Finding Good Configurations in High-Dimensional Spaces: Doing More with Less

Manually tuning tens to hundreds of configuration parameters in a complex software system like a database or an application server is an arduous task. Recent work has looked into automated approaches for recommending good configuration settings that adaptively search the full space of possible configurations. These approaches are based on conducting experiments where each experiment runs the system with a selected configuration to observe the resulting performance. Experiments can be time-consuming and expensive, so only a limited number of experiments can be done even in systems with hundreds of configuration parameters. In this paper, we consider the problem of finding good configurations under the two constraints of high dimensionality (i.e., many parameters) and few experiments. We show how certain design decisions made in previous algorithms for finding good configurations make them perform poorly in this setting. We propose a new algorithm called MOWILE (MOre WIth LEss) that addresses these limitations, and outperforms previous algorithms by large margins as the number of parameters increase. Our empirical evaluation gives interesting insights that will benefit system administrators who apply experiment-driven approaches for configuration tuning.