QoS measurement of Real Time Multimedia services on the Internet

Real-time applications such as IP telephony, Internet radio stations and video conferencing tools require certain levels of QoS (Quality of Service). Because the Internet is still a best-effort network, the QoS of these applications must be measured and monitored in order to provide feedbacks to applications for rate and/or error control, and to both end-users and service providers. A standardized objective measurement technology makes it possible to compare between service providers in a fair way. We address the problems of packet delay and loss measurement, since they are the major determining factor of multimedia quality. We first describe the problems and techniques in obtaining good measurement results. Then, we discuss the modeling and analysis of delay and loss. Our goal is to establish feasible metrics that can reliably predict perceived quality. We find that the extended Gilbert model (2-state being a special case) is a suitable loss model,and the inter-loss distance metric is useful in capturing the burstiness between loss runs. For delay, besides the autocorrelation metric, a conditional cumulative distribution function may be useful. We apply these models to some of our Internet packet traces,and find that losses are generally bursty, and that delays usually have a strong temporal dependency component. We also find that the final loss pattern after applying playout delay adjustment (and FEC if used) still corresponds well to the extended Gilbert model