C:\tvb\tsc036\phm-8607> dir/os | tail
LOG186~1 GZ      2,799,341  05-09-05  5:47a log18676.txt.gz

C:\tvb\tsc036\phm-8607> gzip -d < log18676.txt.gz | wc
7.333 MB, 7509 KB, 7689048 chars, 548573 words, 548573 lines

C:\tvb\tsc036\phm-8607> dir/od | more
LOG18~12 GIF        23,900  05-06-05 11:00a log18689v.gif
LOG18~13 GIF        21,739  05-06-05 11:02a log18690v.gif
LOG18~14 GIF        14,245  05-06-05 11:03a log18691v.gif
LOG18~15 GIF        35,136  05-06-05 11:04a log18692v.gif
LOG186~1 GZ      2,799,341  05-09-05  5:47a log18676.txt.gz
LOG18~16 GIF        23,980  05-10-05  1:02a log18698v.gif
LOG18~17 GIF        22,299  05-10-05  1:03a log18699v.gif
LOG187~1 GIF        14,228  05-10-05  1:03a log18700v.gif
LOG187~2 GIF        35,978  05-10-05  1:04a log18701v.gif

Notes: open log file as phase, with tau 1 s, with scale by 2e-7 (= 200 ns = 5 MHz); plot

• This is a boring plot, but at least shows no obvious glitches.
• Note the large accumulating time offset due to OCXO being off-frequency; about 2 ms time drift over the 6+ day run.
• Note a 1×10^-8 offset in frequency is 10 ns per second offset in time, or almost 1 ms accumulated time error per day.
• Note also with 500k points these plots take a while to render and it's tempting to average down by 10 or 100. However, ...
• Averaging hides glitches which may not be a good idea until the data is known good.
• Averaging also eliminates low tau from stability calculations.

Notes: phase slope is frequency offset; calculate and remove frequency offset (linear fit); re-plot

• Calculated frequency offset is -4.124e-9.
• With the linear term removed (frequency offset) the data is much clearer.
• Time drift is limited to about 3 µs over the run.
• The hump shape of the curve indicates frequency drift is present.
• No glitches visible at this scale either; looks like a clean run.

Notes: frequency slope is drift rate; calculate and remove frequency drift (quadratic fit); normalize residual phase; re-plot

• Calculated frequency drift is -6.984e-17 (change in frequency per second).
• Multiplying by 86400 makes this -6.034e-12 / day, in conventional drift units.
• Time drift is confined to 1 µs over the run.
• Note that it makes little difference in calculated coefficient c if a frequency offset is first removed before the quadratic fit.

Notes: run ADEV from raw phase data (no unfair drift removal at this point); 1-2-4-decade scale; plot

• Stable32 computes Freq Drift/Day=-6.036e-12

Notes: run OADEV & MDEV from raw phase data; re-plot

• Very little difference between the two and the plain ADEV plot above.

Notes: run ADEV from raw phase data, selecting remove drift; re-plot

• Not much difference between this no-drift plot and the normal plot above.
• This suggests there is some long-term instability that isn't purely linear drift.
• Suggests long-term performance is not solely due to ageing.
• A look at a frequency plot will confirm this.

Notes: conv raw phase to frequency; plot

• This isn't pretty for a couple of reasons.
• The y-scale is absolute frequency, hiding relative frequency.
• But if accuracy were more important than stability this would be a good thing.
• There is a noticeable amount of noise in the graph (probably occasional maser glitches).
• For a better frequency plot, remove some short-term noise with averages.

Notes: average by 10x (10 s samples); normalize (remove mean frequency); add linear trend line (frequency drift) but do not remove drift; re-plot

• Much better looking frequency plot.
• Line fit says drift is -7.226e-16 (per 10 s) = -6.243e-12/day.
• Short-term frequency noise is small enough that long-term trend is clearly visible.

Notes: average again by 10x (now 100 s samples); remove frequency drift; normalize; center scale to ±10×10^-12; re-plot

• This represents best-case potential performance if
• There were no systematic linear frequency drift over time (ageing).
• Other random variations are present (it is quartz, after all).

Notes: open phase, conv phase to freq, use check and stats to look for spikes. Note sample numbers, go back to phase, use stats to zoom in; extract nice 10 minute record using part; [i5500 n600]; use 599 points (not 600); remove slopes and normalize both phase and frequency; plot each

• This nicely shows the glitches in both frequency and time domain.
• Frequency spikes are often one of two types.
• Those with a single spike out of the normal noise, which correspond to a phase jump.
• And those with a double spike, which corresponds to a bad phase data point or a
• This example shows one of each.
• The frequency spikes in this example are about 5×10^-12 in magnitude over a few seconds.
• The corresponding phase jumps are about 12 ps and ±3 ps.
• Frequency glitches are often one of two types; those with a single spike above or below the line (which corresponds to a one-way phase jump), and those with a double spike, first one direction and then the other,