Performance evaluation of automatically generated data-parallel programs

The problem of evaluating the performance of parallel programs generated by data-parallel compilers is studied. These compilers take as input an application written in a sequential language augmented with data distribution directives and produce a parallel version, based on the specified partitioning of data. A methodology for evaluating the relationships existing among the program characteristics, the data distribution adopted, and the performance indices measured during the program execution is described. It consists of three phases: a "static" description of the program under study, a "dynamic" description, based on the measurement and the analysis of its execution on a real system, and the construction of a workload model, by using workload characterization techniques. Following such a methodology, decisions related to the selection of the data distribution to be adopted can be facilitated. The approach is exposed through the use of the Pandore environment, designed for the execution of sequential programs on distributed memory parallel computers. It is composed of a compiler, a runtime system and tools for trace and profile generation. The results of an experiment explaining the methodology are presented.