Performance Modeling and System Management for Multi-component Online Services Christopher Stewart and Kai Shen Department of Computer Science, University of Rochester Many dynamic-content online services are comprised of multiple interacting components and data partitions distributed across server clusters. Understanding the performance of these services is crucial for efficient system management. This paper presents a profile-driven performance model for cluster-based multi-component online services. Our offline constructed application profiles characterize component resource needs and inter-component communications. With a given component placement strategy, the application profile can be used to predict system throughput and average response time for the online service. Our model differentiates remote invocations from fast-path calls between co-located components and we measure the network delay caused by blocking inter-component communications. Validation with two J2EE-based online applications show that our model can predict application performance with small errors (less than 13% for throughput and less than 14% for the average response time). We also explore how this performance model can be used to assist system management functions for multi-component online services, with case examinations on optimized component placement, capacity planning, and cost-effectiveness analysis.