Data Compression Techniques for Improving Memory Hierarchy Performance

University Of Arizona, Tucson AZ

Investigators

Abstract

Abstract: This project is aimed at developing data compression techniques for substantially reducing the data memory footprint of memory intensive applications. We will study the behavior of a wide range of applications to identify characteristics of data held in data structures by these programs that can be exploited in carrying out data compression. We will develop a suite of data compression transformations that will be implemented by the compiler. Our goal is to develop data transformations that can be applied to partially compressible data structures, that is, data structures in which all the data is not compressible. This approach will not only result in transformations that are widely applicable, but also lead to techniques for applying them that are efficient. In particular, instead of relying on complex compile-time analysis for proving their applicability, we will be able to use simple value profiling techniques to identify data structures that are mostly compressible and then apply the transformations to them. We will develop data compression extensions (DCX) to a RISC-style instruction set for efficiently manipulating compressed data. In addition, we will also address other low level code generation issues, such as impact of transformations on register pressure and instruction cache behavior, which can have a significant impact on performance.