Parallel programming using functional languages

It has been argued for many years that functional programs are well suited to parallel evaluation. This thesis investigates this claim from a programming perspective; that is, it investigates parallel programming using functional languages. The approach taken has been to determine the minimum programming which is necessary in order to write efficient parallel programs. This has been attempted without the aid of clever compile-time analyses. It is argued that parallel evaluation should be explicitly expressed, by the programmer, in programs. To do achieve this a lazy functional language is extended with parallel and sequential combinators. The mathematical nature of functional languages means that programs can be formally derived by program transformation. To date, most work on program derivation has concerned sequential programs. In this thesis Squigol has been used to derive three parallel algorithms. Squigol is a functional calculus from program derivation, which is becoming increasingly popular. It is shown that some aspects of Squigol are suitable for parallel program derivation, while others aspects are specifically orientated towards sequential algorithm derivation. In order to write efficient parallel programs, parallelism must be controlled. Parallelism must be controlled in order to limit storage usage, the number of tasks and the minimum size of tasks. In particular over-eager evaluation or generating excessive numbers of tasks can consume too much storage. Also, tasks can be too small to be worth evaluating in parallel. Several program techniques for parallelism control were tried. These were compared with a run-time system heuristic for parallelism control. It was discovered that the best control was effected by a combination of run-time system and programmer control of parallelism. One of the problems with parallel programming using functional languages is that non-deterministic algorithms cannot be expressed. A bag (multiset) data type is proposed to allow a limited form of non-determinism to be expressed. Bags can be given a non-deterministic parallel implementation. However, providing the operations used to combine bag elements are associative and commutative, the result of bag operations will be deterministic. The onus is on the programmer to prove this, but usually this is not difficult. Also bags' insensitivity to ordering means that more transformations are directly applicable than if, say, lists were used instead. It is necessary to be able to reason about and measure the performance of parallel programs. For example, sometimes algorithms which seem intuitively to be good parallel ones, are not. For some higher order functions it is posible to devise parameterised formulae describing their performance. This is done for divide and conquer functions, which enables constraints to be formulated which guarantee that they have a good performance. Pipelined parallelism is difficult to analyse. Therefore a formal semantics for calculating the performance of pipelined programs is devised. This is used to analyse the performance of a pipelined Quicksort. By treating the performance semantics as a set of transformation rules, the simulation of parallel programs may be achieved by transforming programs. Some parallel programs perform poorly due to programming errors. A pragmatic method of debugging such programming errors is illustrated by some examples