Carl Herder San Francisco State University
This presentation describes a white-box approach to workload characterization applied to Java workloads. We will investigate techniques for computing differences between individual Java benchmarks and describe two models: a model based on bytecode frequencies and execution times, and a model that uses only bytecode frequencies. These models are useful for decomposition and aggregation of Java workloads and for design and evaluation of Java benchmarks at various levels of granularity. Procedures for classifying benchmarks according to granularity and complexity are explained, and illustrated through application to the SPEC JVM98 benchmarks.