Predictive modeling of streaming servers

Abstract

In this paper, we describe our approach to deriving saturation models for streaming servers from vector-labeled training data. If a streaming server is driven into saturation by accepting too many clients, the quality of service degrades across the sessions. The actual saturating load on a streaming server depends on the detailed characteristics of the client requests: the content location (local disk or stream relay), the relative popularity, and the bit and packet rates [1]. Previous work in streaming-server models has used carefully selected, low-dimensional measurements, such as client jitter and rebuffering counts [2], or server memory usage [3]. In contrast, we collect 30 distinct low-level measures and 210 nonlinear derivative measures each second. This provides us with robustness against outliers, without reducing sensitivity or responsiveness to changes in load. Since the measurement dimensionality is so high, our approach requires the modeling and learning framework described in this paper.