Provisioning and analysis of application-specific Internet measurements

Real-time applications such as videoconferencing and Grid computing are being widely deployed on the Internet. To effectively meet the Quality of Experience (QoE) expectations of end-users, Internet Service Providers (ISPs) are increasingly relying on automated frameworks. These frameworks use Internet measurements pertaining to Quality of Service (QoS) metrics (e.g., delay and loss) and estimation techniques such as forecasting for initial selection and dynamic adaptation of resources. The degree of success achievable by these frameworks is being limited by the inability of ISPs to suitably provision and analyze Internet measurements. In this dissertation, measurement scheduling techniques are presented that provision QoS metrics on the Internet with the accuracy, regulation, frequency and flexibility desired by the frameworks. Also, a tool is presented that analyzes provisioned QoS metrics to estimate end-user QoE in real-time without involving actual end-users, a key requirement for framework automation.;Provisioning QoS metrics using measurement scheduling involves careful orchestration of active measurement tools. Since these tools consume non-negligible amount of network resources, it is essential to schedule them in a regulated and non-conflicting manner. In addition, QoS metrics need to be sampled with strict periodicity for accurate network weather forecasting. To address these measurement requirements, offline/online measurement scheduling algorithms are described. Simulations show that our algorithms can improve the schedulable utilization of offline measurements by 300% and the response time of online measurements by 50%. To evaluate their utility for performance forecasting, the algorithms are implemented in the popular Network Weather Service (NWS) framework. From the implementation experience, it is shown that no measurement scheduling algorithm that avoids measurement conflicts can cater strict periodicity. However, it is demonstrated how a piece-wise linear interpolation technique used with the algorithms can improve the forecast accuracy by reducing the mean square error by half. Lastly, an active measurement tool called Vperf is developed that uses provisioned network QoS measurements to estimate real-time QoE of applications, specifically involving Voice and Video over IP (VVoIP). Vperf uses a novel psycho-acoustic/visual cognitive model called GAP-Model. Validation experiments show that online Vperf measurements have matching performance when compared with time-consuming offline Peak Signal-to-Noise-Ratio (PSNR) measurements.