I set up a small cluster. The master makes NFS shares available. For
normal operation of e.g. make on the NFS share, the time needs to
be very tightly synchronized.
I start timed as a master on the master PC and as a slave on each
(master) in /etc/rc.d/rc.local I added
/usr/sbin/timed -M -n cluster.priv -t
(slave) I created a script /etc/rc.d/init.d/timed which issues
daemon timed -t
and is called via appropriate symbolic links.
After periods of long operation of the master system (days or even
weeks), timed on master appears to go bonkers (a technical term). On
different occasions I observed the following phenomena:
(a) master system time slows down or even goes backwards (recently
a few days in just a few minutes),
(b) timed uses up a lot of processor time (50%) until it is killed, (happened
once -- this is how I figured out who may be the culprit),
(c) the system becomes unstable and requires reboot (always).
As for (c) above, different things are happening on different occasions:
(a) X started displaying something like a TV signal between the
stations -- like modulated white noise, I had to kill it, kill timed and
reboot (this was on the same occasion that timed was taking up 50%
(b) unrelated applications (e.g. F95 compiler) start behaving strangely,
(c) other things which I can't recall now.
The problem *may* be related to manual or automatic adjustments of
time: once a day a cron process adjusts the time on the master
with an external clock (much more accurate):
/usr/bin/rdate -sp aurora
Here's a workaround for your timed problems:
Try using xntpd, it handles time warps much more gracefully than
The hardest part about xntp is choosing a server. See the url or
/usr/doc/xntp* for guidance there.
Then on the local server, you will need to do
chkdonfig --add xntpd
echo your_xntp_server > /etc/ntp/step-tickers
Add to /etc/ntp.conf
If you can, enable multicast by commenting out multicastclient and
adding broadcast as below:
#multicastclient # listen on default 126.96.36.199
broadcast 188.8.131.52 key 65535
On each client, then do
chkconfig --add xntpd
echo the_local_server > /etc/ntp/step-tickers
Restart the antpd daemons of local server and clients, wait 15
and look us "ntpq -p" to verify operation (an asterisk precedes the
server that is currently being used for synchronization).
If you do the above, all your clients will initially set the clock
from your_local_server using ntpdate, and then receive synchronization
from multicast packages.
This problem appears resolved. Please reopen if I'm wrong.