Optimal Scheduling Strategy of Multi-physics Coupled Computing Resources Based on Machine Learning

Abstract

Multi-physics coupled simulations imposes stringent demands on high-performance computing(HPC) resources, particularly regarding physical memory consumption. The inappropriate allocation of HPC resources may result in computational task failures or inefficient resource utilization. To enhance the utilization rate of the scarce large-memory computing resources in HPC clusters, this study explores the prediction methods for resource requirements and computation time of multi-physics coupling calculations, as well as rational allocation strategies for computing resources.By analyzing the characteristics of multi-physics coupling computational tasks and utilizing actual data collected from fluid-structure-acoustic(FSA) coupling computations, this study establishes an effective resource prediction model for FSA coupling calculations based on various machine learning algorithms. Subsequently, based on the forecasting results and HPC node configurations, an optimal allocation model for computing nodes is developed using an improved selection and elimination method based on weight estimation. This approach achieves a balanced optimization between computation time and resource allocation.Experimental results indicate that the proposed methods can effectively predict the resource requirements and computation time for multi-physics coupling calculations and provide optimal strategies for resource allocation, thereby resulting in an average decrease of 17.5% in computation time and an improvement of 20.4% in resource utilization efficiency.