Productivity and Performance of the HPC Challenge Benchmarks with the XcalableMP PGAS Language

The present paper introduces designs of the XcalableMP PGAS language for improved productivity and performance of a high performance computing (HPC) system. The design of a unique XcalableMP programming model is based on both local-view and global-view models. This design allows programmers to easily develop HPC applications. Moreover, in order to tune HPC applications, XcalableMP provides inquiry functions for programmers to obtain local memory information of a global array. In the present paper, we evaluate the productivity and the performance of XcalableMP through implementations of the High-Performance Computing Challenge (HPCC) Benchmarks. We describe the implementations of three HPCC benchmarks, including RandomAccess, High-performance Linpack (HPL), and Fast Fourier Transform (FFT). In order to evaluate the performance of XcalableMP, we used the K computer, which is a leadership-class HPC system. As a result, we achieved 163 GUPS with RandomAccess (using 131,072 CPU cores), 543 TFlops with HPL (using 65,536 CPU cores), and 24 TFlops with FFT (using 262,144 CPU cores). These results reveal that XcalableMP has good performance on the K computer.