An accurate ACOSSO metamodeling technique for processor architecture design space exploration

Processor architects usually design uniprocessor or chip multiprocessor (CMP) by using a platform-based approach. One of the major challenges in this approach is to explore the exponential-size design space composed of many tunable and interacting architectural parameters. An exhaustive search of the design space is prohibitive because of the expensive run-time of simulations. So an efficient design space exploration (DSE) strategy that can fast find the multi-objective architectural configurations (points in design space) in terms of system metrics like performance and energy is needed. In this paper, we propose an accurate and efficient adaptive component selection and smoothing operator (ACOSSO) metamodel assisted NSGA-II (MA-NSGA-II) multi-objective optimization (MOO) technique for processor DSE. We show the effectiveness of our methodology by comparing with linear regression (LR), restrict cubic splines (RCS), natural cubic splines (NCS) and artificial neural network (ANN) metamodeling techniques for processor design metrics prediction and architecture optimization. The experimental results show that, the proposed methodology achieves higher prediction accuracy and better architecture optimization results.