Experience in the automatic parallelization of four Perfect-Benchmark programs

Abstract

Abstract.This paper discusses the techniques used to hand-parallelize, for the Alliant FX/80, four Fortran programs from the Perfect-Benchmark suite. The paper also includes the execution times of the programs before and after the transformations. The four programs considered here were not effectively parallelized by the automatic translators available to the authors. However, most of the techniques used for hand parallelization, and perhaps all of them, have wide applicability and can be incorporated into existing translators.

An Overview of Symbolic Analysis Techniques Needed for the Effective Parallelization of the Perfect Benchmarks

“…Other work has shown that current commercial parallelizing compilers perform poorly on real codes 4], and that a compiler can theoretically achieve good speedups for these codes 12,11]. Motivated by this, we examined the Perfect Benchmarks R to determine what symbolic analysis techniques are required to get the observed good speedups.…”

“…We h a ve quanti ed and analyzed this situation in previous reports 4]. There have also been studies of potential improvements by manually parallelizing real programs and reporting the necessary transformation techniques 12,11]. Table 1 shows some of these results.…”

A scalable method for run-time loop parallelization

“…(In contrast, privatization needs only to recognize privatizable variables by performing data dependence analysis, i.e., it is contingent only on the access pattern and not on the operations.) Since parallel methods are known for performing reductions operations (see, e.g., [12,17,18,20,44]), the difficulty encountered by compilers arises from recognizing the reduction statements. So far this problem has been handled at compile-time by syntactically pattern matching the loop statements with a template of a generic reduction, and then performing a data dependence analysis of the variable under scrutiny to guarantee that it is not used anywhere else in the loop except in the reduction statement [44].…”

Polaris: Improving the effectiveness of parallelizing compilers

“…P arallelizing compilers have been playing an important role in this quest. The present project has its early roots in a compiler evaluation e ort of the late 80s, where we h a ve found that despite the success on kernel benchmarks, available compilers were not very e ective on large programs EHLP91,BE92]. New measurements on a representative set of real programs were made possible, thanks to the Perfect Benchmarks R e ort, which w as initiated by CSRD, with participation from many other institutions BCK + 89].…”

“…Based on these observations, we h a ve hand parallelized the program suite as a major new approach to identifying e ective program transformations EHLP91,EHJP92]. As a result we h a ve found that not only can real applications be parallelized e ectively, but the transformations can also be automated in a parallelizing compiler.…”