CSR-AES: Collaborative Research: Behavior-Based Speculative Parallelization and Optimization on Desktop Multiprocessors

University Of Rochester, Rochester NY

Investigators

Abstract

The on-going shift in processor technology to favor multicore multiprocessors is opening new opportunities for software speculation, where program code is speculatively executed to improve speed at the cost of having to monitor for errors and the risk of having to re-execute code when an error happens. This project develops a new, behavior-based approach, which allows a user or a profiling tool to parallelize or optimize a program based on partial information about the program code and the input. It mainly develops two programming techniques: Behavior-oriented parallelization (BOP), which speculatively executes possibly parallel regions, and Fast track, which uses software speculation to support the use of unsafely optimized code. The exponential increase in microprocessor performance over the past 30 years has had incalculable impact on science, commerce, government, and quality of life. Continuing this revolution through the coming decade will depend on a large degree of processor-level parallelism. Behavior-based oftware speculation promises to improve the performance of existing, sequential software and of new software that reuses an existing code base. It simplifies parallelization and should thus improve the productivity of future software development. The outcome of this proposal will be a suite of techniques for general-purpose software, and new tools that will be transitioned to industrial partners wherever possible, and will be used in both undergraduate and graduate courses.