Transformation of functional programs for identification of parallel skeletons

Hardware is becoming increasingly parallel. Thus, it is essential to identify and exploit inherent parallelism in a given program to effectively utilise the computing power available. However, parallel programming is tedious and error-prone when done by hand, and is very difficult for a compiler to do automatically to the desired level. One possible approach to parallel programming is to use transformation techniques to automatically identify and explicitly specify parallel computations in a given program using parallelisable algorithmic skeletons. Current existing methods for systematic derivation of parallel programs or parallel skeleton identification allow automation. However, they place constraints on the programs to which they are applicable, require manual derivation of operators with specific properties for parallel execution, or allow the use of inefficient intermediate data structures in the parallel programs. In this thesis, we present a program transformation method that addresses these issues and has the following attributes: (1) Reduces the number of inefficient data structures used in the parallel program; (2) Transforms a program into a form that is more suited to identifying parallel skeletons; (3) Automatically identifies skeletons that can be efficiently executed using their parallel implementations. Our transformation method does not place restrictions on the program to be parallelised, and allows automatic verification of skeleton operator properties to allow parallel execution. To evaluate the performance of our transformation method, we use a set of benchmark programs. The parallel version of each program produced by our method is compared with other versions of the program, including parallel versions that are derived by hand. Consequently, we have been able to evaluate the strengths and weaknesses of the proposed transformation method. The results demonstrate improvements in the efficiency of parallel programs produced in some examples, and also highlight the role of some intermediate data structures required for parallelisation in other examples