A subspace optimizing data parallel complier

Abstract

Scientific programs with large data sets are an important class of computer application, requiring large amounts of memory and computational power. Massively parallel processing hardware and data parallel programming techniques are increasingly used to meet these requirements. A new approach to data parallel compilation, the Subspace compilation model, is introduced. This model is based on the idea that the shapes of data objects and how these shapes change represent higher-level performance considerations that the alignment of individual data elements. This model also removes the ad hoc restrictions of the prevalent Single Program, Multiple Data (SPMD) model. A Subspace compiler is designed based on the Subspace model, employing subspace trees as its program representation. A significant part of this compiler is implemented, with the CM-5 CM Fortran platform as its target. The implementation is tested on benchmark code, and the results discussed.