CAREER: Compilation Techniques for Customizing Software Libraries

University Of Texas At Austin, Austin TX

Investigators

Abstract

This project presents a vision for satisfying the world's demand for software: increase programmer productivity by building programs out of large software components that can provide powerful abstractions, that can be seamlessly integrated with other components, and that can easily conform to different execution environments. Thousands of software libraries already exist to provide such components, but most fall far short of the goal of seamless efficiency on all platforms. This project will define and experimentally evaluate a new approach for tuning libraries using both domain-specific and machine-specific compiler optimizations. The key to the approach is an annotation language that conveys three types of information: library semantics, machine characteristics, and information about how library components will be composed. This advances other library techniques by encapsulating and exploiting domain-specific information to reap benefits of programming abstractions without incurring their usual efficiency costs. The project will focus on three important opportunities for improving performance and portability: the extension of conventional scalar optimizations to library optimizations, the optimization of sequences of library calls, and methods for customizing libraries on different machines. The methods and results will be validated in experiments using the PLAPACK dense linear algebra library, MPICH message-passing library, and object-oriented languages such as Java. Research results expected from this project include a deeper understanding of the use of domain- and machine-specific information in the compilation process, new optimizations based on that information, and a compiler and language that can be used in testing the ideas. Just as this research seeks to improve the flow of information across interfaces, the project's educational component will attempt to bridge interfaces across disciplines and communities of the student population by creating a "Fundamentals of Computer Science" course for the institution's interdisciplinary Computational and Applied Mathematics program.