NPB UA benchmark scaling to thousands of Blue Gene/P cores using PGAS-like OpenMP extension

This paper demonstrates a parallel implementation of the Unstructured Adaptive Benchmark from the NASA Parallel Benchmark collection in the PGAS paradigm extending the OpenMP paradigm for distributed memory supercomputers. Some details of the modified OpenMP benchmark version generalized to the PGAS variant are discussed. The PGAS-like SHMEM library implementation for IBM Blue Gene/P through the DCMF messaging layer is considered. A number of results obtained on up to 4096 cores of the Blue Gene/P system installed at the Moscow State University are analyzed.