A portable and customizable profiling framework for java based on bytecode instruction counting

Abstract

Prevailing profilers for Java, which rely on standard, native-code profiling interfaces, are not portable, give imprecise results due to serious measurement perturbation, and cause excessive overheads. In contrast, program transformations allow to generate reproducible profiles in a fully portable way with significantly less overhead. This paper presents a profiling framework that instruments Java programs at the bytecode level to build context-sensitive execution profiles at runtime. The profiling framework includes an exact profiler as well as a sampling profiler. User-defined profiling agents can be written in pure Java, too, in order to customize the runtime processing of profiling data. © Springer-Verlag Berlin Heidelberg 2005.