STTR Phase I: Commercial Grade Automatic and Manual Parallelization and Performance Tools

Interactive Supercomputing, Inc., Waltham MA

Investigators

Abstract

This Small Business Technology Transfer (STTR) Phase I research project aims to create a software development toolkit for semi-automatic parallelization of software written in high-level languages for serial architectures. The goal is to not only ensure high performance on parallel applications, but to do so while minimizing development time. Using a series of programs previously developed as a test-bed, the intent is to identify high payoff opportunities for semi-automatic parallelization. Then prototype a series of semi-automatic parallelization techniques and use them on the Northeastern application test-bed to show proof of concept. The chosen transformations will require the solution of complicated inferencing problems in dynamic computer languages. Novel parallelization tools such as a calculus of distributions and the lazy determination of parallel variables are among the techniques proposed. Techniques will range from vectorization to full blown mathematical operations on the data. Additional proposed innovations include novel expression templates that can grow with the software library. A rigorous statistical methodology for evaluating user performance gains will define success. These methods will be tested on the Star-P parallel computing platform thereby providing the first rigorous testing of such approaches in a mature parallel high-level language that is commonly available on the desktop. This research project addresses the needs of the millions of scientists and engineers who wish to access the speed and memory capabilities of high performance computers, but do not possess a computer science background. High performance computing is increasingly used in a diverse group of life science, engineering, physical science, and mathematical domains. The proposed toolkit allows applications developed in high-level languages for serial architectures to be easily transitioned to parallel high performance computing architectures. Given the growing importance of high performance computing, this research should have broad and deep impact in many scientific disciplines.