Early Evaluation of Directive-Based GPU Programming Models for Productive Exascale Computing

TitleEarly Evaluation of Directive-Based GPU Programming Models for Productive Exascale Computing
Publication TypeConference Paper
Year of Publication2012
AuthorsLee, Seyong, and Vetter Jeffrey S.
Conference NameSC12: ACM/IEEE International Conference for High Performance Computing, Networking, Storage, and Analysis
Date Published11/2012
PublisherIEEE press
Conference LocationSalt Lake City, Utah, USA
ISBN Number978-1-4673-0804-5
KeywordsDirective, GPU, programming model

Graphics Processing Unit (GPU)-based parallel computing systems have increased popularity as a promising building block for high performance computing, even possibly for future Exascale computing. However, their programming complexity remains as the biggest hurdle for developers. To provide better abstractions on the GPU computing, several directive-based GPU programming models have been proposed by both industry and academia. The directive-based models provide different levels of abstraction, and programming effort required to conform to their models and optimize the performance also vary. Understanding the differences in these new models will give valuable insights on the general applicability and performance of the directive-based programming models for both the current GPUs and future Exascale systems. This paper evaluates the existing directive-based GPU programming models by porting thirteen applications from various scientific domains into CUDA GPUs, which allows to identify important issues on the functionality, scalability, tunability, and debuggability of the existing directive-based approaches.