ITR: Resource Management Techniques for Internet Services on Large-Scale Clusters

University Of California-Santa Barbara, Santa Barbara CA

Investigators

Abstract

Web-based services delivered on scalable cluster platforms are playing an increasingly important role in today's society. A key aspect that has not been sufficiently addressed is delivering scalable performance with quality of service guarantees. It is all too common that web services suffer from outages due to unexpected demand. In the future, as even more computing devices (e.g., Sensor Networks) depend on services, ensuring scalability, reliability, fault-tolerance and quality of service will become paramount, and will allow our society to function normally. This proposal aims at solving the critical research necessary to achieve this vision. This project will study the problem of resource management for Internet services on large-scale clusters, and develop novel solutions that address the shortcomings of traditional approaches (i.e., poor resource utilization, high cost, and low flexibility). This work will develop models that express how clusters share the resources, and will devise metrics that capture the needs of Internet services. Performing extensive simulations using workloads of real web services, a variety of scheduling algorithms and techniques for managing resources will be analyzed. Results from these studies will be applied in building an actual cluster research prototype capable of delivering the desired service guarantees.