[ntp:questions] ntpd sensitivity to ordering of servers in ntp.conf?
wsdl at osengr.org
Thu Feb 25 10:49:48 UTC 2016
Bored? Need something to do? You could try helping me with this puzzle
regarding ntpd. This is a rather long message, so if you're otherwise
busy I won't be offended if you skip it.
-- Executive Summary --
The delay and offset values measured by ntpd appear to change slightly
when the order of two servers is swapped in ntp.conf.
-- Introduction --
I'm working to verify the performance of two NTP servers I've built, and
have run into some curious behavior with ntpd on a current release of
Fedora. This is regarding a few 10's of microseconds at most and is
certainly not a serious problem.
-- The NTP Servers --
The two NTP servers being tested are embedded systems (Arduino-based)
and each has an independent GPS reference clock. One uses an HP55300A
while the other uses a Trimble timing-specific GPS module (ICM SMT 360).
There are hardstamps on received packets and (oscilloscope-calibrated)
drivestamps for transmit.
I suspect these servers are accurate to something on the order of a
microsecond or so and am trying to get some verification of this at the
network level. This estimate of accuracy only applies when the NIC can
get onto the wire w/o delay. When there is other network traffic I can
routinely measure variable delays on the order of a hundred microseconds
I realize this sort of accuracy is not perhaps all that useful given the
variable delays that occur in a typical loaded network. On the other
hand, by placing one of these servers on an unused Ethernet port it
could be possible to get reference-clock accuracy without a
hardware-connected reference clock. And well, like climbing a
mountain...just because it's there.
-- Test Setup --
The NTP client is a 32-bit Pentium-class PC (an old one with IDE disks)
running Fedora 23. I believe it has kernel discipline and probably the
nano kernel but am not positive about that. The kernel build is
The PC and two NTP servers are connected to a 100 base-T network hub.
The hub is not connected to any other networks.
ntp.conf is specifies both servers with minpoll 4/maxpoll 4. Peer and
loop statistics are enabled.
-- loopstats --
Both offset and frequency offset are very stable in loopstats (as long
as the CPU temperature stays constant). Over a 6-hour test run, the mean
value of offset was -0.63us and the standard deviation was 5.5us.
Frequency jitter hovered around 5ppb with occasional jumps up to about
10ppb. As I understand it, this is about as good as it gets, even with a
good local ref clock.
-- The Puzzle --
The curious part is in peerstats. Server A shows a delay averaging
around 85us and server B is running around 60us. Offsets are also
different, but by about half as much. Offset on A is about +5us and -5us
I was hoping to see closer to identical delays and offsets on the two
servers. I'm fairly certain they should be closer than 10us in offset. I
tried the following experiments trying to track this down:
1) Swap network hub ports between servers A and B. No change.
2) Cycle power on network hub after (1). No change.
3) Re-check transmit timestamp timing with oscilloscope. Looks good.
4) Change the order of the two servers in ntp.conf. Yikes, that's it!
By just changing the order of servers in ntp.conf the delay and offset
values in peerstats are swapped. Now it is A with 60us delay and B has
85us. Similarly, A's offset is not -5us and B is showing +5us.
It appears there is something in ntpd where measurements on server A in
ntp.conf come out slightly different depending its ordering in ntp.conf.
So yes, this is really mouse-nuts. It is however setting a bound on how
accurate I can conclude the NTP servers are.
Any ideas what might be causing this?
(I have some data plots and could e-mail them if desired)
More information about the questions