Study On The Method And Key Technology Of The Runtime Parameter Optimization Of The MPI In The Parallel Computing System

With the wide application of high performance computing(high performance computing)in aerospace,Astrophysics,meteorological material science,nuclear engineering,biomedicine and other fields,its high computing ability requirements bring new challenges to the software and hardware design of the computing platform.At present,message passing interface(message passing interface)has become the most common information transmission standard in high performance computing.Although it has hundreds of system parameters to be adjusted to adapt to different running environments,because of the diversity and complexity of scientific applications and operating platforms,it is very difficult and inconvenient to determine the appropriate parameter configuration.In this paper,based on the research of the run-time parameter setting in MPI operation,and the analysis of 3 different communication models,19 communication algorithms,2 communication protocols and the influence of the relative parameters on the application performance during the operation of MPI,we proposed a method for conveniently and simply determining the parameters of the MPI operation in the Tianhe-2 subsystem platform.The optimal parameter configuration is obtained according to the MPI colletive algorithms and the implementation protocols,which helps the real world appliction to get performance improvement for test and verify.This paper is devoted to optimizing the configuration of MPI runtime parameters by using the empirical method obtained from the measured data.The main contributions are as follows:1.The 3 communication models of the 8 collectives in the MPI are analyzed.With the model parameters of the system platform obtained through the testing tool,we calculated the performance difference between the different implementation algorithms of the collective operation by using the model formula.And then the benchmarks are run on the Tianhe-2 subsystem,to show the actual change of the communication performance of the algorithm realized in different message sizes and process sizes by the 8 collective operations.2.According to the analysis of the communication characteristics of MPI operations,the switching points of different implementation algorithms are obtained,and a simple and intuitive method is put forward,that is,the runtime parameters of MPI collective operation are set as the switching points corresponding to different implementation algorithms,which are regarded as the best runtime parameters of different collective operations under different application characteristics.And the parameters obtained are measured and compared with the default situation in actual application programs.The average performance improvement was 22.7%,and up to 40.7%.3.With the point-to-point communication test program,the performance influence on the point-to-point communication is obtained under the condition of various message sizes and process sizse by adjusting the run-time parameters for the MPI implementation protocols.In the same way,we set the runtime parameter as the switching point of the performance of the 2 communication protocols.The performance improvement of the Alltoall operation under different application characteristics is measured on the experimental platform,which is up to 16.0%.