Parallel execution of LISP programs
Joseph S. Weening · 1990
This dissertation considers several issues in the execution of Lisp programs on shared-memory multiprocessors. An overview of constructs for explicit parallelism in Lisp is first presented. The problems of partitioning a program into processes and scheduling these processes are then described, and a number of methods for performing these are proposed. These include cutting off process creation based on properties of the computation tree of the program, and basing partitioning decisions on the state of the system at runtime instead of the program. An experimental study of these methods has been performed using a simulator for parallel Lisp. The simulator, written in Common Lisp using a continuation-passing style, is described in detail. This is followed by a description of the experiments that were performed and an analysis of the results. Two programs are used as illustrations--a Fast Fourier Transform, which has an abundance of parallelism, and the Cocke-Younger-Kasami parsing algorithm, for which good speedup is not as easy to obtain. The difficulty of using cutoff-based partitioning methods, and the differences between various scheduling methods, are shown. A combination of partitioning and scheduling methods which we call dynamic partitioning is analyzed in more detail. This method is based on examining the machine's runtime state; it requires that the programmer only identify parallelism in the program, without deciding which potential parallelism is actually useful. Several theorems are proved providing upper bounds on the amount of overhead produced by this method. We conclude that for programs whose computation trees have small height relative to their total size, dynamic partitioning can achieve asymptotically minimal overhead in the cost of process creation.