Passive Estimation of Quality of Experience Denis Collange, Jean-Laurent Costeux (Orange Labs, Sophia Antipolis, France {denis.collange, jeanlaurent.costeux}@orange-ftgroup.com) Abstract: Quality of Experience (QoE) is a promising method to take into account the users' needs in designing, monitoring and managing networks. However, there is a challenge in finding a quick and simple way to estimate the QoE due to the diversity of needs, habits and customs. We propose a new empirical method to approximate it automatically from passive network measurements and we compare its pros and cons with usual techniques. We apply it, as an example, on ADSL traffic traces to estimate the QoE dependence on the loss rate for the most used applications. We analyze more precisely the correlations between packet losses and some traffic characteristics of TCP connections, the duration, the sizes and the inter-arrival. We define different thresholds on the loss rate for network management. And we propose a notion of sensitiveness to compare these correlations on different applications. Keywords: Quality of experience, passive performance measurements, internet performance, end-to-end performance, user behaviour, traffic analysis Categories: C.2.3 - Network Operations, C.4 – Performance of Systems 1 Introduction More and more new applications arise in the Internet. Some of them may be used by relatively large communities. These applications may have very different traffic characteristics and performance requirements. Internet users show then a wide range of behaviours and needs depending on the applications they use, and how they use them. These users' requirements depend also on their access mode, their terminal, etc. Some users may "suffer" on a performance level satisfying others. An Internet Service Provider (ISP) must meet these customers' needs despite these heterogeneities. So the ISP cannot rely only on the classical network level performance criteria, such as the load and the loss rate on routers' interfaces, to manage its resources. These classical criteria only ensure that the network resources are not congested. They give very few insights on the end-to-end performance, and even less on customers' satisfaction. So a new criterion has been proposed, namely the Quality of Experience or QoE [Nokia 04] [Fiedler et al. 04]. Its aim is to characterize and to measure how the users perceive the network performance. QoE is a subjective criterion, but it must be correlated with technical network parameters managed by the network operator. It may depend on many parameters, for example the availability, the security level etc. This paper focuses on the correlations of QoE with end-to-end network level performance criteria. These correlations may be modelled by network utility functions [Fiedler et al. 05]. The challenge is then to find these correlations between the subjective QoE and the technical Quality of Service (QoS). Indeed, an ISP usually designs, monitors and manages its resources in order to respect some network level performance criteria. To be sure that its customers experience a good Journal of Universal Computer Science, vol. 14, no. 5 (2008), 625-641 submitted: 1/11/07, accepted: 28/2/08, appeared: 1/3/08 © J.UCS