FAILURE PREDICTION IN VIDEO-STREAMING SERVERS THROUGH PERFORMANCE ANALYSIS OF SERVER AND CLIENT-SERVER INTERACTIONS

摘要

Video-streaming services are nowadays one of the biggest contributors to Internet traffic. Due to the real time characteristic of video-streaming, very low failure de tection and recovery delays are required. Proactive recov ery in face of performance degradation is a prominent area to explore. Current performance analysis work in video streaming focuses mostly on capacity planning and session admission through complex workload and resource model ing. These models have limited application when degrada tion is not explained by client workload (e.g., dynamic re source reallocation, software faults and misconfiguration). We explore server-side monitoring of performance degradations in video-streaming servers, based on statis tical analysis of server metrics and client-server interaction messages. Statistical analysis of event logs show that an alyzed metrics can be used as symptoms of failures to an ticipate them and thus enabling proactive recovery. Excep tion is server overloading caused by streaming of unpopular videos, which are exposed by metrics when QoS degrada tion is close to accepted quality thresholds.