[ntp:questions] ntpdate works, but ntpd doesn't (reach = 0)

Unruh unruh-spam at physics.ubc.ca
Mon Feb 16 23:50:24 UTC 2009


mayer at ntp.org (Danny Mayer) writes:

>Martin Burnicki wrote:
>> Unruh wrote:
>> 
>>> Harlan Stenn <stenn at ntp.org> writes:
>>>
>>>>>>> In article <ST4ll.11327$Db2.8064 at edtnps83>, Unruh
>>>>>>> <unruh-spam at physics.ubc.ca> writes:
>>>> Unruh> Why is -g inappropriate from a resttart. If the clock is still very
>>>> Unruh> close to right, it will make no difference. If for some reason on
>>>> Unruh> restart the clock is out, then -g is appropriate. Ie, I cannot see
>>>> Unruh> why it would hinder anything if it were run on restart.
>>>> I'm getting tired of repeating myself.
>>> Repeating yourself is different from explaining yourself. You feel that
>>> ntpd should have "not -g" as the default. A number of us suggest that -g
>>> is a more appropriate default (ie so that to stop ntp from doing a time
>>> reset once would then require some flag).
>>> You say I can rewrite ntpd, or set up a special ntp.conf file or...
>>> But I still do not understand why -g should not be the default.
>> 
>> Again, I absolutely agree to the above.
>> 
>> I don't see any restrictions in cases where ntpd is re-(!)started with -g
>> even though this case is obviously which prevents -g from becoming the
>> default.
>> 
>> I'd just like to know a condition where -g is Bad in the particular case of
>> restarting ntpd.

>There are a couple of situations where a step rather than a slew would
>be really bad on a restart of ntp (as opposed to a reboot of the O/S).
>One of these cases is with databases timestamping a transaction where a
>step backward would result in a transaction taking place before some
>other transaction. These are real financial issues that are involved.
>Note that on a regular startup, ntpd should be started before the
>database so that the issue is not a problem. A restart *after* things
>are running becomes a real big problem. Another instance is in a lab
>network where time is extremely important. In addition I hear in TICTOC
>that there may be problems with different segments that they are trying
>to deal with.

If ntp is out by 128ms it steps no matter what. This is a situation whee on
a restart it is out by far more than that for the -g to become operative.
Is it better for the time to be out by many many seconds or to be stepped?
Is it better for ntp to die and refuse to keep running or to step?

Note that using an ultra fast slew rate ( 2000PPM or 10000 PPM say) to get
rid of such gross errors might well be better than a step for your reasons.

The cases where the system must absolutely not be stepped are relatively
rare I would expect so as to make -g the default ( with of course the
possibility of switching off the default in your special cases where a gory
death of ntp is preferable to a step) seem the more generally applicable
instance.

In a lab instance where time is extremely important and the time is out by
seconds/minutes/hours on a restart, would you rather have the step or ntp
dead doing nothing to discipline that time?

>Danny




More information about the questions mailing list