ITR: Dynamic Cooperative Performance Optimization

University Of Massachusetts Amherst, Amherst MA

Investigators

Abstract

Object-oriented programming languages, notably Java, are gaining broad use because of their benefits, which come largely from their flexibility. But this same flexibility makes Java programs more difficult to optimize in advance. One sample collection of programs, optimized in advance, spent 40-95% of their time on an aggressive processor waiting for the memory to provide data, illustrating the need to improve memory behavior. Future processors will only make the problem worse. The project makes an integrated attack on this problem, incorporating new program analyses and optimizations, profile feedback, run-time techniques including adaptive garbage collection algorithms, and ways of communicating high-level predictions and observations of program behavior to the hardware. The project aims to design and build a compiler, run-time system, and enhanced architectural and operating system features that react quickly and gracefully to compiler predictions and actual run-time behavior to achieve high performance. The goal is synergy via cooperation between the system components versus solving each problem within an individual component. While the research focuses on improving memory performance, the project's envisioned framework is suited to a wide range of performance optimization techniques, so the expected research results and software products have broader impact.