Performance of a high-level parallel programming layer defined on top of the Ada tasking model

Abstract

In earlier work, we proposed extending Ada with a high-level parallel programming layer consisting of three pragmas that indicate where standard sequential constructs can be executed in parallel without changing the sequential semantics of a program. The three pragmas specify: {i) parallelizable sequences of statements, (ii) parallelizable for-loop statements, and (iii) single-assignment types for synchronization. A precise definition of these pragmas was given by means of transformations to standard Ada tasking constructs. In this paper, we improve our definition of the parallel programming layer and demonstrate that the pragmas can be implemented efficiently on top of an existing Ada compiler by straightforward preprocessing using the defining transformations. We present three simple parallel programs that exhibit good speedups over sequential programs, when executed on a multiprocessor computer. Development of these programs was considerably easier than the direct use of Ada tasking, because the parallel programs are identical to equivalent sequential programs except for the added pragmas.