The aim of this paper was to analyze the cellular network
performance based on an extensive measurement study. In particular
the analysis of performance limiting factors and their
root cause was targeted. The data was collected during an
extensive crowd-sourcing study and additional targeted measurement
campaigns, generating reference measurements using
a more extensive test set. From this data, a number of observations
regarding the cellular network and its performance are
derived. The crowd-sourced measurements are well suited to
derive general, location based network parameters like cell size
and density, as well as available technologies and the respective
signal strength. Compared to simple RTT and throughput
measurements, the network performance related parameters can
only be derived using dedicated measurements and a more
complex measurement setup. From these, the implications of
network management decisions – namely PoP selection – on the
end-to-end performance of cellular network access have been
shown. These were confirmed for general web access by measuring
the performance of the 25 most popular web sites and
determining the influence of PoP selection on the response time.
The main findings are summarized along the research questions
posed in the beginning for which the answers are outlined
in the following:
How does the network configuration and management influence
the cellular network performance? The analysis of the
stationary measurements shows that the network performance
(i.e. RTT) is highly dependent on network configuration and
management decisions. In the data sets, four different PoPs
were discovered. One of these is selected for each device connecting
to the mobile network. The assignment was observed
to be random but with unequal probabilities. The resulting RTT
between the mobile devices and the server at TU Darmstadt also
strengthens this observation. Due to the random assignment of
the mobile devices to the PoPs the performance experienced on
the mobile devices is often sub-optimal. This performance also
depends on the location of the application server. Considering
that a large number of application servers and Internet services
are located near PoP 1, and only 19% of the traffic is
routed using this PoP, the resulting performance shows a clear
optimization potential.
Can the performance of the cellular network be predicted
based on the available measurements? Analyzing the assignment
of the mobile devices to the PoPs provides an indication
of periodicity. Hence, a Markov Chain was derived to model
the starting state probability and transition probabilities for the
PoPs. This is also exploited to predict the assignment of the
devices to the PoPs. As time patterns were also visible in the
Spanish measurements, discovering similar effects in other networks
is probable. Applying a binary classification tree, a 97%
accuracy predicting the correct PoP has been achieved.
What is the optimization potential in the given network configuration,
and how can this be exploited? The analysis in
Section V shows a large additional RTT when using a suboptimal
PoP. An overhead of 58% to 73% was visible in
more than 57% of the measurements. This overhead is nearly
constant for the full time of the PoP association. Automatic
re-assignments are only carried out after a connection duration
of 36 hours. Hence, the performance of the current connection
can be considered stable. The performance does neither change
with handovers in between cells. Still, forcing a re-assignment
to another PoP is possible by disabling and re-enabling cellular
data. Such, the performance penalty imposed by the
sub-optimal PoP assignment can be mitigated.
How is the user-perceived mobile service quality affected
by the routing decisions of the mobile operator? Differences
in absolute response times between different websites were
observed in the collected data set, but cannot be accounted to
the PoP selection alone. Instead, these are caused by processing
on the server side. Comparing the relative difference in
response times of the 25 most popular websites in Germany
shows a similar behavior to the previous observations. This