[2854] in SIPB-AFS-requests
Re: xntpd running on ra and rc
daemon@ATHENA.MIT.EDU (John Hawkinson)
Thu Sep 4 02:12:38 1997
Date: Thu, 4 Sep 1997 02:12:22 -0400
To: Tom Yu <tlyu@MIT.EDU>
Cc: sipb-afsreq@MIT.EDU
In-Reply-To: "[2853] in SIPB-AFS-requests"
From: John Hawkinson <jhawk@MIT.EDU>
| I've compiled and installed xntpd for SunOS, and it's currently
| running on ronald-ann and reynelda. I've also updated the config file
| for rosebud. I may be tweaking these a little, so I'm not putting the
| configs in afs just yet.
It looks like this isn't working very well, and there's a very large frequency
offset on ra and rc:
ntpq> host ra
current host set to RONALD-ANN2.MIT.EDU
ntpq> rv
status=06f4 leap_none, sync_ntp, 15 events, event_peer/strat_chg
system="SunOS", leap=00, stratum=3, rootdelay=21.77,
rootdispersion=36.24, peer=41541, refid=BITSY.MIT.EDU,
reftime=b7b8ca4c.25e9c000 Thu, Sep 4 1997 1:49:00.148, poll=6,
clock=b7b8ca53.77f4b000 Thu, Sep 4 1997 1:49:07.468, phase=-16.613,
->FREQ=343041.61, error=21.03
ntpq> host rc
current host set to REYNELDA.MIT.EDU
ntpq> rv
status=06f4 leap_none, sync_ntp, 15 events, event_peer/strat_chg
system="SunOS", leap=00, stratum=3, rootdelay=25.54,
rootdispersion=17.85, peer=9805, refid=BITSY.MIT.EDU,
reftime=b7b8ca35.d4f17000 Thu, Sep 4 1997 1:48:37.831, poll=6,
clock=b7b8ca55.b197f000 Thu, Sep 4 1997 1:49:09.693, phase=-5.357,
->FREQ=338710.10, error=6.44
[caps mine]
For comparison, rb has:
status=06f4 leap_none, sync_ntp, 15 events, event_peer/strat_chg
system="UNIX/Ultrix", leap=00, stratum=3, rootdelay=4.94,
rootdispersion=13.69, peer=63185, refid=AGAMEMNON.MIT.EDU,
reftime=b7b8ca71.621f9000 Thu, Sep 4 1997 1:49:37.383, poll=6,
clock=b7b8ca8d.c70d3000 Thu, Sep 4 1997 1:50:05.777, phase=-1.229,
freq=-5328.87, error=7.14
Which is much more reasonable. I believe that a reasonable expectation
is that freq should be within +- 50,000, assuming no special adjustment
of "tick".
I believe the reason for this is that the "dosynctodr" kernel variable
is set to 1 (the default). dosynctodr controls whether the kernel
periodically syncs the system clock to the time-of-day (TOD) clock. (The
time-of-day clock is the battery-backed clock).
The xntpd documentation recommends that you set dosynctodr to zero
prior to starting xntpd, so that there are not multiple competing sources
setting the time. The recommended way to do this is to run "tickadj -s"
at boot time.
I asked Matt to check (with adb -- "echo dosynctodr/D | adb -k /vmunix
/dev/mem" is sufficient) and we determined that ra and rc both had
dosynctodr set to 1.
I think that's sufficient to account for this brand of lossage.
Could you please zero dosynctodr, kill xntpd, remove /etc/ntp.drift,
and restart xntpd? (And also please fix the startup file?)
Incidently, I observed some odd behavior on agamemnon's part earlier
today and Mike checked the zone cell servers for me and confirmed that
they didn't seem to be zeroing dosynctodr, so that may account for
problems there. So there may be some changes on the ops front, too.
The xntpd docs are somewhat vague as to whether zeroing dosynctodr
is appropriate under Solaris (they specify for Solaris 2.1...). Since
the stated freq offset on world-wide-weed's clock is 87,000, I'm
going to zero dosynctodr and see if it improves...
Thanks.
--jhawk
