Empirical Performance Models for Java Workloads

Abstract

Java is widely deployed on a variety of processor architectures. Consequently, an understanding of microarchitecture level Java performance is critical to optimize current systems and to aid design and development of future processor architectures for Java. Although this is facilitated by a rich set of processor performance counters featured on several contemporary processors, complex processor microarchitecture structures and their interactions make it difficult to relate observed events to overall performance. This, coupled with the complexities associated with running Java over a virtual machine, further aggravates the situation. This paper explores and evaluates the effectiveness of empirical modeling for Java workloads. Our models use statistical regression techniques to relate overall Java system performance to various observed microarchitecture events and their interactions. Multivariate adaptive regression splines effectively capture non-linear and non-monotonic associations between the response and predictor variables. Our models are interpretable, easy to construct and exhibit high correlation/low errors between predicted and measured performance. Furthermore, empirical models afford additional insights into the characteristics of Java performance and the use of statistical techniques throughout this study allow us to assign confidence levels to our estimates of performance.