Merit Network
Can't find what you're looking for? Search the Mail Archives.
  About Merit   Services   Network   Resources & Support   Network Research   News   Events   Home

Discussion Communities: Merit Network Email List Archives

North American Network Operators Group

Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical

Re: Standalone Stratum 1 NTP Server

  • From: David G. Andersen
  • Date: Wed Aug 28 00:54:26 2002

On Tue, Aug 27, 2002 at 11:57:39PM -0400, John Todd mooed:
> 
> Hmm... $2400 is still in the "pricey" range to be throwing out 
> bunches of these across a network in wide distribution.  (Pardon me 
> [...]
> 
> One would think that a vendor could come up with a 1u rackmount box 
> with a GPS and single-board computer (BSD or Linux-based) for ~$500 
> total cost.   Add 150% for profit and distribution costs, you're 
> still in the $1300 range, which is more reasonable.  I suppose my 
> oversimplification is the reason I'm not in the hardware business. 

   You might be imagining a somewhat larger market for standalone
stratum-1 timeservers than you might imagine.  For real accuracy, 
you don't want standalone -- you want a locally connected source that 
you can use to tightly discipline the local clock.  (When I say "real," I
mean sub-milisecond).  The difference is .. substantial.  Taken
from two of my machines on the same subnet:

     remote       st poll reach  delay   offset    disp
========================================================
Local CDMA         0   32  377 0.00000 -0.000011 0.00047
100Mbps Ethernet   1 1024  377 0.00035  0.001103 0.01866

And if you want paranoia, go by ntp's estimate of its accuracy:

Local  maximum error 5449 us, estimated error 3 us, TAI offset 0
Ether  maximum error 584994 us, estimated error 1241 us, TAI offset 0

With that on the board... why do you need, or even want,
a standalone NTP server if you're on a budget?  Almost certainly
you have a local computer in your POP -- you can even get
Cisco routers to talk with a local time receiver, if all you
want to do is discipline your routers.  If you've got a caching
nameserver or something else in your POP, that will do just as
well.

> I'd be even happier with a PCI-bus card that I could put into an old 
> (reasonably fast) PC and a CD-ROM with an OpenBSD distribution that 
> automatically did the Right Thing.   There is a case to be made about 

  Grab a serial CDMA/GPS unit (I use the EndRun Praecis Ct because it's
CDMA;  I mention some GPS units below), plug it into your serial port, and
stick:

server 127.127.29.0 prefer
fudge 127.127.29.0 refid CDMA

   in ntp.conf.  It's about as simple as you can get.  But remember --
regardless of how nifty your local clock is, you still need to
have a good server mesh with NTP.  Clocks go bad.  CDMA base stations
screw up (we've found one so far) or change protocols unexpectedly
(three).  GPS has serious visibility issues unless you can get an actual
roof antenna (two).  Configuring this mesh in an intelligent way takes work.
Would make a great research project. :)

  The Ct costs something like $1100.  endruntechnologies.com.
synergy-gps.com sells a really nice GPS timing unit based on the
Motorolla UT+ chipset (designed for timing), including all the parts
you need, for .. eh, 600?  I forget.  Maybe a bit less.  Plug into
serial port, go.  Requires a recompiled kernel under FreeBSD and
Linux, but it's fairly easy to set up.  If you want something for
a bit less work, look at the Trimble units.

  (For reference:  I've got two of the UT+ GPS units, and 20
EndRun Praecis Ct's.  Like them both.  The Ct is a heck of a lot
easier to deploy in a datacenter, as would be the CDMA TrueTime model)

  If you're really broke, and want a stratum 1 server, host one of
our network measurement boxes.  We'll ship it to you, you provide
the network.  In return, you get a local stratum-1 timeserver, 
managed by yours truly.  (I'm serious about this offer, btw.)

  As a second option:  If you manage the connections between
your POPs, you can get really decent remote NTP performance.
The places in which NTP dies are where latencies are asymmetric.
With priority assigned to inter-POP NTP traffic and known
symmetric links, life could be quite happy.

   -Dave (time is very cool)

-- 
work: dga@lcs.mit.edu                          me:  dga@pobox.com
      MIT Laboratory for Computer Science           http://www.angio.net/
      I do not accept unsolicited commercial email.  Do not spam me.




Discussion Communities


About Merit | Services | Network | Resources & Support | Network Research
News | Events | Contact | Site Map | Merit Network Home


Merit Network, Inc.