Date Added: Mar 2012
This paper presents the initial steps toward a distributed system that can optimize its performance by learning to reconfigure CPU and memory resources in reaction to current workload. The authors present a learning framework that uses standard system-monitoring tools to identify preferable configurations and their quantitative performance effects. The framework requires no instrumentation of the middleware or of the operating system. Using results from an implementation of the TPC Benchmark W (TPC-W) online transaction-processing benchmark, they demonstrate a significant performance benefit to reconfiguration in response to workload changes.