Performance Evaluation of Algorithms for Optimizing Processor Simulator Parameters

With Moore's Law coming to an end, the improvement of processor performance largely depends on the optimization of processor microarchitecture, so processor simulator plays a more and more important role in modern high-performance processor design. As the simulation task becomes complex and sophisticated, more and more tunable parameters with different data types have been exposed to human beings, and affect the quality-of-result (QoR) metrics of the simulator outcome. A fully functional processor simulator, such as gem5, often has huge parameter space and long execution time, which makes it almost impossible for engineers to perform thorough manual tuning. This paper mainly evaluates the performance of SA, TPE, and CMA-ES optimization algorithms in the above scenarios and gives some suggestions for reference based on the results.