[time-nuts] Next step up from basic GPS/PPS timekeeping
Esa Heikkinen
tn1ajb at nic.fi
Wed Feb 24 13:33:54 EST 2016
> What would be my next step up be, hardware-wise, in terms of improving precision, stability, etc? A GPSDO? Budget is limited as far as these things go - about £150 UK/$210 US.
I have Symmetricom TS2100 with OCXO and GPS upgrades as primary server,
accessible only at local network. It has about +-50 ns nominal PPS
accuracy. Then there's another server based on Raspberry Pi2, which is
accessible from Internet. The PPS from symmetricom is fed to Rpi2 and
the actual time and date information is passed from TS2100 to Rpi2 with NTP.
Rpi2 can serve about 10000 NTP queries per second, according my stress
tests with ntpload.c test program by Kasper Pedersen. For stress test,
several PC's are needed, each running few threads of ntpload.c, all
bombarding the server under test simultaneously. Rpi2 has quad core
processor, but unfortunately it looks that ntpd can utilize only one
core. With about 10000 ntp queries per second there's only one core with
100% load and others are basicly idling. Good thing is that you can run
another tasks simlataneously, without noticeable slowdown.
I also joined the NTP pool (only Finnish one) but then I noticed that
it's was really bad idea to run any NTP server behind router/firewall.
When there's rush peaks in NTP pool, there will be several hundreds (I
had more than 600) NTP queries per second and each from different
address of course. This will create a lot of sessions in router NAT
table, in my case there was more than 13000 active sessions according to
conntrack -C. This may overwhelm the router and in any case it will eat
lot of RAM on the router. The lesson learnt was that NTP server should
have it's own direct WAN access without any router. That might also
decrease the delays. This will be my next setup. The Rpi2 NTP server
itself seems to be very stable and there was no issues when I did the
stress test with continuous ~10000 queries/sec for about one day.
Regards,
--
73s!
Esa
OH4KJU
More information about the time-nuts
mailing list