW, thermal drift as the computers and their crystals
> warm up is another factor that means that initial performance will be so-so.
More information about the questions