Architecture-independent scientific programming in data parallel C: three case studies

Dataparallel C is a SIMD programming language suitable for solving many problems arising in science and engineering. We have developed compilers that translate Dataparallel C programs into C code that can be compiled and executed on multicomputers and multiprocessors. We describe the Dataparallel C programming language, the implementation of three scientific programs in Dataparallel C, and the performance of the compiled programs on three MIMD computers: the Intel iPSC/2, the nCUBE 3200, and the Sequent Symmetry S81. These case studies-an ocean circulation model, the "shallow-water" atmospheric model, and the Sharks World model-illustrate that nontrivial SIMD programs can achieve good speedup on MIMD computers and still be portable.