E ective Load Management for Scalable

Advances in storage technologies are making it feasible to design video servers that are capable of supporting continuous, real-time delivery of multiple video streams simultaneously. One of the major design considerations for a large-scale video server is scalability, to admit and service thousands of subscriber requests simultaneously. In this paper, we deene and formulate various policies that can be used for load management in a video server. We propose a predictive placement policy that determines the degree of replication necessary for selected popular video objects using a cost-based optimization procedure that is based on a priori predictions of expected subscriber requests. For scheduling subscriber requests, we propose an adaptive scheduling policy that enables the video server to admit and service a maximum number of subscriber requests simultaneously. By comparing the relative utilization of resources at diierent data sources, this policy determines an assignment of requests to replicas. Since storage space is one of the key resources in the video server, we also devise methods for dereplication of storage for video objects, depending upon change in video object popularities or changes in server usage patterns. Optimizations to exploit dynamic changes in server access are also discussed. We present performance evaluations to compare the eeectiveness of these load management policies for diierent video server conngurations. Our analysis indicates that a load management procedure that uses a judicious combination of the diierent policies out performs the other policies for most server conngurations.