On 2015-08-29 15:23, Mike Cook wrote:
Le 29 août 2015 à 17:41, Charles Swiger <cswi...@mac.com> a écrit :
On Aug 29, 2015, at 1:03 AM, Mike Cook <michael.c...@sfr.fr> wrote:
For me this would not be considered acceptable. I would NEVER have ntp servers
in virtual machines.
+1 to this point.
Your maxpoll default of 10(1024s) for the pool servers is too high. Try
dropping it to 6 (64s) or 7(128s).
-1 to this.
experience tells me otherwise.
Here’s a test:
2 clients, same hardware, same LAN segment, same version NTP, same ntp config,
with exception of their own GPS PPS ref clock.
The 2 clients reference 4 identical servers with the same versions of ntpd, all
4 have GPS PPS ref clocks in addition to their PPS refs and preferred GPS
sync’d server.
All servers are on the same LAN segment as the clients. These servers are
configured with maxpoll 10 on bb2, and 6 on bb3.
ntpq -pn data
bb2
Sat Aug 29 20:59:19 UTC 2015
remote refid st t when poll reach delay offset jitter
==============================================================================
o127.127.22.0 .Neo6. 0 l 13 16 377 0.000 0.002 0.002
*192.168.1.4 .PPS1. 1 u 9 16 377 0.712 0.119 0.015
-192.168.1.15 .PA6H. 1 u 583 1024 377 1.163 0.358 0.056
-192.168.1.16 .MAX7. 1 u 597 1024 377 1.260 0.369 189.435
+192.168.1.17 .ResT. 1 u 493 1024 377 1.261 0.312 0.040
+192.168.1.18 .Neo8. 1 u 600 1024 377 1.136 0.250 0.088
bb3
Sat Aug 29 20:57:18 UTC 2015
remote refid st t when poll reach delay offset jitter
==============================================================================
o127.127.22.0 .NS-T. 0 l 12 16 377 0.000 0.003 0.002
*192.168.1.4 .PPS1. 1 u 10 16 377 0.591 0.064 0.029
+192.168.1.15 .PA6H. 1 u 10 16 377 0.437 0.021 0.037
-192.168.1.16 .MAX7. 1 u 9 16 377 0.583 0.074 0.029
+192.168.1.17 .ResT. 1 u 10 16 377 0.448 0.017 0.030
+192.168.1.18 .Neo8. 1 u 10 16 377 0.593 0.055 0.022
You can see that the maxpoll 10 servers seen from bb2 have delays at >2 times
that of the same servers seen from bb3.
Reported offset on bb3 is up to 5 times that seen from bb3.
Reported jitter (ntpd’s error bars) is in 3 cases of the same order, though
greater, but in one is completely anomalous on bb2 .
I have not made a complete analysis of this and give the following as a
probable cause.
Long intervals between client server exchanges allow arp caches to be cleared
in both clients, servers, routers causing arp resolution (who has IP, I have
IP, here’s my MAC).
This extra path delay will possibly be asynchronous and offset and jitter will
be increased .
Here are the clients arp caches to show what I mean:
bb2
mike@bb2:~$ arp -a
livebox-router (192.168.1.1) at 3c:81:d8:db:4e:b4 [ether] on eth0
muon.stratum1.d2g.com (192.168.1.4) at 00:00:24:c6:20:60 [ether] on eth0
electron.home (192.168.1.13) at 34:15:9e:01:e5:9c [ether] on eth0
cubieez.stratum1.d2g.com (192.168.1.124) at 02:cd:07:c1:82:5f [ether] on eth0
mike@bb2:~$ exit
bb3
mike@bb3:~$ arp -a
cubieez.stratum1.d2g.com (192.168.1.124) at 02:cd:07:c1:82:5f [ether] on eth0
livebox-router (192.168.1.1) at 3c:81:d8:db:4e:b4 [ether] on eth0
raspb3.stratum1.d2g.com (192.168.1.17) at b8:27:eb:71:05:50 [ether] on eth0
raspb2.stratum1.d2g.com (192.168.1.16) at b8:27:eb:2b:ab:90 [ether] on eth0
raspb4.stratum1.d2g.com (192.168.1.18) at b8:27:eb:fe:15:fa [ether] on eth0
electron.home (192.168.1.13) at 34:15:9e:01:e5:9c [ether] on eth0
muon.stratum1.d2g.com (192.168.1.4) at 00:00:24:c6:20:60 [ether] on eth0
raspb1.stratum1.d2g.com (192.168.1.15) at b8:27:eb:7e:0b:f2 [ether] on eth0
So in certain application environments using long poll intervals IS detrimental.
I am going to run the same configs overnight with the net loaded to see what
difference that makes.
Unless your hardware is broken, it shouldn't need to ask what the time is every
minute just to manage decent timekeeping accuracy. Admittedly, running on a VM
often resembles running on real hardware with a broken clock, but to me that in
turn means one should be running ntpd in the hypervisor so that it is managing
a real HW clock and let the VMs inherit time from the hypervisor / Dom0 / etc.
Reports indicate min/maxpoll 4 improve results on LAN segments,
minpoll 6 improves remote network server offsets, maxpoll 10
improves system clock frequency and offset estimation using only
remote network servers.
I have seen the latter pull Windows systems within 10-100us of UTC,
much better than the few to tens of ms expected on this platform.
--
Take care. Thanks, Brian Inglis
_______________________________________________
questions mailing list
questions@lists.ntp.org
http://lists.ntp.org/listinfo/questions