Compression Techniques for Improved Bandwidth and Static Code Size in Computer Systems

Doktorsavhandling, 2008

Technological improvements in integrated circuits have for a long time allowed the performance of computer chips to grow exponentially, allowing for more and more advanced systems. Unfortunately, the I/O-pins connecting the processor core/cores with memory and other devices have not seen the same rate of improvement. In this thesis I study the opportunities and challenges with using compression to help solve some of the challenges that have emerged. These include reducing the bandwidth needed and the static code size. The first part of the thesis addresses data-link compression. By studying the data transferred between the last-level cache and main memory, several types of localities that can be used to compress the data are identified. Current state-of-the-art compression techniques are analyzed in the context of this categorization. Using this categorization, I show that it is possible to combine techniques that work on different types of locality into a more efficient compression algorithm. Moreover, I show how stateful compression schemes can be implemented in multi-node systems. This thesis also considers efficient program representation. In most programs, identical sequences of instructions often appear in several places, making dictionary based compression an efficient scheme for reducing the static code size. In this thesis I propose a new, more flexible scheme in which similar sequences of instructions can be represented by the same dictionary entry and executed with low hardware overhead. Finally, a code compression scheme targeting wide instruction formats is presented. The scheme is based on dynamic look-up tables, which allow the compiler to adapt the compression to different phases of the application. Also presented is a methodology for dimensioning the decompression engine and an algorithm for generating compressed programs that dynamically manage the look-up tables with little run-time performance overhead. The compression scheme is evaluated using FlexCore, an architecture with exposed datapath control, and shown to efficiently reduce the control-bits of the instruction word by more than 50%.