Program and storage transformations for improving memory performance

Louisiana State University, Baton Rouge LA

Investigators

Abstract

Jagannathan Ramanujam, Louisiana State University The performance of programs on modern processors depends critically on how their memory access characteristics can be matched to the multi-level memory hierarchy commonly used in these processor architectures. The goal of this project is derive compiler transformations to improve the memory performance of scientific computations. In particular, a combination of program restructuring and memory layout transformations of data will be derived to handle a larger class of programming constructs than perfect nests and regular memory accesses. This project will study several important problems, including: (a) strategies to integrate tiling and data shackling in order to effectively orchestrate the movement of data through memory hierarchies; (b) issues in the design of a sophisticated locality-enhancing compiler for regular and irregular codes; (c) extensive experimental evaluation of locality-enhancing transformations; (d) insights on the interaction between techniques for exploiting instruction-level parallelism and register-level reuse; and (e) possible insights on improvements in the design of memory systems for applications, including the design of application-specific cache architectures. Most importantly, these compiler techniques will allow users to easily exploit the enormous computation power in modern processor architectures.