home *** CD-ROM | disk | FTP | other *** search
- This is a rough compilation from mail messages I've sent giving advice
- about configuring a site. It has appeared in even rougher form as the
- "ntpinfo" file on louie.udel.edu.
-
- Last edited on 7 May 1990.
-
- - Cary Gray
- gray@pescadero.stanford.edu
-
- --------------------------
- [1] one description of configuration
-
- Here's the approach I suggest, which maximizes redundancy while
- minimizing extraneous traffic. It also keeps you from ganging up on
- any one server.
-
- Pick three local hosts to be your campus servers.
- Pick six stratum-1 hosts, and pair them off. For each pair, pick an
- external stratum-2 that syncs with primaries other than that pair.
- Each of your local servers is then configured as
- peer stratum-1-from-a-pair
- peer other-stratum-1-in-the-pair
- peer corresponding-stratum-2
- peer another-of-the-local-servers
- peer the-third-local-server
-
- This is an extremely robust configuration. The second stratum-1 in
- each pair can be farther away (as the packets fly), since the traffic
- will normally be light. It is more resistant to falsetickers and
- generates less traffic than having multiple machines all talking to the
- same set of external peers.
-
- (By "external" I mean outside your organization/site--not one of the
- three you're configuring.
-
- The external stratum-2 should provide your host a path to two
- stratum-1 hosts in addition to those it talks to directly. If you're
- dealing with the public stratum-2 fuzzballs, that's pretty well
- guaranteed, since each of them squawks with lots of primaries. When
- you're dealing with other stratum-2 machines, you can get the info from
- their keeper--who you want to keep in touch with, anyway.)
-
- The remainder of your larger machines can sync from your three
- stratum-2s; the rest can just peer with a pair of stratum-3s.
-
- For example, kermit.stanford.edu's configuration looks like this:
-
- peer clepsydra.dec.com
- peer ncarfuzz.ucar.edu
- peer holmes.lcs.mit.edu # by arrangement
- peer neon.stanford.edu
- peer ahwahnee.stanford.edu
-
- Holmes squawks with three stratum-1 servers: bitsy.mit.edu,
- umd1.umd.edu, and suzuki.ccie.utoronto.edu. With holmes, neon,
- and ahwahnee-doda, kermit is just two hops away from 7 different radio
- clocks should both clepsydra and ncarfuzz go south.
-
- My others aren't quite so lovely. Ahwahnee's primaries are wwvb.isi.edu
- and truechimer.cso.uiuc.edu, and its external stratum-2,
- lilben.tn.cornell.edu, does talk to truechimer along with a couple of
- other primaries.
-
- If you want to learn about peer sets on machines that don't answer
- ntpdc--fuzzballs and host running xntpd--grab the 'ntpq' program from
- the xntpd distribution.
-
- ------------------
- [2] another description
-
- You probably want three levels in your configuration. Here's the way I've
- set it, which gives a lot of redundancy, but without generating a lot of
- traffic.
-
- (1) Campus primaries
-
- These talk to the outside world; you'll want two or three of these.
- If you have three (A, B, and C), the configuration for A should look
- like this:
-
- peer first-stratum-1
- peer second-statum-1
- peer external-stratum-2
- peer B
- peer C
-
- B and C should be similar, but with different outside hosts. Also, you
- should make sure that 'external-stratum-2' talks to at least two
- stratum-1 hosts other than 'first-stratum-1' and 'second-stratum-1'.
-
- If you're a small site, you might just two of these machines, in which
- case you probably want three stratum-1's each. (If you're really small or
- rather distant from the backbone, you might want to team up with another
- nearby site (or sites) to bring in three or four stratum-2s; then add another
- level or half-level at your site.)
-
- A note on picking statum-1's: give each of your hosts one that is
- "nearby" (as the packets fly), with its others spread around.
-
- (2) Secondaries
-
- You want a pair of these per group of hosts, where a group is easily
- upt to 60 or so. Each should peer with all of the local primaries,
- with its partner, and with a secondary in a another group. For
- example, if the pair for a group are X and Y, then X would have
-
- peer A
- peer B
- peer C
- peer Y
- peer a-secondary-in-a-different-group
-
- As general rule, X and Y's outher secondaries shoudl come from
- different groups.
-
- (3) End systems
-
- Someday multicast will handle this level automatically. In the
- meantime, the remaining systems in the group with X and Y just need
-
- peer X
- peer Y
-
- Some general notes:
-
- Using "server" instead of "peer" has two results: the named host won't
- synch to you, and it won't keep track of you either. The latter makes
- it polite when you're banging on a machine that has lots of other hosts
- chiming with it, as do most stratum-1 hosts. At one time that mattered,
- but the implementations have been improved; the consensus now favors using
- "peer" (almost) everywhere.
-
- There are also two differences between "peer" and "passive". First, if
- you configure "passive", your host won't send out packets to the other
- unless it hears from it first. So you can use it to cut down traffic if
- one machine out of a pair is down a lot.
-
- The other difference comes into play only when you configure a server
- with "trusting no". It then will synch only to hosts mentioned in the
- configuration file: one's you trust can then be named as passive.
- (I've made my local primaries non-trusting because some other folks
- peering with the public one have really brain-damaged configurations
- that could otherwise confuse false-ticker detection.)
-
- ----------------
- [3] configuring ntpd for a reference clock
-
- ... [from a message to someone on the west coast]
- ... You'll still need to pick external peers, both
- stratum-1 and -2. pub/ntp/clock.txt on louie.udel.edu lists a number
- of public servers. I think you'll do well with all of the good
- west-coast primaries: wwvb.isi.edu, fuzz.sdsc.edu, and
- clepsydra.dec.com (which is here in Palo Alto). You'll also probably
- do well with the other fuzzballs on the NSFnet backbone, which are
- listed in clock.txt. I can offer stratum-2 from here at Stanford:
- kermit.stanford.edu is public, and ahwahnee and neon are
- available by arrangement. My personal advice is to steer clear of the
- WWV clocks advertised as stratum-1, since they don't seem to give very
- stable time, and the three I'm aware of near me (apple.com,
- hp850.mbari.org, and norad.arc.nasa.gov) are the reason my hosts have
- gone non-trusting.
-
- To configure a local clock, you have to compile in the reference clock
- support, and put something like this into your ntp.conf:
-
- peer /device reference-id stratum precision type
-
- For the system clock, /device just needs to start with a "/" (it would
- be a tty line if using a radio clock), and type is "local". Precision
- is the same as configured (-7 on vaxen), and stratum is whatever you
- want, e.g., 6. I'm not sure what the blessed name for reference-id
- is---up to four characters, maybe SYS? So that comes out looking like
-
- peer /dev.null unix 6 -7 local
-
- If you do this, keep the stratum pretty high (> 5) so that you don't confuse
- anyone. If you want to order a set of hosts in this manner, increment
- stratum by 2 as you head down the ranking.
-
-