[ntp:questions] Re: Effects of mindist setting

Jeff W. Boote boote at internet2.edu
Mon Sep 25 23:14:09 UTC 2006


Kevin Oberman wrote:
> I have a bunch Endrun Technologies CDMA clock attached to small servers
> using PPS. To prevent the jitter on the serial port from causing the
> clock to be marked as falsetick and disabling the PPS, I have added "tos
> mindist" to my configuration.
> 
> I am unclear on the actual effect of this configuration and that makes
> me a bit unsure of how I should set the value. 0.010 was one
> recommendation and 0.015 was another.
> 
> Any place I can look for an explanation of this? I don't see in in the
> documentation. 

I see Dr Mills pointed you to the mathematical description...

For the 'code' description - see ntp_proto.c:3448 (the middle of the clock 
intersection algorithm). The root_distance() function uses the mindist value to 
basically state that no clock is 'better' than the given dispersion distance.

I recommended the 0.015 value for the owamp systems simply because our 
calibration tests of the CDMA serial interface put it between 0.012 and 0.013 
off from the PPS. (fudge time1)

After thinking this over more - I don't think mindist should be as large as I said.

First - once we set the time1 value the magnitude of the dispersions should be 
decreased. (Was the 0.010 recommendation in 
http://ntp.isc.org/bin/view/Support/ConfiguringHPZ3801ARefclocks
based on any tests? Or perhaps assumptions based on Unix scheduling?)

Second, after more closely examining the code, I realize mindist actually sets 
the root distance for each and every peer to the max of the real distance, and 
this value. (This makes perfect sense - I just did not think through it.)

Unfortunately, this will effect the error estimate. For the one-way network 
latency measurements - we really want to minimize the error estimates as much as 
we can.

So, even the 0.010 value is probably larger than we want. We should experiment 
with our hardware/OS combinations and try and use the smallest value we can that 
removes the clock hopping problem.

jeff




More information about the questions mailing list