Compression Techniques for Code Size and Data Bandwidth Reduction

Licentiate thesis, 2006

A challenge in the design of high performance computer systems is how to transfer data efficiently between main memory and the faster and smaller memory located on the processor chip. Main memory holds both the program to be executed and the data it needs to perform its tasks. This thesis focuses on how compression techniques can be used to store programs more efficiently and how the data can be encoded in such a way that the effective bandwidth of the link between memory and the processor chip is improved. In the first part of the thesis, different dictionary based compression schemes are evaluated and a new flexible scheme for efficient compression and execution of compressed programs is proposed. Since some sequences of instructions are more common than others, it is possible to reduce the memory needed for the program with little hardware overhead. In order to transfer data efficiently, this thesis also analyzes and classifies what types of value locality we can exploit in the data. Current state-of-the-art compression techniques are analyzed in the context of this categorization. Using this information, I show that it is possible to efficiently combine different techniques that work on different types of locality into a more efficient compression algorithm. Finally, I identify that the data-link compression schemes scale poorly with the number of nodes in a multiprocessor system. By studying frequent value encoding in such a framework I show that in some configurations, it is possible to reuse frequent values at each node and achieve significantly better bandwidth reductions than the baseline case.