Evolution of the Multicore Adaptability of Scientific Software Systems

摘要

An empirical study that examines the challenges that scientific software systems developed in C/C++ have that prevent them from efficiently exploiting the full potential of the new multicore technology is presented. The study is conducted on 12 open source scientific systems comprising more than 5.4 million lines of code and containing over 84.5 thousand for- loop statements. Static analysis methods are applied to each system to determine the number of for-loops and free-loops (i.e., loops that can be parallelized). Additionally, each system is analyzed and the challenges and inhibitors to parallelization from a software engineering perspective are detected and presented. Some challenges towards adapting and re- engineering scientific software systems to better utilize modern multi-core architectures are determined including function side effects, data dependency and jumping statements. The results show that the most prevalent inhibitors are functions called within for-loops that have side effects, followed by data dependency as the second most prevalent inhibitor. These inhibitors pose the greatest roadblock to re-engineer and transform systems to better utilize parallelization. Results also show that data dependency has a more significant impact on scientific systems compared to general purpose systems that have been studied in previous studies. Historical data over a 5- year period of inhibitor counts for the set of systems studied is also presented. It shows that there is an insignificant change in the potential for parallelization of for-loops over time in general. The study suggests some software engineering techniques that have the potential to improve the parallelizability of scientific systems.