Programmability and Performance of New Global-View Programming API for Multi-Node and Multi-Core Processing

Various partitioned global address space (PGAS) languages capable of providing global-view programming environments on multi-node computer systems have been proposed to improve programming productivity in high-performance computing. However, several PGAS languages often require a detailed description of the remote data access, similar to descriptions used in message passing interface one-sided communications. Some PGAS languages have limitations pertaining to remote data access and recommend their local-view programming models, rather than the global-view ones, due to performance-related reasons. In this study, we propose SMint, which is an application programming interface that provides a global-view programming model with a software distributed shared memory mSMS as the runtime. Using stencil computation as a typical processing method, the performance and programmability of SMint have been compared with those of XcalableMP and Unified Parallel C, which are well-known examples of PGAS languages based on the C language. It was found that SMint achieved the best performance under the ideal global-view programming model.