Designing and developing tools to automatically identify parallelism

TL;DR

This paper introduces a dynamic analysis tool and an abstract method for identifying parallelism in program execution graphs, aiming to improve automatic parallelization techniques.

Contribution

It presents a novel dynamic analysis tool and a quotient graph approach for analyzing and detecting parallelism in code execution graphs.

Findings

Effective identification of parallel regions in execution graphs

Successful application to four computational dwarfs

Potential for enhancing automatic parallelization tools

Abstract

In this work we present a dynamic analysis tool for analyzing regions of code and how those regions depend between each other via data dependencies encountered during the execution of the program. We also present an abstract method to analyze and study parallelism in a directed graph, by studying a Quotient Graph of the execution graph of a program, and give a simple algorithm for searching parallelism in execution graphs with a high degree of symmetry. Finally, we evaluate our approach selecting four dwarfs out of 13 Berkeleys computational dwarfs or otherwise known as parallel patterns.