Performance Prediction Of Parallel Programs Based On Runtime Features And Machine Learning

As the computing power of high-performance computing systems continues to increase,the complexity and scale of their architectures and software systems continue to increase,which poses significant challenges to the design and optimization of various large-scale parallel applications.Research on massively parallel application performance modeling is becoming more and more important.Accurately predicting the performance of massively parallel programs not only enables users to analyze program performance,enabling them to efficiently execute applications on high-performance computing systems,but also helps users manage and schedule jobs,allocate scheduling strategies appropriately,and reduce job waiting time,and being able to conduct resource assessments to guide users to apply for resources.This paper proposes a parallel program performance prediction framework,which consists of feature acquisition,performance modeling and performance prediction.The framework uses the basic block frequency as a program feature and then uses machine learning algorithms to construct a performance prediction model for multiple input parameters.In the process of acquiring features,firstly,the smallscale parallel program runtime features are acquired by the instrumentation technique,and then the pre-processing technique is used to retain useful features.In the performance modeling part,the basic block frequency and process are input,and the program running time is output for performance modeling.In the performance prediction part,in order to reduce the overhead of acquiring large-scale parallel program features,a hybrid instrumentation algorithm and a program deletion algorithm are proposed.At the end of the paper,the proposed performance prediction framework is validated by executing six commonly used parallel test procedures on the Tianhe-2 supercomputer.The experimental results show that when the performance prediction framework adopts support vector machine regression modeling,the prediction effect is the best,the average prediction error is less than 10%,and the average prediction standard error is less than 7%.This paper also compares with two typical modeling methods based on input parameters.Experiments show that the proposed method is superior to the other two methods.In addition,in the prediction phase,the prediction overhead is less than 0.13% of the total program overhead.In summary,in the parallel program forecasting framework proposed in this paper,the generated forecasting model has strong generalization ability and low forecasting overhead.