An empirical study of data decomposition for software parallelization

Context: Multi-core architectures are becoming increasingly ubiquitous and software professionals are seeking to leverage the capabilities of distributed-memory architectures. The process of parallelizing software applications can be very tedious and error-prone, in particular the task of data decomposition. Empirical studies investigating the complexity of data decomposition and communication are lacking. Objective: Our objective is threefold: (i) to gain an empirical-based understanding of the task of data decomposition as part of the parallelization of software applications; (ii) to identify key requirements for tools to assist developers in this task, and (iii) assess the current state-of-the-art. Methods: Our empirical investigation employed a multi-method approach, using an interview study, participant-observer case study, focus group study, and a sample survey. The empirical investigation involved collaborations with three industry partners: IBM's High Performance Computing Center, the Irish Centre for High-End Computing (ICHEC), and JBA Consulting. Results: This article presents data decomposition as one of the most prevalent tasks of parallelizing applications for multi-core architectures. Based on our studies, we identify ten key requirements for tool support to help HPC developers in this area. Our evaluation of the state-of-the-art shows that none of the extant tool support implements all 10 requirements. Conclusion: While there is a considerable body of research in the area of HPC, a few empirical studies exist which explicitly focus on the challenges faced by practitioners in this area; this research aims to address this gap. The empirical studies in this article provide insights that may help researchers and tool vendors to better understand the needs of parallel programmers.