Calibration and Prediction of Streaming-Server Performance

Streaming media is gaining in popularity for viewing both, video-on-demand content as well as live Webcasts. Streaming servers must meet strict data-delivery timing constraints in order to provide acceptable viewing quality. These constraints can be achieved only if the servers are not allowed to exceed their operational saturation point. At the same time, providers of streaming services need to maximize the use of their infrastructure to remain cost-effective. These competing goals motivate development of detailed models that predict server saturation points under extremely diverse workloads. Due to the intricate effects of distinct usage patterns on low-level measurements, no single server-side or client-side metric can adequately predict saturation for a non-controlled mixture of workloads. Furthermore, the dynamically changing nature of streaming workloads render simple linear statistics inadequate. Instead, we propose a methodology that can build predictive models using a relatively small number of calibration workloads. These models include both serverand client-side metrics and are accurate in predicting server performance, not only for the calibration workloads but also for arbitrary mixtures. We contend that the strength of our approach to modeling streaming-server behavior is its highly data-driven nature. The same calibration regime and modeling method are shown to be applicable to different streaming servers and across the wide variety of workloads seen in today’s environments.