On using Aksum for semi-automatically searching of performance problems in parallel and distributed programs

There has been an increasing research interest in formalizing and automating the search for performance problems. In previous work we have introduced Aksum, which tries to automatically locate all performance problems in parallel and distributed applications based on multi-experiment performance analysis and user-provided machine and problem sizes. In this paper we report on experiences with Aksum for performance analysis of three realistic message passing, shared memory and mixed parallelism codes taken from laser physics, material science, and financial modeling. Aksum has been used to automatically decide which code regions must be instrumented, what performance information must be collected, and to launch experiments and test performance hypotheses as soon as more performance data becomes available. Multiple-experiment performance analysis enables the user to compare the performance behavior across a range of problem and machine sizes.