[ntp:questions] PPS kit for 2.6 linux kernels?
Michael L. Semon
mlsemon33 at verizon.net
Sat Apr 14 22:22:37 UTC 2007
ramadog at gmail.com wrote:
> Michael L. Semon wrote:
>> 1) Use my current Linux 22.214.171.124 + LinuxPPS and sit and watch as ntpd
>> darts +/- 120 us over the course of a few hours, even when idle.
>> 2) Reboot to FreeBSD and let the kernel, hardpps, and the NTP kernel
>> code make more radical short-term adjustments, and ntpd will dart +/- 10
>> us over a few hours.
>> 3) Recompile glibc with support with older kernels without breaking my
>> system, then fetch Linux 126.96.36.199, fetch Ulrich's last PPSkit alpha, and
>> hope it all works.
> If you do try a 2.6.17.X kernel maybe try with linuxpps. It seems
> there were changes to the linux clock in 2.6.18 that make convergence
> a lot slower. With 2.6.17 or 2.6.16 + linuxpps and default minpoll
> maxpoll my time usually stays within +/- 10us. Minpoll 4 maxpoll 4 and
> it is down to +/- 3us. For me 2.6.18 and later with default minpoll
> maxpoll means +/- 250us. Minpoll 4 maxpoll 4 and I am still looking at
> +/- 40us.
> My drift changes by less than 0.6ppm during the temperature cycle of a
> day. Freebsd or 2.6.17 (and lower) + linuxpps have no trouble
> following that change. 2.6.18 and later react too slowly.
I might try that, one day in the near future.
Thank you for motivating me to go back to kernel 188.8.131.52 kernel. I
tried the PPSkit patch that is meant for that kernel, and it's a
definite improvement over my old setup. [BTW, there was a typo in my
previous post. I was using 184.108.40.206, not 220.127.116.11.] The PPSkit patch
is marked as alpha for a reason--it burps a little bit, and it does not
like the 1000-Hz kernel timer frequency at all--but overall, it does a
decent job at the 100-Hz and 250-Hz timer frequencies.
More information about the questions