The goal of this project is to investigate the performance of GPU programming models, across all of their implementations and for various GPU hardware architectures, to establish which ones offer the best portability with minimal loss of computational performance. In order to render the outcomes of our study widely applicable, this performance portability study will be conducted using the 13 Berkeley Dwarfs, which embody the most common and critical computation/communication idioms of high-performance scientific computations. The outcomes of this project are poised to change the landscape of GPU and, by extension, of HPC software development, by guiding and informing effective programming model choices for performance-critical applications — dramatically shortening time and expenses associated with their generation.
Experience in C/C++, at least one GPU programming model, OpenMP and HPC environment is required.
Experience in profiling code is desirable.
High-Performance Computing, GPU programming, Profiling, Performance Evaluation