Automated benchmarking of Java APIs

Abstract

Performance is an extra-functional property of software systems which is often critical for achieving sufficient scalability or efficient resource utilisation. As many applications are built using application programmer interfaces (APIs) of execution platforms and external components, the performance of the used API implementations has a strong impact on the performance of the application itself. Yet the sheer size and complexity of today's APIs make it hard to manually benchmark them, while many semantical constraints and requirements (on method parameters, etc.) make it complicated to automate the creation of API benchmarks. Additionally, modern execution platforms such as the Java Virtual Machine perform extensive nondeterministic runtime optimisations, which need to be considered and quantified for realistic benchmarking. In this paper, we present an automated solution for benchmarking any large APIs that are written in the Java programming language, not just the Java Platform API. Our implementation induces the optimisations of the Just-In-Time compiler to obtain realistic benchmarking results. We evaluate the approach on a large subset of the Java Platform API exposed by the base libraries of the Java Virtual Machine.