timed/ntpd

[Vernon Schryver]

In article <9010102100.AA02185 at frodo.Physics.McGill.CA>, loki at physics.mcgill.ca (Loki Jorgenson Rm421) writes:
> 
> 	Can anyone comment on whether IRIX 3.3 supports date/time
> via ntp either with the existing timed or with some appropriately
> hacked version of ntpd?  I understand that the new IRIX 3.3 has
> improved timing but I would still like to tie the PIs in with
> our SUNs which use NTP.


I think NTP is a good protocol.  Someday, it may be in the standard release.

Until then, if I were in your situation, I would use timeslave to connect
one of my IRIS's to a NTP ticker at a fairly good stratum, and the use
timed and a netgroup synchronize the rest.  You could use more than one
IRIS-NTP pair for redundancy.

Timeslave was developed before NTP was public knowledge, when the SGI
network was connected by a heavily congested, somewhat unreliable 9600 b/s
TCP/IP link to a machine at another company with a "satelite receiver."  It
uses the UDP time port for the date and ICMP timestamps for milliseconds.
It has some (probably naive) filtering to fight things like occassionally
assymmetric round trip delays.  It can be an entertaining, independent
measure of the accuracy of other time mechanisms.

After NTP is available (errrr, "if"--no promises, no dates, etc.  It's been
on the list for years), I would still use timed to slave the majority of a
network of IRIS's to scattered NTP players.  Last I looked, NTP is still
weak on automatically discovering and using a good time server.  Timed,
despite grievious bugs in the protocol (not to mention astonishing bugs in
the 4.3BSD implementation, which I think are fixed in IRIX), works pretty
well "out of the box."  That is the excuse used for shipping timed turned
on by default.

To avoid the wrath of the Net Gods, please do not use timeslave accross the
Internet.  It can be tuned use lots of packets.


Vernon Schryver,   vjs at sgi.com