----- On Jun 19, 2017, at 5:10 PM, Andreas Krüger [email protected] wrote:

> Hello, all,
> 
> I'm just reading the advice we offer to ntp client vendors,
> http://www.pool.ntp.org/de/vendors.html#basic-guidelines ,
> section "Implementation specifics":
> 
>> Do have your devices query the NTP servers at random times of
>> the day. For example every 43200 seconds since boot is good, at
>> midnight every day is bad.
> 
> I completely agree that "at (some hard-wired time) every day" is
> bad for any vendor that puts out more than a few devices.
> 
> But I'm afraid of the "every 43200 seconds since boot" stuff,
> too...
> 
> I imagine a big city with many devices. Or maybe just a big
> factory yard deploying thousands of little devices integrated
> into heaven-knows-what, all with the same ntp client build
> in. That client diligently follows the rule.
> 
> Doesn't matter much whether the devices are from the same vendor
> or not.
> 
> Say, power goes off at some time and comes back up at some
> time. For the sake of example, say comes back at 3:17 UTC.
> 
> Many devices boot - simultaneously.
> 
> Two load spikes each day result, at 3:17 UTC and 15:17 UTC. Those
> two spikes are there to stay, during the life-time of the
> devices. Unless the factory yard finds a way to switch off and
> back on the little devices, one at a time.
> 
> Am I being paranoid here, or is this something we should consider?
> 
> Regards, Andreas


I support an ecosystem of (a really big number of) clients that are all 
intentionally rebooted every night at the same local time. There is indeed a 
huge flood of time checks immediately following them since there is no 
battery-backed "CMOS-type" clock function on the device (thank you 
bean-counters). This is actually kind of nice since it's a really accurate 
approximation of worst case scenario. 

In addition to that first check, the on-board scheduler in the client has some 
pseudo-randomization capability built in such that it can be told to spread its 
sync checks out over seconds or minutes or hours (compared to other adjacent 
nodes). Each time it reboots, it calls home. Each time it calls home, it gets 
seed data to alter behavior for next time. Do you have to plan it effectively? 
Sure you do, which I'll admit many CE vendors probably won't do. It's not the 
end of the world if you plan for the contingency and have some reasonable flood 
protections built into your hosts and networks. 

Dan

-- 
Dan Geist dan(@)polter.net

_______________________________________________
pool mailing list
[email protected]
http://lists.ntp.org/listinfo/pool

Reply via email to