On Robust Estimation of Network-Wide Packet Loss in 3G Cellular Networks

In this work we address the problem of estimating the network-wide packet loss rate across the radio access section of a 3G cellular network. The reference scenario consists of a passive monitoring probe located in the Core Network. The probe counts the number of TCP packets directed to each individual mobile terminal and, from the analysis of (un)acknowledged packets, infers the loss ratio for each individual terminal. The problem is then to derive a synthetic indicator representative of the network-wide packet loss, which can be used to detect large- scale performance drifts and network anomalies. We show that common simplistic indicators like the total rate of lost packets (across all terminals) and the average per-terminal loss rate do not work well in the general case. The key problem is the large disparity of traffic volume across individual terminals. In this contribution we formulate the problem in terms of optimal statistical inference and provide a set of robust near-optimum estimators that are relatively simple to implement. We validate the proposed estimators with simulations in synthetic scenarios and provide results from a real operational 3G network. Complex network infrastructures like third-generation (3G) cellular networks are exposed to errors and attacks. Passive packet-level monitoring can play an important role in support- ing the operation and troubleshooting of such systems. Sum- mary statistics of network-layer performance variables (delay, RTT, loss, throughput etc.) can be extracted from passive probes and used as synthetic indicators for the network health status. Sudden deviations in the otherwise regular temporal behavior of such indicators might point to network problems. Change-point detection methods can be used to automatically identify anomalies and raise timely alarms. In this work we focus on the use of network-wide packet loss statistics obtained from the passive analysis of TCP packets at a single measurement point. The reference scenario consists of a passive monitoring probe located in the Core Network (ref. Fig. 1). The probe counts the number of TCP packets directed to each individual mobile terminal and, from the analysis of (un)acknowledged packets, infers the number of packet loss events. The problem is then to derive a synthetic indicator representative of the network-wide packet loss. The ultimate goal is to use such indicator to detect network problems and large-scale performance drifts. A common simple approach is to use as the summary indicator the ratio of loss events over the total transmitted packets across all the terminals. Another possible method is to compute the individual per-terminal loss ratios and take the arithmetic average. We show that such simplistic approaches do not always work well in practical scenarios. The key issue is the large disparity found in the distribution of the number of packets per terminal, which is due to the typically large disparity in traffic intensity of individual users: many users generate very low traffic, while a few ones generate very high volumes. For the two simplistic summary indicators introduced above, such disparity translates directly into an increased variance of the estimate. In this work we formalize the problem at hand as an optimal estimation problem. The optimal solution can be found via Bayesian approach (9), but at the cost of a relatively high resolution complexity which might hamper the practical adoption of such schemes. In this contribution we focus on the derivation of simple yet robust near-optimum estimators.