[ntp:questions] Answer: PPS on a rack server
John Ioannidis
ntp at tla.org
Tue Jun 3 01:20:39 UTC 2008
A few days ago I had asked this list for assistance with the following
problem: my rack-mounted IBM x3650 servers only have one serial port,
and the data I was feeding them from my GPS clock were interfering with
the serial console feature, the BIOS in particular. Thanks to several
people who replied to my original message, I now have a working setup.
Here it is:
1. A Spectracom 9283 netclock feeds *only* the PPS signal to the
servers. Pins 1 (DCD) and 5 (GND) of a DB-9F connector are connected to
the core and shielding of the coax (through T connectors, and the cable
is terminated with a 50 Ohm terminator from the old thinlan days!). The
signal seems to be strong enough to drive the DCD pin, but for longer
runs I would use this: http://www.tapr.org/kits_tadd-3.html.
2. NTP is used to get the time reference from the Spectracom.
3. The harder part was figuring out what to put in /etc/ntp.conf. Here
is what I have (the address of the Spectracom is not the real one, of
course)
server 192.2.200.2 minpoll 5 maxpoll 5 prefer # 9283
server 127.127.22.0 minpoll 5 maxpoll 5 # PPS
fudge 127.127.22.0 flag3 1
Of course, the kernel is configured with option PPS, and /dev/cuad0 is
symlinked to /dev/pps0.
That's it! The machines appear to be within a few microseconds of each
other, so I'm happy.
4. I can still use the serial console, after a fashion -- the machines
support IPMI, so I can use serial-over-lan with ipmitool to get console
i/o (for the BIOS, which is all I care about) without actually needing
to plug anything to the serial ports. Of course, what I really should
have done is paid the extra $1K or so and gotten the RSA-II cards for
the boxes, but that's another story.
Thanks again to all who replied
/ji
More information about the questions
mailing list