[ntp:questions] Re: Ntpd accuracy

Richard B. Gilbert rgilbert88 at comcast.net
Wed Jun 8 19:15:54 UTC 2005


Eric Liu wrote:

>Hi all,
>
>I am testing Nptd on office LAN. The server is my office PC(a dell computer,
>P4 2.8G, 256M RAM, Windows 2000 sp4). Below is its ntp.conf:
>----------------------------------------------------------------------
>driftfile "C:\Program Files\NTP\etc\ntp.drift"
># local system clock
>server 127.127.1.1
>fudge 127.127.1.1 stratum 12
>----------------------------------------------------------------------
>
>The client is a PC104(586CPU) running embedded Linux. Below is its ntp.conf:
>-----------------------------------------------------------------------
>server	192.168.0.121 minpoll 4	# The dell computer is time server
>driftfile /etc/ntp.drift
>----------------------------------------------------------------------
>
>Below are some data output by client Ntpd:
>offset (s)               jitter
>-0.000106       0.000836552
>-0.0001345      0.00079245
> 0.0001785      0.000786682
>-0.0001575      0.000726
>-0.0011095      0.000707
>-0.0005755      0.0001365
>-0.000439       0.000005
> 0.004018       0.0032485
> 0.005664       0.001646
> 0.008183       0.002519
> 0.0094615      0.0012785
> 0.0114985      0.002037
> 0.012597       0.0010985
> 0.013009       0.000412
> 0.011785       0.001224
> 0.01082        0.000965
> 0.0073825      0.002506541
> 0.003579       0.002915856
> 0.016884       0.001566162
> 0.0151845      0.008989183
> 0.004747       0.014612339
>-0.023417       0.009749554
>-0.019694       0.002824951
>-0.012324       0.004500296
>
>We can see that the offset varys in a very large range. How can I make the
>offset much more regualr?  After our hard work, can I have 1ms accuracy?  If
>there are many clients(in fact, there will), 1ms accuracy means among all
>clients, relative to each other.
>
>Thanks and regards
>Eric
>
>  
>
Your largest offset is 23 milliseconds (absolute value)  which is not 
bad for a machine synchronized to a server using its local clock as a 
reference.    The local  clock on virtually any computer will not 
perform quite as well as a typical cheap wristwatch!!!!!!!!   The wrist 
watch, after all, is designed to keep time and has no other purpose.

The situation is analogous to one drunken driver trying to follow 
another!  The server's clock drifts but the client assumes that the 
errors are caused locally. . . .


The fix is to use a stable and accurate reference for the the server.  
The cheapest, assuming you have a full time internet connection, is to 
configure one, or four stratum two network servers and run your local 
server at stratum three.  The single server configuration means that 
your site will follow that server, right or wrong.  With four servers 
the NTP algorithms protect you against a single "falseticker". 

A better choice is to install a GPS timing receiver as a reference 
clock.  The cost is reasonable and performance is superb if you can site 
an antenna where it will have an unobstructed view of the sky.  If this 
is not feasible there are other alternatives.  Roughly equivalent 
performance, at much greater cost, can be had from rubidium or cesium 
frequency standards (with clock option).  A clock driven by an  Oven 
Controlled Crystal Oscillator (OCXO) is not as good as GPS, rubidium or 
cesium but could be a good deal cheaper and will still be more stable 
and more accurate than the local clock.



More information about the questions mailing list