Auto-tuning of Large-Scale Parallel Computing Applications Based on Bayesian Optimization

The performance of large-scale parallel computing applications is highly dependent on the parameter settings within complex systems. Due to the high-dimensional and nonlinear nature of kernel environment parameter spaces, traditional manual tuning methods are both labor-intensive and inefficient when exploring optimal parameter combinations, often leading to suboptimal performance. To address this issue, this paper proposes an automated parameter tuning framework based on Bayesian optimization, aimed at efficiently exploring complex high-dimensional parameter spaces and predicting better-performing parameter combinations. This framework builds a surrogate model, iteratively selects potentially efficient parameter configurations, and integrates SHAP to analyze the influence of each parameter on performance optimization. We validated the effectiveness of this framework in several typical large-scale parallel computing application scenarios, including NAS Parallel Benchmarks (NPB), quantum chemistry simulation software Quantum ESPRESSO and VASP, and the Weather Research and Forecasting model (WRF). Experimental results show that Bayesian optimization significantly shortens the overall tuning time and performs traditional methods such as genetic algorithms in terms of performance optimization. Experimental results show that Bayesian optimization significantly shortens the overall tuning time by 50% and improves application performance by up to 11%, outperforming traditional methods such as genetic algorithms in terms of performance optimization. Furthermore, SHAP analysis reveals the importance of key kernel parameters in enhancing application performance.