Method for constructing skeleton program and realizing prediction of operation time of high-performance calculation program

Abstract

The invention discloses a method for constructing a skeleton program and realizing prediction of operation time of a high-performance calculation program. The method is improved on the basis of the conventional skeleton program method, and a skeleton program capable of predicting the operation time of the program under different input is constructed by virtue of a machine learning method and the operation time data of the program. Compared with the existing skeleton program, the skeleton program constructed by the method can predict the operation time of the program under different inputs; andcompared with the regression method which uses a machine learning technology, the method has the advantage of predicting the operation time of the program in various operation environments.

Description

technical field [0001] The invention relates to the technical field of program running time prediction, in particular to a method for constructing a skeleton program and realizing high-performance computing program running time prediction. Background technique [0002] In the field of high performance computing, the prediction of the running time of parallel programs is an important research problem. Accurate prediction of program running time can effectively improve the task scheduling efficiency of parallel computing systems, thereby improving the performance of the entire system. [0003] A commonly used method for program runtime prediction is the skeleton program method. The skeleton program refers to the program obtained by replacing the calculation process of the parallel program with a sleep process based on the communication process of the parallel program, and reducing the communication amount and calculation time. The running time of the skeleton program is much...

AI Technical Summary

Problems solved by technology

Using runtime data as features to build a model can better reflect the impact of the running process of the program on the running time, so it usually has a better prediction effect, but it will make the prediction have to be performed after the program starts running for a period of time

The main problem with this approach is that whether or not the program's runtime data is used as features, the built model is only applicable to the current running environment

Images