Analysis of Distributed Java Applications Using Dynamic Instrumentation

Although new Java virtual machines provide an API to obtain raw performance data, it is still the task of a skillful performance analysis tool to take all the strategic decisions for instrumentation and performance analysis of distributed Java programs. In this paper we demonstrate two new tools. Twilight and Aksum, which try to automatically instrument code regions, to determine what performance data to collect, to interpret performance data, and to relate the bottlenecks found back to source code. We present experiments with a widely distributed Java application running on a heterogeneous set of machines with different operating systems to demonstrate the efficacy of our tools