Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!swrinde!zaphod.mps.ohio-state.edu!van-bc!ubc-cs!uw-beaver!apollo!hpway!sommerfeld From: sommerfeld@apollo.hp.com (Bill Sommerfeld) Newsgroups: comp.sys.apollo Subject: Re: ncs-clocks skew warning Message-ID: Date: 20 Mar 90 00:10:00 GMT References: <9003191219.AA12175@apo.esiee.fr> Sender: root@apollo.HP.COM Organization: HP Apollo Systems Division, Chelmsford, MA. Lines: 44 In-reply-to: bonnetf@apo.esiee.fr's message of 19 Mar 90 11:15 GMT In article <9003191219.AA12175@apo.esiee.fr> bonnetf@apo.esiee.fr (bonnet-franck) writes: What does it EXACTLY means ? It means exactly what it says :-). The replication algorithm used by the GLB depends on the clocks on the nodes running the GLB being loosely synchronized; the "clock skew warning" means that the clocks are drifting too far apart. The GLB won't correct this for you automatically; instead, you have to synchronize the clocks in some other way. Doing it manually: - using a wristwatch and /bin/date whenever you see the "clock skew warning". Doing it automatically (more work to set up; less work over time): - If both glbd's are on the same physical network and both have IP (some say "TCP") enabled, set the nodes up to run /etc/timed all the time. Timed should keep your clocks synchronized to within a second or so. It works by "averaging" all your clocks together, so it works best if there are a relatively large number of clocks participating. I believe timed is new as of SR10.2; the binary included in SR10.2 might work on 10.1, but I haven't tried it. - If you have a flair for overkill, and you're connected to the Internet or have access to a proper time source (typically a radio clock), you can bring up NTP. This will tend to keep your node clocks synchronized to within about 10-20 milliseconds of the correct time. There are two different implementations of NTP for UNIX which are generally available, "ntpd", done at the University of Maryland, and "xntpd", by Dennis Ferguson of the University of Toronto. I've managed to get both working on SR10.2, and if enough people are interested in patches, I could be convinced to make them available, although porting them was really very trivial. "xntpd" is a bigger, more featureful implementation; "ntpd" is kind of minimalist. - Bill Sommerfeld, amateur clock watcher sommerfeld@apollo.com sommerfeld@apollo.hp.com