Machine Learning-Based Prediction of MPI Communication Time in High-Performance Computing Applications

The rapid advancements in high-performance computing (HPC) have made large-scale parallel computing feasible. As a commonly used parallel programming model, Message Passing Interface (MPI) plays a crucial role in HPC systems and directly affects the performance and efficiency of these applications. Therefore, modeling and predicting the performance of MPI communication is critically important. This paper utilizes a machine learning-based approach to predict MPI communication time. It uses performance analysis tools to collect real-time communication data and select critical features related to execution time. The data is categorized, then, for each type of MPI communication, a dedicated regression model is trained. Firstly, XGBoost and Artificial Neural Networks (ANN) are employed for function level modeling. Then, fusion techniques are used to combine XGBoost and ANN models to improve prediction accuracy. Finally, these function level models are integrated into an application level model to predict the overall communication time. This model is evaluated using CG, FT, and CPMD. Experimental results show that the average prediction errors of the model are 12.6%, 13.2%, and 13.5%, with a minimum error of 7%.