Investigating the statistical properties of the captured data
closer, we provide the statistics for the captured traffic of all
three measurement runs at the second, minute and hour time
scales in Table I. We initially note that the September measure-
ment is about half the duration of the other two measurement
periods, resulting in a lower amount of bandwidth samples.
We furthermore note that in each measurement period, the
minimum bandwidth observed per second is zero. Comparing
the average average bandwidth amongst the three measurement
series, we note that the November average bandwidth is
approximately 33 time the average bandwidth observed during
the December period. We furthermore note that the average
bandwidth slightly changes with the aggregation level, which
is explained by incorporating only complete aggregates into
the calculations, which typically drops the last aggregate of
each measurement. The peak-to-mean ratios observed are ex-
tremely high for the September and December measurements,
whereas the November measurement exhibits a lower level
for the second timescale. With increasing timescales, however,
we note a significant drop in the peak-to-mean traffic ratios
for the September and December measurements, whereas the
November decreases only slightly. These behaviors indicate
that the short time traffic is typically highly variable, but
only in certain periods for September and December, whereas
November exhibits overall high traffic variability. As more
unbiased estimator for the traffic variability, we observe the
coefficient of variation in Table I. We note a similar trend
between measurements and timescales, confirming the previ-
ous observations for the peak-to-mean ratio. Lastly, we eval-
uate the overall self-similarity of the utilized server network
bandwidth in form of the estimated Hurst parameters using
the estimation approaches outlined in Section II. We observe
that all three measurements exhibit an overall level of self-
similarity, which is the most pronounced for the November
measurement. We additionally note a trend for the increase in
the Hurst parameter as the aggregation level is increased to
minutes and a subsequent decrease as the aggregation level is
further increased to hours.
In the remainder of this paper, we focus on the Novem-
ber and December measurements, as the statistics for the
September measurement resemble those of the December
measurement.