high performance computing on graphics processing units: hgpu.org

As technology advances, computing resources also gain benefits in many aspects: larger capacity, increased capability as well as rapidity. However, with heterogeneously distributed resources in Grid computing environment, the development an application to fully utilize the resources is a challenge. Especially, the computing resources themselves regularly upgrade their computing power for example by recruiting General Purpose Graphics Processing Unit (GPGPU) resources. The challenge in developing an application on computing environment like that becomes even greater. In this paper, we propose an approach to unify the programming paradigms in Grid computing and GPGPU computing as well as further our investigation on the performance of an application developed on such environment. To maximize its efficiency, the grid application is developed based on multi-level parallelism together with multi-level topology-aware techniques and the Global Arrays toolkit. We have successfully implemented the grid application with the proposed approach and the performance of the application depends directly on how the computing loads are distributed over those resources. The direct portability of a GPGPU application/module in order to be integrated into a comprehensive grid computing code is also observed in our approach.