From elog:
  http://midtwist.triumf.ca:8081/EL/020829.16670?exp=twist

I've noticed that there is a systematic bias between readings of the M
visual scaler and the corresponding ODB/history variable: 
when running with average M(visual)=1.9kHz the corresponding value on
the history plot was about 2.1kHz.  

Glen did a quick test by putting a generator signal in an unused CAMAC
scaler input. The corresponding ODB variable fluctuated and had a
higher value than the generator frequency.

We put a reference frequency into CAMAC scaler 18 (17 in the
zero-based slow control numbering). It's RF/1024 ~= 22514Hz.  So the
actual rates can be recovered.  (This scaler is labelled "Reference".)
The same signal is now sent to the counting room in CRPCH08.

An investigation by Konstantin showed that indeed, the scaler rates we
record in history and in the data stream are not real frequencies.
Instead, slow control reads cumulative counters and subtract values
obtained in consequent invocations.  The reads do not occur exactly
every second, and this explains the observed effect.