Compilation techniques for multicomputers

Abstract

This thesis considers problems in process and data partitioning when compiling programs for distributed-memory parallel computers (or multicomputers). These partitions may be specified by the user through the use of language constructs, or automatically determined by the compiler. Data and process partitioning techniques are developed for two models of compilation. The first compilation model focusses on the loop nests present in a serial program. Executing the iterations of these loop nests in parallel accounts for a significant amount of the parallelism which can be exploited in these programs. The parallelism is exploited by applying a set of transformations to the loop nests. The iterations of the transformed loop nests are in a form which can be readily distributed amongst the processors of a multicomputer. The manner in which the arrays, referenced within these loop nests, are partitioned between the processors is determined by the distribution of the loop iterations. The second compilation model is based on the data parallel paradigm, in which operations are applied to many different data items collectively. High Performance Fortran is used as an example of this paradigm. Novel collective communication routines are developed, and are applied to provide the communication associated with the data partitions for both compilation models. Furthermore, it is shown that by using these routines the communication associated with partitioning data on a multicomputer is greatly simplified. These routines are developed as part of this thesis. The experimental context for this thesis is the development of a compiler for the Fujitsu AP1000 multicomputer. A prototype compiler is presented. Experimental results for a variety of applications are included. Show more