Accuracy of packet loss monitoring over networked CPE

In this paper we study the accuracy achievable when using probing to measure packet loss probability. Active monitoring using packet probing is widely used to measure network performance at the packet level, especially where direct measurements from network nodes is not possible. In this paper we develop an analytical solution that allows us to predict the number of samples (probes) needed to measure packet loss probability. This formula relates the number of probes needed to the networking scenario, the traffic characteristics, the target packet loss probability and the accuracy to which this loss probability must be resolved. We show that the number of probes required to accurately measure the packet loss probability is a highly sensitive function of the link bandwidth, the traffic load and burstiness, and the packet loss probability itself. We quantify how many probes are needed in certain typical scenarios involving either VoIP or generic data traffic models.