The Study Of Fine-grained Performance Modeling For Large-scale HPC Applications

Large-scale HPC applications play an important role in intensive computing and mass data processing,and are widely applied in various scientific domains,such as climate forecast,seismic modeling,pharmaceutical development,and nuclear experiment simulation.They usually require a lot of computation resources.However,the computational performance of the application is not satisfactory in real-world applications.Most HPC applications have low machine utilization,and the peak performance that can be achieved is mostly much lower than the peak performance of the machine.Performance modeling is an effective method that can quickly analyze the performance on a specific platform.The performance information and prediction results provided by performance modeling can help researchers understand the performance characteristics of applications and guide the optimization.There are still some key issues that need to be discussed for performance modeling,such as the appropriate granularity of performance modeling,the comprehensive modeling of computational behaviors and the changes of communication performance with cluster and application scaling.Facing the above-mentioned challenges,this paper proposes a performance modeling approach for large-scale HPC applications with fine-grained profiling,and conducts research from the following three aspects:First,the comprehensive computation performance modeling.This paper proposes a performance modeling approach with fine-grained profiling,building performance models for all user-defined functions of the applications.The approach comprehensively and precisely describing the computational behavior,effectively improving the accuracy,and keeping low overhead.Second,characteristic-aware communication performance modeling.According to communication types,communication functions can be divided into five types.This paper automatically builds the communication function performance model based on the performance characteristics of different types of communication functions,and has achieved accurate prediction results.Last,automatic performance modeling and tuning for large-scale HPC applications.This paper designs and implements an automatic performance modelling method for large-scale HPC applications.The method could automatically collect performance data,perform performance analysis and modeling.In addition,this paper proposes a model-driven process layout optimization algorithm.After filtering all layouts for deduplication,the algorithm searches for the optimal process allocation under each layout based on the performance model.This paper applies the method to the benchmark NPB and the real-world applications of different scales.Experimental results show that the maximum relative error is less than 20%,the average relative error is less than 10%,and the overhead is less than 15%.According to the results,the method has better robustness and low overhead compared with the baselines.