Program and storage transformations for improving memory performance

点击翻译按钮获取中文摘要

Jagannathan Ramanujam, Louisiana State UniversityThe performance of programs on modern processors depends critically onhow their memory access characteristics can be matched to themulti-level memory hierarchy commonly used in these processorarchitectures. The goal of this project is derive compilertransformations to improve the memory performance of scientificcomputations. In particular, a combination of program restructuringand memory layout transformations of data will be derived to handle alarger class of programming constructs than perfect nests and regularmemory accesses. This project will study several important problems,including: (a) strategies to integrate tiling and data shackling inorder to effectively orchestrate the movement of data through memoryhierarchies; (b) issues in the design of a sophisticatedlocality-enhancing compiler for regular and irregular codes; (c)extensive experimental evaluation of locality-enhancingtransformations; (d) insights on the interaction between techniquesfor exploiting instruction-level parallelism and register-level reuse;and (e) possible insights on improvements in the design of memorysystems for applications, including the design of application-specificcache architectures. Most importantly, these compiler techniques willallow users to easily exploit the enormous computation power in modernprocessor architectures.