On Tue, Mar 03, 2009 at 01:32:14AM -0800, Ask Bjjjrn Hansen wrote:
> Sadly it seems like a HUGE amount of users are fetching time right at  
> the top of the hour.
> 
>       http://tmp.askask.com/2009/03/84.45.68.23-daily.png

RRDtool isn't really very good for this sort of thing; you really want
to do some sort of time series analysis to get useful information out.

That'd allow you to correct for things like the gentle upward drift
over time as the service accumulates users, the intra-annual changes as
people go on holiday and hence computers are on for more or less time,
weekly and daily effects and so on.  I've not been logging this data for
a while now, but remember the hourly spikes being big.  Also there were
very reliable other effects, 14minutes past for some reason sticks out
in my memory.  Some people like to query at specific seconds past the
hour as well, which was slightly strange but would suggest that they are
maintaining their clock reasonably well.

I seem to have chucked all my data away, but I'll get my logging going
and see if I can put something like this together again.  I think I
found the best was to dump out the binary stream of the incoming NTP
"response" packets.  I can't remember how I identified them as being
response packets now, but it shouldn't be too hard.

> It's almost tempting to tweak the nameserver so the TTL in the minute  
> just around the top of the hour is 3-4 hours instead of the usual 20  
> minutes.

Sounds like an interesting experiment!  If you've got a host that could
take a bit of load, you could try redirecting some subset of the traffic
towards it at specific times and build up a more complete profile of the
traffic over time.

-- 
  Sam  http://samason.me.uk/
_______________________________________________
timekeepers mailing list
[email protected]
https://fortytwo.ch/mailman/cgi-bin/listinfo/timekeepers

Reply via email to