[ntp:questions] Allan deviation survey
David L. Mills
mills at udel.edu
Tue Sep 14 02:45:01 UTC 2010
You don't need a week for that, since the anticipated intercept is in
the order of 200 s (trace 3). However, plots such as these are really
susceptible to little hidden resonances, so I tend to prefer a long tail
and lots and lots of samples. For comparison, the averaging time for PPS
signal in the kernel is 256 s, which is close to the expected Allan
intercept for modern systems.
Don't get fooled by the MINSTEP. Precision is defined by the time to
read the system clock at the user interface and I have never seen
anything less than 500 ns for that, more typically 1000 ns.
Miroslav Lichvar wrote:
>On Fri, Sep 10, 2010 at 08:48:58PM +0000, David L. Mills wrote:
>>I've done this many times with several machines in several places
>>and reported the results in Chapter 12 and 6 in both the first and
>>second editions of my book, as well as my 1995 paper in ACM Trans.
>>Networking. Judah Levine of NIST has done the same thing and
>>reported in IEEE Transactions. He pointed out valuable precautions
>>when making these measurements. You need to disconnect all time
>>disciplines and let the computer clock free-wheel. You need to
>>continue the measurements for at least a week, ten times longer than
>>the largest lag in the plot. You need to display on log-log
>>coordinates and look for straight lines intersecting at what I have
>>called the Allan intercept. I have Matlab programs here that do that
>>and produce graphs like the attached.
>For the simulation and development purposes I'm interested in the most
>important part of the graph is the point at which the line starts to
>divert from the -1 slope. With good PPS signal one day of collecting
>data should be enough.
>>For those that might want to repeat the experiments, see the
>>attached figure. Trace 1 is from an old Sun SPARC IPC; trace 2 is
>>from a Digital Alpha.
>Thanks, that's very helpful.
>>Traces 3 and 4 were generated using artificial
>>noise sources with parameters chosen to closely match the measured
>>characteristics. Phase noise is generated from an exponential
>>distribution, while frequency nose is generated from the integral of
>>a Gaussian distribution, in other words a random walk. Trace 4 is
>>the interesting one. It shows the projected performance with
>>precision of one nanosecond. The fastest machines I have found have
>>a precision of about 500 ns. Note, precision is the time taken to
>>read the kernel clock and is not the resolution.
>With current CPUs the precision is well below 100 ns. (thus the
>MINSTEP constant used in ntpd's precision routine is too high)
More information about the questions