Debunking Statistical Simulation in HLS
Robert H. Bell Jr., Lieven Eeckhout, Lizy K. John
To appear at
Workshop on Duplicating, Deconstructing, and Debunking (WDDD04), Munich, Germany, June 19-20, 2004
Abstract
Statistical simulation systems can provide an accurate and efficient way to carry out early design studies for processors. The HLS system provides a rapid simulation capability, but our experiments demonstrate that several modeling improvements are needed. The front-end graph structure in HLS is hampered by workload modeling at the instruction level that does not accurately reflect program behavior when simulated. The workload and processor models require significant changes to provide accurate results for a variety of benchmarks. We improve HLS by modeling the workload at the granularity of the basic block and by changing the processor model to more closely reflect components in modern microprocessors. The specific techniques improve HLS accuracy by factor of 3.78 at the cost of increased storage and runtime requirements.
Our examination of HLS points to a major pitfall for simulator developers: reliance on a single small sets of benchmarks to qualify a simulation system. A simple regression model shows that the SPECint95 benchmarks, the only benchmarks used to calibrate HLS, have characteristics that yield to very simple modeling.