Author retrospective: compilation techniques for block-cyclic distributions

Compilers for data-parallel languages use data distribution specifications to guide code generation for distributed-memory machines. Our 1994 paper described how to generate efficient code for programs that employ block-cyclic data distributions. In subsequent work at Rice University, Darte, Mellor-Crummey, Fowler, and Chavarría-Miranda developed a more general form of block-cyclic partitionings known as generalized multipartitionings. Generalized multipartitionings provide some additional balance properties that make them useful for parallelizing computations that solve recurrences along spatial dimensions. These partitionings were subsequently implemented in Rice University's dHPF compiler for High Performance Fortran. In the years since, the field has changed in many ways; however, variants of block-cyclic data distributions are still used today by modern parallel programming models, algorithms, and compilers. Original paper: http://dx.doi.org/10.1145/181181.181572