Reproducing Software Experiments

Abstract

Computational science software experiments are hard to reproduce because external data sets could have changed, software used in the original experiment cannot be reconstructed, or the input parameters for an experiment may not be documented. We have developed a set of tools called Rex to aid in reproducing software experiments. They enable one to record an experiment and archive its apparatus, replay experiments, run new experiments on a recorded apparatus, and compare two recorded experiments. Rex can handle sequential, multiprocess, and multithreaded programs. It does not require any modification to applications or the operating system on which they execute. The implementation of the Rex tools is based on being able to trap and compare the system calls made by an experiment. This dissertation discusses the challenges in reproducing software experiments and describes Rex's design and implementation. It also evaluates the execution and space overhead of the Rex tools.