PhD

Abstract

This thesis investigates the suitability of using functional programming for building parallel rule-based systems.

A functional version of the well known rule-based system OPS5 was implemented, and there is a discussion on the suitability of functional languages for both building compilers and manipulating state. Functional languages can be used to build compilers that reflect the structure of the original grammar of a language and are, therefore, very suitable. Particular attention is paid to the state requirements and the state manipulation structures of applications such as a rule-based system because, traditionally, functional languages have been considered unable to manipulate state.

From the implementation work, issues have arisen that are important for functional programming as a whole. They are in the areas of algorithms and data structures and development environments. There is a more general discussion of state and state manipulation in functional programs and how theoretical work, such as monads, can be used. Techniques for how descriptions of graph algorithms may be interpreted more abstractly to build functional graph algorithms are presented. Beyond the scope of programming, there are issues relating both to the functional language interaction with the operating system and to tools, such as debugging and measurement tools, which help programmers write efficient programs. In both of these areas functional systems are lacking.

To address the complete lack of measurement tools for functional languages, a profiling technique was designed which can accurately measure the number of calls to a function , the time spent in a function, and the amount of heap space used by a function. From this design, a profiler was developed for higher-order, lazy, functional languages which allows the programmer to measure and verify the behaviour of a program. This profiling technique is designed primarily for application programmers rather than functional language implementors, and the results presented by the profiler directly reflect the lexical scope of the original program rather than some run-time representation.

Finally, there is a discussion of generally available techniques for parallelizing functional programs in order that they may execute on a parallel machine. The techniques which are easier for the parallel systems builder to implement are shown to be least suitable for large functional applications. Those techniques that best suit functional programmers are not yet generally available and usable.