NGS: Resource-Aware Off-line and On-line Empirical Optimization

University Of Southern California, Los Angeles CA

Investigators

Abstract

EIA-0204040 Mary Hall University of Southern California Resource-Aware Off-line and On-line Empirical Optimization The research we propose herein explores a new paradigm that fundamentally restructures the compilation process to enable resource-aware optimization. We will demonstrate how this system can alleviate some of the problems that lead to inefficiencies in big science and engineering codes today: register pressure, cache conflict misses, the trade-off between synchronization, parallelism and locality in SMPs, and process migration in computational grids. There are three essential components of this research program: Compiler-directed off-line empirical optimization; Dynamic feedback, for user-and compiler-directed on-line optimization; and Experience base to guide optimization from history: The results of experiments from both off-line and on-line empirical optimization will be maintained in an experience base, which can be used to generalize prior optimization results and plan a future optimization strategy.