An evaluation of global address space languages

摘要: Co-array Fortran (CAF) and Unified Parallel C (UPC) are two emerging languages for single-program, multiple-data global address space programming. These boost programmer productivity by providing shared variables inter-process communication instead of message passing. However, the performance these still has room improvement. In this paper, we study variants NAS MG, CG, SP, BT benchmarks on several modern architectures to identify challenges that must be met deliver top performance. We compare CAF UPC programs with original Fortran+MPI code. Today, scalable clusters only when written use bulk communication. our experiments uncovered some significant bottlenecks codes all platforms. account root causes limiting such as synchronization model, efficiency strided data, source-to-source translation issues. show they can remedied language extensions, new constructs, and, finally, adequate optimizations back-end compilers.