Hello all,
Miroslav, David W: Thank you for your help!
David L: Thank you for creating NTP :)
I experimented with the settings Miroslav suggested.
For the record:
I could not find the SHIFT_PLL in either of both system kernel's
.config, so I just tried with and without "disable kernel".
On the six (on-site) MIPS Routers I use for testing (WRT54GS v1.1), I am
getting an offset in the order of one tenth of a millisecond up to tens
of microseconds; "disable kernel" seems to degrade the accuracy quite a bit.
On the Avila (Intel XScale ARM) router boards in the (remote) testbed,
the offset is in the order of tens of microseconds or well below with
"disable kernel", still in the order of milliseconds without that setting.
That is, Miroslav's
> server 172.17.255.254 iburst minpoll 4 maxpoll 4
> disable kernel
> tinker allan 16
did the trick quite nicely.
Thank you all very much!
Florian
Miroslav Lichvar wrote:
On Mon, Aug 02, 2010 at 10:10:15PM +0200, Florian Sesser wrote:
Do you have advice for me how to improve the accuracy on the nodes to
the order of 1 ms (or below)? I have googled quite a bit, but the amount
of information is overwhelming, and I cannot afford nor attach GPS or
other external clock sources to the nodes. (In fact, I cannot change the
cabling or HW setup in any way.)
We are running ISC ntpd 4.2.4p6. Here's our ntp.conf config
(omitting #comments):
restrict 127.0.0.1
driftfile /tmp/ntp.drift
server 172.17.255.254 iburst
So the ntpd setup is quite simple at the moment.
I am willing to try any patches to the kernel / ntpd, I just don't know
where to go at the moment. Maybe there is much lower hanging fruit, but
being the ntp newbie I am, I cannot see it. If I absolutely need more
ntpd stratum 2 servers in the network, I can maybe lobby to get some,
but any easier option is of course preferred (what about that broadcast
option again? does it only reduce load or also improve accuracy?)
Unless the ethernet network is heavily loaded, getting the nodes
synchronized to 1ms should be possible, without kernel patching or
any extra NTP servers.
Don't use the broadcast mode, it's mainly useful to reduce the load.
The most important thing is to set minpoll and maxpoll for the server
specified in ntp.conf. Generally, lower is better unless the network
load to the NTP server is a concern. With ntp-4.2.4 the allowed
minimum is 4, with 4.2.6 and it's 3.
This alone probably will be enough to get the accuracy below 1 ms.
If not, depending on network jitter and how stable are the clients'
clocks, you might want to consider disabling the kernel discipline and
using the daemon loop in FLL mode instead. With low jitter and low
quality clocks and/or high temperature variations, this can improve
the accuracy up to factor 20, in the opposite case it can make it
about 8 times worse, assuming the kernel is compiled with SHIFT_PLL=4.
For example:
server 172.17.255.254 iburst minpoll 4 maxpoll 4
disable kernel
tinker allan 16
The allan parameter specifies at what update interval will be FLL
used, set it to the polling interval to ensure it's always enabled.
With ntp-4.2.4 it's specified in seconds, with 4.2.6 it's a power of
two.
_______________________________________________
questions mailing list
[email protected]
http://lists.ntp.org/listinfo/questions
