ITR: A High-Performance Compression Infrastructure for Extended Program Traces

Cornell University, Ithaca NY

Investigators

Abstract

Burtscher A High-Performance Compression Infrastructure for Extended Program Traces Abstract Program execution traces are widely used by researchers and educators to study program and processor behavior. Unfortunately, even capturing only a byte of information per executed instruction generates on the order of a gigabyte of data per second on a modern high-end microprocessor. Hence, nontrivial traces need to be stored in compressed form to be manageable. While good compression schemes exist for traces that capture only the PCs of the executed instructions, these schemes can be ineffective on extended traces that include important additional information such as the content of registers, effective addresses, values on a bus, etc. We propose to employ techniques from the value-prediction literature to compress extended traces. Preliminary results show that our approach delivers substantially improved compression rates on the traces where it matters the most, i.e., on the traces that other algorithms cannot compress well. The extended information is important because it encapsulates the parameters that are of interest in many current research endeavors. To enable the utilization of this kind of information in the classroom and the laboratory, high-performance compression tools need to be developed and made publicly available. This is the primary goal of this proposal. The tool will allow students and researchers alike to gain a better understanding of real programs and processors, and will promote the teaching and learning about these topics. It will be made available on the Web along with commented source code, documentation, and a tutorial. Moreover, the inner workings of the compression algorithm will be publicized and disseminated in research papers to encourage the usage as well as further studies of the compression algorithm.