Trace File Compression Using Value Prediction Techniques

Abstract

 

Value prediction techniques have been widely-researched in recent years to predict the results of register writing instructions in CPUs.  In this project, we apply these techniques to a new domain: trace file compression.  Since the files we are interested in contain sequences of instruction addresses and their corresponding load values, value predictors are ideal for this purpose because they are known to yield a high percentage of correct predictions.   We compress the trace files by encoding the sequences of binary values in the files with predictor numbers.  To decompress, we run the compressed file through the same set of predictors in the order specified in the file to recreate the original values.  Our approach delivers substantially better compression rates than existing commercial software such as Winzip and gzip.