A Survey of Portability in Scientific Computing

Portability is a measure of the ease with which software can be moved between heterogeneous computing platforms. Software with a high degree of portability is valuable because it reaches more users. Commercial developers expend a considerable effort to make software that can run on multiple platforms. Scientific programmers have the same motivation to make their software portable, but their requirements differ in some respects from those of commercial software. Scientists have a unique need for collaboration and the independent replication of results at geographically diverse sites. Scientific algorithms tend to be expensive computationally, so scientific developers are more likely to use parallel computing. The fast pace of evolution for High-Performance Computing (HPC) systems affects scientific organizations inordinately, causing diversity in their computing platforms. These factors combine to make portability a critical issue for scientific computing. In addition, scientific institutions may also lack the resources which are available to commercial developers, and this can influence selection of a portability strategy. In this paper we propose criteria for evaluating the portability of a scientific application and we survey existing portability techniques with respect to those criteria. We identify the set of portability methods which are most applicable to scientific computing, and we identify future research problems.