A Model for Automatic Data Partitioning

In order to efficiently exploit global parallelism, it is essential to find a good way to distribute data among the processors in distributed-memory parallel computer systems. A formal technique utilizing augmented data access descriptors (ADADs) to determine this distribution is presented. This technique differs from previous approaclies in that it views the problem of finding a good distribution as an extension of data dependence analysis. The importance of this difference is demonstrated through an explanation of how ADADs facilitate interprocedural analysis, directed loop transformations, and incremental analysis, which may lead to improvements in the eficieiicy of both program developn~enta nd the program itself.