Restructuring Arrays for E cient Parallel Loop Execution

In a sequential program, data are often structured in a way that is optimized for a sequential execution. However, when the program is parallelized, the data access pattern may change drastically. If the structure of the data is not changed accordingly, parallel performance will su er. In this paper, we consider this problem in the context of runtime loop parallelization [8, 9], which is a general technique to parallelize loops not amenable to compile-time analysis. In a parallel execution of a loop, iterations may be performed in a very di erent order than in the sequential execution. This may result in undesirable cache e ects on distributed shared-memory multiprocessors, unless the structure of the arrays accessed by these iterations is changed accordingly. We discuss what these problems are and how they arise. We then describe two data restructuring techniques to address them: the restructuring of read-write arrays to reduce inter-processor communication due to false sharing, and the restructuring of read-only arrays to improve spatial locality. We also report experiments on a KSR1 [3] to evaluate the e ectiveness of these techniques and the preprocessing and postprocessing overheads they entail. The results show that the restructuring techniques can substantially improve performance of the parallelized loop. When restructuring overheads are ignored, we see a doubling of parallel speedups. While restructuring overheads can be quite signi cant, they can often be amortized across multiple loop executions so that they do not outweigh the performance bene ts. In our experiments, it takes only two loop executions to achieve this.