[ntp:questions] Allan deviation survey
unruh at wormhole.physics.ubc.ca
Fri Sep 10 22:01:33 UTC 2010
On 2010-09-10, David L. Mills <mills at udel.edu> wrote:
> This is a multi-part message in MIME format.
> Content-Type: text/plain; charset=UTF-8; format=flowed
> Content-Transfer-Encoding: quoted-printable
> I've done this many times with several machines in several places and=20
> reported the results in Chapter 12 and 6 in both the first and second=20
> editions of my book, as well as my 1995 paper in ACM Trans. Networking.=20
> Judah Levine of NIST has done the same thing and reported in IEEE=20
> Transactions. He pointed out valuable precautions when making these=20
> measurements. You need to disconnect all time disciplines and let the=20
> computer clock free-wheel. You need to continue the measurements for at=20
> least a week, ten times longer than the largest lag in the plot. You=20
> need to display on log-log coordinates and look for straight lines=20
> intersecting at what I have called the Allan intercept. I have Matlab=20
> programs here that do that and produce graphs like the attached.
What was the load on those computers? Were they running just this time
measurement software or were they being used for real work by normal people?
>From what I have seen from machines that are in use, they heat up
during the day when people use them and cool off at night when they are
idle. The amplitude of this component depends on how much work is being
done. This does not fit the model of random phase noise/random walk
frequency noise. It has a strong periodic component with period of a
day. I see this in most of my systems. Such a periodic noise is not part
of the noise model on which the Allan intercept is based.
Also this assumes a very particular model of how the measurements are
made, of how the time corrections are made and of what is desired from
> Most of the prior work was done 15 years ago. I strongly suspect the=20
> Allan intercept has moved to lower time values due to the fact that=20
> modern processors are faster and the interrupt latency is smaller. The=20
> current NTP distribution includes a NTP simulator that can be excited=20
> with white phase noise and random-walk frequency noise that very nicely=20
> models the real noise sources.
> For those that might want to repeat the experiments, see the attached=20
> figure. Trace 1 is from an old Sun SPARC IPC; trace 2 is from a Digital=20
> Alpha. Traces 3 and 4 were generated using artificial noise sources with=20
> parameters chosen to closely match the measured characteristics. Phase=20
> noise is generated from an exponential distribution, while frequency=20
> nose is generated from the integral of a Gaussian distribution, in other=20
> words a random walk. Trace 4 is the interesting one. It shows the=20
> projected performance with precision of one nanosecond. The fastest=20
> machines I have found have a precision of about 500 ns. Note, precision=20
> is the time taken to read the kernel clock and is not the resolution.
You graph of course did not make it through to newsnet.
Have you archived the figure somewhere?
More information about the questions