Packet Loss Estimation Using Distributed Adaptive Sampling

Packet losses are a critical metric for network performance assessment. In this paper we present a novel methodology to accurately estimate the packet loss ratio in realtime in a fully distributed scenario. The constraint such systems must face is the large amount of resources required for keeping live performance assessment. Our contribution has two main parts. On the one hand we study the behaviour of packet losses among the traffic sharing a path, and extend the classical definition of loss burst by the concept of density. On the other hand, with the knowledge acquired in the loss distribution study, we present an adaptive sampling technique that schedules the network resources in order to distributely estimate the packet losses with reasonable accuracy. In order to validate the proposal we perform some real tests over an European-wide testbed. The results show a great improvement in packet loss estimation over previous research, while using a controlled amount of resources.