Research On Optimization Technology Of YHGSM Based On One-sided Communication

With the increasing computational complexity of numerical weather prediction models,it is necessary to design parallel algorithms aiming at scalability,so as to effectively utilize processors of ultra large scale and to meet the real-time requirement of NWP.YHGSM adopts semi-implicit semi-Lagrange advection scheme,and parallel algorithms for this scheme have already been successfully implemented on massively parallel machine.In addition,HALO problem is addressed by an ?on-demand? scheme in the interpolation stage,which reduces the communication volume to a certain extent.However,the computational efficiency and scalability are still insufficient.Taking into account the usability under our existing experimental environment of the code,this paper uses MPI-based one-sided communication to optimize the semi-Lagrange algorithm in YHGSM,and focus on the scalability and efficiency.Firstly,the one-sided communication models based on MPI and Coarray are both studied in detail,to understand the mechanism of underlying communication patterns aiming to analyze their advantages and disadvantages.Secondly,Cannon algorithm is used to find the best way to modify codes in YHGSM.Then,the algorithm flow chart is established according to the original communication scheme and detailed analysis of data structure of interpolation buff and send buffer is presented in this paper.After that,we summarize the key communication parameters and design a parallel algorithm based one-sided communication.Finally,we find that one-sided communication can reduce the transfer volume in HALO exchange and prove the scalability and efficiency of the model...