On Fri, Nov 04, 2016 at 11:52:47AM -0700, Denny Page wrote:
> The offset is being taken at 15 second intervals via a script that calls 
> "chronyc -c sourcestats” and feeds the data to rrd. My polling was not quite 
> as aggressive, at 2 for the device shown in the graphs I sent.
> 
> I wouldn’t think I have much in the way of asymmetry, but anything’s 
> possible. Both the monitoring system and the and the NTP servers under test 
> are connected to the same switch. Here’s what my sources/sourcestats looked 
> like:

In my testing on LAN the asymmetry in jitter doesn't seem to come from
network, but rather the kernel, drivers, or maybe HW.

Anyway, your results look much better than what I normally see, even
between machines connected directly without switch. Without HW
timestamping the std dev I see is at least few microseconds.

What HW and kernel do you use?

> 210 Number of sources = 7
> MS Name/IP address         Stratum Poll Reach LastRx Last sample              
>  
> ===============================================================================
> ^+ 192.168.230.240               1   2   377     3   -218ns[ -240ns] +/-   
> 23us
> ^x 192.168.230.241               1   2   377     0    -40us[  -40us] +/- 
> 2930us
> ^x 192.168.230.242               1   2   377     0    -40us[  -40us] +/- 
> 2930us
> ^* 192.168.230.244               1   2   377     0   -146ns[ -168ns] +/-   
> 23us
> =x 192.168.230.1                 2   3   377     1  -7027ns[-7049ns] +/- 
> 1530us
> =x 192.168.230.2                 2   2   377     4  +4821ns[+4799ns] +/-   
> 86us
> =x 192.168.225.10                2   2   377     5    -24us[  -24us] +/-  
> 206us

Hm, why are so many sources marked as falsetickers? Their measurements
look good.

> Back on the switch, if I bring the polling down to 1 then things tighten up a 
> bit more. But still nowhere near the 151ns standard deviation that you are 
> getting.

FWIW, from those 150 ns a large part seems to be instability of the
server's clock, which is synchronised to a GPS with PPS over serial
port. If I disable the PPS refclock, the std dev observed on the
client drops to about 50 ns. I have another server, which has clock
stable to maybe 20 ns (PPS over GPIO), but it doesn't have a NIC with
HW timestamping, so it's useless for these tests.

-- 
Miroslav Lichvar

-- 
To unsubscribe email chrony-dev-requ...@chrony.tuxfamily.org with "unsubscribe" 
in the subject.
For help email chrony-dev-requ...@chrony.tuxfamily.org with "help" in the 
subject.
Trouble?  Email listmas...@chrony.tuxfamily.org.

Reply via email to