Using Statistical Models for Embedded Java Performance Analysis
This paper proposes and evaluates rigorous statistical regression techniques for Java performance analysis on a high-end embedded processor (MPC7447A). The models relate overall Java system performance to various microarchitecture metrics and their interactions. The authors show that the models they develop in this paper are easy to construct, are interpretable and have high prediction accuracies. Java applications are prevalent on a wide variety of processors, many of which provide on chip Performance Monitoring Units (PMU) to measure and track system performance. Due to additional complexities associated with running Java applications on a Virtual Machine (JVM), it is necessary to clearly understand the influence of the various factors that affect performance.