Profiling Data-Dependence to Assist Parallelization: Framework, Scope, and Optimization

摘要: This paper describes a tool using one or more executions of sequential program to detect parallel portions the program. The tool, called Par wiz, uses dynamic binary instrumentation, targets various forms parallelism, and suggests distinct parallelization actions, ranging from simple directive tagging elaborate loop transformations. first part details link between program's static structures (like routines loops), memory accesses performed by program, dependencies that are used highlight potential parallelism. also instrumentation involved, general architecture system. second puts framework into action. study focuses on targeting OpenMP parallel-for directives, including privatization when necessary. is an adaptation well-known vectorization technique based slightly richer dependence description, where transformation. third views loops as graph (hopefully lightly) dependent iterations. explains how overall cost data-dependence profiling can be reduced. has two major causes: first, instrumenting slows down second, turning graphs consumes processing time. wiz analysis original (binary) provide data at coarser level, moving individual complete whenever possible, thereby reducing impact both sources inefficiency.