Is SIMD enough for scientific and engineering applications on massively parallel computers?

Some basic issues involved in matching an application to a distributed memory parallel machine are addressed. In particular, data communication and processor synchronization as they relate to MIMD (multiple instruction-multiple data) and SIMD (single-instruction-multiple data) architectures are discussed. To illustrate the differences, the authors describe the implementation and performance of several engineering and scientific applications that have been coded for both kinds of machines. They find that many problems are well suited to both architectures. However, when the natural parallelism in a simulation requires a loose synchronization between processors, the MIMD paradigm offers a greater programming flexibility than SIMD.<>