Dynamics of Internet load are investigated using measured statistics o
f round-trip delays, packet losses and out-of-sequence acknowledgement
s. Several segments of the Internet are studied. They include a region
al network (the Jon von Neumann Center Network), a segment of the NSFN
et backbone and a cross-country network consisting of regional and bac
kbone segments. Issues addressed include (a) the probability distribut
ion of round-trip delay; (b) the correlation between packet loss and v
arious statistics of delay (average, minimum, maximum and standard-dev
iation); (c) the relationship between out-of-sequence acknowledgements
and various statistics of delay; and (ii) the dominant frequencies in
the spectral decomposition of delays. A comparison of results across
different network segments (regional, backbone and cross-country) is a
lso given. Results show that (a) the distribution of delay isa shifted
gamma with scale and shape parameters that change with time of day, (
b) the conditional probability of packet loss increases with different
statistics of delay; however, the correlation between the two call be
low, especially For low values of delay and/or loss, (c) the conditio
nal probability of out-of-sequence acknowledgements increases with dif
ferent statistics of delay, but again the correlation between the two
is low for low values of delay and/or out-of-sequence acknowledgements
, and (d) there exist dominant low frequency components in the delay d
ata, The relationship between packet loss and delay indicates that the
signal-to-noise ratio in the delay data (e.g. for the purposes of ada
ptive routing or congestion control) is low for small to moderate valu
es of delay, but increases for higher values of delay. The distributio
n of delay may have applications in modeling studies of feedback-based
algorithms.