[ntp:questions] Re: peer trouble ? reach going down

David Schwartz davids at webmaster.com
Thu Aug 7 08:51:55 UTC 2003


"Cyril" <cyril_mtl at hotmail.com> wrote in message
news:ad5ec9c.0308061606.61b07f6b at posting.google.com...

> Everything work well, but the 'reach' for the peer associations never
> stay a 377.

    Probably because packet loss is not zero.

> I snoop NTP communication I find some strange delay.
> When a peer ask a server, the answer is very fast :
>    fwdrp-dev -> chrivh40.cch.com NTP  client (Wed Aug  6 19:53:47
> 2003)
> chrivh40.cch.com -> fwdrp-dev    NTP  server (Wed Aug  6 19:53:47
> 2003)
>    fwdrp-dev -> SUE.CC.UREGINA.CA NTP  client (Wed Aug  6 19:53:58
> 2003)
> SUE.CC.UREGINA.CA -> fwdrp-dev    NTP  server (Wed Aug  6 19:53:58
> 2003)

    That's because the peer is asking a question.

> but message between peer are not the same :
>        fwdrp -> gestor01     NTP  symmetric active (Wed Aug  6
> 19:45:08 2003)
>     gestor01 -> fwdrp        NTP  symmetric passive (Wed Aug  6
> 19:46:07 2003)
>     gestor01 -> fwdrp        NTP  symmetric passive (Wed Aug  6
> 19:47:15 2003)
>        fwdrp -> gestor01     NTP  symmetric active (Wed Aug  6
> 19:47:16 2003)
>     gestor01 -> fwdrp        NTP  symmetric passive (Wed Aug  6
> 19:48:23 2003)
>        fwdrp -> gestor01     NTP  symmetric active (Wed Aug  6
> 19:49:24 2003)
>     gestor01 -> fwdrp        NTP  symmetric passive (Wed Aug  6
> 19:50:31 2003)
>        fwdrp -> gestor01     NTP  symmetric active (Wed Aug  6
> 19:51:32 2003)
>     gestor01 -> fwdrp        NTP  symmetric passive (Wed Aug  6
> 19:52:39 2003)
>        fwdrp -> gestor01     NTP  symmetric active (Wed Aug  6
> 19:53:40 2003)
>     gestor01 -> fwdrp        NTP  symmetric passive (Wed Aug  6
> 19:54:47 2003)
> 30s to 2 mn between messages.

    If every symmetric query got a reply, then every reply would get a
reply, and you'd have an endless blast of traffic.

> with ntpq peer, I see the reach for all my peer at 376 or 367 or 267,
> but all my
> 'public' internet servers get 377 for reach.

    A reachability of '376' or '367' indicates one packet dropped out of
five. A reachability of '267' indicates two packets dropped out of five.
This is certailny tolerable.

> Can you explain to me why peer exchange are slow ? (and sometime
> one-way,
> because they are not the same stratum 2/3 ?)  and my reach not 377 ?

    Because peering relationships do not consist of queries and replies.
Your reachability is probably due to packet loss. By the way, I don't
recommend setting 'maxpoll'. You may force the servers to talk to each other
too often and the newer data will push older data out of the averaging
window. This will mean averaging over a smaller window, which can be less
accurate.

    Suppose you have two clocks and you want to know if they run at the same
speed. So you set them both to 12:00, wait one day, and compare them. This
will likely be very accurate, even if you set 12:00 off by a second or so.
By contrast, if you compare them one hour later, your one second setting
error may swamp the measurement. By limiting the poll frequency and not
letting it climb as high as it might need to, you are forcing a smaller
window. Bad idea.

    If you force the poll interval to 128 seconds, average 8 samples, and
the network jitter is 20 mS, you may your accuracy at 20 ppm. In the real
world, it will probably be worse than this. (20 mS is twenty millionths of 8
* 128 seconds, so a sample window of 8*128 seconds with an accuracy of 20mS
can't get better stability than 20ppm.)

    DS





More information about the questions mailing list