Synthesizing scalable computations from sequential programs

Advocates an approach that supports decomposition and scalable synthesis of a parallel computation. The decomposition is achieved with the aid of annotation languages that enable one top annotate programs written in various programming languages. The authors have implemented annotations for the Equational Programming Language (EPL). The synthesis is achieved with the aid of a simple configuration language that describes the computation in terms of interactions of programs and their fragments created by annotations. The decomposition and synthesis simplify the process of: (1) determining the grain size for efficient parallel processing, (2) data distribution, and (3) run-time optimization. The authors discuss annotations and configurations suitable for parallel programs written in EPL and FORTRAN and their use in scalable synthesis. They first discuss how annotations can define different computational blocks from a single program and how these blocks determine data distributions across processors. They outline a design of the configurator and show how FORTRAN programs can be configured into a hierarchical structure of computational blocks. An example of LU decomposition written in both EPL and FORTRAN illustrates the process of program decomposition and synthesis. The authors discuss code generation for synthesized computations, and some possible extensions.<>