A Graph Representation for Automatic Dynamic Data Distribution with Control Flow Analysis

Data parallel programming languages, such as HPF, are the easiest way to program Distributed Memory Machines. These languages allow the programer to write code using global data references, but require the specification of data layout over the individual memories of the physical processors. The compiler will translate global data references into local and non-local references based on the data layout specified by the programmer. However the choice of a good data layout is not a trivial task, and can dramatically affect performance because of the non-uniformity of the memory system. This paper describes a graph representation used in a data distribution tool which automatically derives a static or dynamic data layout and parallelization strategy for the arrays used in a Fortran 77 program. All the information regarding data movement and parallelism is contained in a single data structure named Communication-ParallelismGraph (CPG). The tool takes into account control flow statements in t..