[ntp:questions] Re: Allan variance plots

David L. Mills mills at udel.edu
Sat Sep 20 19:33:19 UTC 2003


Piotr,

To really get into this business, you need to think like a physicist,
engineer and computer scientist and rub your belly at the same time.
However, if you need to compare timekeeping on various machines, you
probably don't need all the tools. From experience, you can probably
assume the random-walk frequency noise is between the limits called out
in my paper. The only question remaining is how much the CPU temperature
change affects the frequency. I hear the CPU temperature can indeed
change if you throw a ton of floating point operations at it. You should
confirm and calibrate if significant.

Just about every modern architecture includes  a PCC or TSC counter with
resolution far better than needed for the usual NTP discipline. So, the
question then is the latency variations, which appear as jitter. My
experience is to run the jitter.c program in two or three processes,
collect the histogram and analyze it. One of the more telling statistics
is the ntpd precision measure, which varies from 42 microseconds in an
old SPARC IPC to just 0.4 microseconds in the Blade 1000. Unless the
operating system has little evil faults, like disabling interrupts to
fondle the disk controller, that is the defining characteristic.

Dave

Piotr Trojanek wrote:
> 
> In article <3F68A9FF.78018661 at udel.edu>, David L. Mills wrote:
> >Here is a Matlab program to read loopstats files and construct Allan
> >deviation plots. You can modify it for another language or file format.
> >The key is in the first while loop.
> 
> Thanks. I modified it for octave (free, mostly Matlab (R) compatible),
> works really great (and much faster compared to soft I was using for
> last few days).
> 
> I would like to answer myself a question, how various OS are capable
> of timekeeping. I was thinking about FreeBSD, NetBSD, Linux and two
> real-time OS, QNX and ECOS. Also I would like to check this on both
> i486 and PentiumII (without and with Processor Cycle Counter).
> 
> I'm still not sure does Allan deviation is what I need, but I just
> wanted to make these plots anyway. Papers about this topis are quite
> difficult for me to understand...
> 
> --
> Piotr Trojanek



More information about the questions mailing list