[ntp:questions] Best practices for leaf nodes

Maarten Wiltink maarten at kittensandcats.net
Wed Nov 28 15:54:58 UTC 2007


"Steve Kostecke" <kostecke at ntp.org> wrote in message
news:slrnfkqvo6.orf.kostecke at stasis.kostecke.net...
> On 2007-11-28, Dag-Erling Smørgrav <des at des.no> wrote:
[...]
>> and how to behave if no answer is received.
>
> Don't start hammering on the server.

This bears fleshing out, since it's not obvious - although it's
one of those ideas that are obvious in hindsight.

If there is no answer, don't start knocking faster. Start knocking
_slower_. The technical term is 'exponential backoff', because the
usual implementation is to double the interval every time.

The idea is that missing a beat isn't immediately disastrous, but
hammering a momentarily or permanently absent server is. Especially
if you are an over-the-counter DSL router and you have half a
million brothers and sisters.

If the server is away for only a moment, you should NOT retry once
a second until it's back. You can live with a lost sample, but if
there are enough clients with the same behaviour, that server may
never get its backlog cleared because it's orders of magnitude
above regular traffic levels.

If the server is gone forever, you _definitely_ don't want to keep
flooding its network for all eternity.

Groetjes,
Maarten Wiltink





More information about the questions mailing list