A deep-ensemble Bayesian optimization with computation budget allocation for design space exploration problems

The micro-architectures of processors are becoming increasingly complex, which introduces a large number of micro-architecture parameters. The design space exploration (DSE) problem, obtaining a set of micro-architecture parameters that could make the processor perform well, is both crucial and challenging. This simulation-based optimization problem involves a vast search space with more than 50 dimensions. Evaluating the performance of a given parameter set requires expensive simulations using the Cycle Accurate Simulator (CAS). The CAS simulates each typical program (referred to as a slice in DSE) included in the benchmark to obtain their respective instruction per cycle (IPC) scores, which are then weighted and aggregated to calculate the overall performance score for the given parameter design. The optimization objective in the DSE problem is the final score. Traditional DSE algorithms use black-box optimization methods, such as Bayesian optimization (BO), to optimize the parameters. However, these approaches neither leverage the information contained in the individual slice scores nor allocate the simulation budget efficiently. In this paper, we propose a deep-ensemble Bayesian optimization with computation budget allocation (DEBO-CBA) algorithm for DSE problems. The numerical and empirical tests demonstrate that the proposed method outperforms state-of-the-art approaches, including the black-box optimization algorithm HEBO and the genetic algorithm (GA). In a practical micro-architecture DSE problem, our algorithm requires 27% fewer iterations than HEBO to achieve a set of good enough parameters.