Load Characterization, Overload Prediction and Load Anomaly Detection for Voice over IP Traffic

We characterize traffic load and provide a method for overload prediction, more generally load anomaly detection, in segments of IP networks that carry almost exclusively voice traffic. The fact that such network segments exist emerges from the architectures of many IP-based networks carrying voice. The need for load anomaly detection arises, for example, when there is a focused overload, e.g., a large number of admitted calls try to reach the same destination number, or when a segment of the network fails and voice traffic either overflows into a normal segment of the network or migrates away from it. To detect such load anomalies, we show it is sufficient to estimate the offered load measured over relatively long aggregation intervals such as 5 minutes. This is done by estimating the offered load in an M/M/inf system, e.g., a trunk group in which calls arrive according to a Poisson process and have exponentially distributed holding times, using a circuit-switch analogy.