Research And Realization On Parallel I/O Optimization Technology For High Resolution Ocean Model LICOM3

High resolution earth system models are of great significance for predicting severe climate and extenuating negative impact of extreme weather on human society.However,I/O part can occupy more run time and deplete more resources as model resolution and process number increase,which consequently affects operating efficiency of those models.We focus on I/O issues of existing earth system model and use LICOM3(LASG/IAP Climate System Ocean Model Version 3)as a typical research case to implement two parallel I/O optimization schemes with horizontal 10 km resolution based on ADIOS(Adaptive I/O System)and MPI-I/O(Message Passing Interface-I/O),respectively.As for ADIOS,we adopt POSIX and MPI methods to improve parallel output performance.A consistent one-to-one match is established between processes and output files in POSIX while a bijection relationship is introduced between communication-coordinators and output files in MPI.For MPI-I/O,we use collective I/O method with blocked independent file pointers.Also,a combination of file viewpoint and distributed array is adopted in MPI-I/O.In addition,we have conducted an enormous number of study in the Tianhe-HPC2 system public network environment to explore I/O performance influence factors and general high performance I/O optimization rules have been proposed for the Earth System Model.As the experiments show,when the process number scales to 4K,I/O part makes up a relatively significant share(50%)of the whole runtime for serial pattern.In contrast,the I/O time of LICOM3 using ADIOS is no more than 10% and that for MPI-I/O scheme is less than 25%.The two parallel schemes both greatly outperform the original serial I/O access pattern.