Value - Based Distributions and Alignments in Fortran DReinhard

Compiling irregular applications written in a data-parallel language, like Fortran D or High Performance Fortran (HPF), presents a challenge of growing importance. A major component of data-parallel programming is the data mapping, i.e., data distribution and alignment. Map-pings may be either regular, in which case they correspond to simple functions between array indices and owning processors, or irregular. Regular mappings can already be speciied fairly conveniently, with keywords such as BLOCK or CYCLIC. However, the mechanisms proposed for irregular mappings so far require the programmer to go down to a very low level, for example by specifying explicit mapping arrays or mapping functions. For irregular problems, this makes it diicult to achieve good load balance or low storage and communication requirements without sacriicing programming convenience. Value-based mappings extend classical, index-based mappings to express for example the locality characteristics of an underlying physical problem. This gives the compiler an opportunity to improve both inter-and intra-processor locality, resulting in better performance and scalability, even when this locality is not inherent in the original data structures and access patterns. This paper reports on the experience gained from implementing value-based mappings in a Fortran 77D compiler prototype. We propose a natural extension of index-based mappings as already present in Fortran D and HPF. This paper addresses the compilation issues involved and makes a quantitative comparison of index and value-based mappings for a Molecular Dynamics application and an Unstructured Mesh kernel.