Comparative Study of Directive-based Programming Models on CPUs and GPUs for Scientific Applications

Abstract

Graphics processing unit-based parallel programming architecture has shown increased popularity as a building block for high performance computing. With the increased processing required in the last years, the search for devices with better performance, started in computing a need for parallel programming, makes it possible to support the performance of software and algorithms requiring high processing patterns. However, the major hurdle is their programming complexity for their widespread adoption. For better performance, many directive-based GPU programming models have been proposed. Directive-based programming model requires a different level of programming efforts to port and optimize applications. In this paper, two popular directive-based programming models have been evaluated using popular scientific benchmark applications. The benchmark applications are implemented using OpenACC for GPUs and OpenMP for CPUs. Results show that the directive-based programming model, OpenACC for GPUs, shows reasonably high performance, compared to directive-based programming model, OpenMP on CPUs.