OpenBSD 4.1 running on a somewhat strange box, a Soekris 5501. The
hardware clock on the Soekris isn't super-accurate, but it is at least
so-so (drifts by maybe a minute per day if I turn NTP off). No, the
well-known hardware bug that made the hardware clock run 3x too fast
has long been fixed. This machine is always on, and only rebooted
every few months.

I'm running the standard NTPD that comes with the default OpenBSD
install. It uses pool.ntp.org. I'm on a permanently-connected pretty
fast network, with typical ping times to pool.ntp.org of 30-40 ms.

Here comes the wierdness: If I check how accurate NTP is running with
"rdate -ncpv pool.ntp.org", it is usually off by 23 seconds. If I
hand-set the clock with rdate -ncv, then the clock becomes perfectly
accurate, but within a few hours drifts back to being off by 23
seconds.

What gives? Can I do something to improve this?

Note that this is not TERRIBLY important, as have a super-accurate
clock is not vital (if it were, I'd run the ntp.org version of NTPD,
and hook up a cheap GPS receiver on the serial port). But it is still
weird; with a good network connection I should get subsecond accuracy
easily.

--
Ralph Becker-Szendy _firstname_@lr_dot_los-gatos_dot_ca_dot_us
735 Sunset Ridge Road; Los Gatos, CA 95033