Abstract submitted to the 4th Annual HPF User Group meeting




Interaction of the solar wind with the earth's magnetosphere has
been studied by using a 3-dimensional global MHD model.

Global MHD Simulation Code of Earth's Magnetosphere Using HPF/JA

    Tatsuki Ogino

    Solar-Terrestrial Environment Laboratory, Nagoya University

    We could execute a 3-dimensional global magnetohydrodynamic (MHD) simulation of interaction between the solar wind and the earth's magnetosphere in almost any kinds of computers by using a fully vectorized MHD code when we used to use the vector supercomputers such as CRAY Y-MP, Fujitsu VP-2600, Hitachi S820 and NEC SX-3. Therefore, we had many chances to work together with other scientists executing the computer simulation. However, as soon as the vector parallel and massive parallel supercomputers have come in simulation community, they began to speak their own dialects. We, simulation scientists are forced to learn the dialect and lost a common language. We would strongly hope to recover a common language in supercomputer world. It has been told that candidates of the common languages would be HPF (High Performance Fortran) and MPI (Massage Passing Interface). We have waited the time when we can use such a compiler in supercomputers.

    From this June, we have had a good chance to use HPF/JA (an extended version of HPF by JAHPF) by a supercomputer, Fujitsu VPP5000/56 in the Computer Center of Nagoya University. We immediately began to rewrite our 3-dimensional global MHD simulation code of the earth's magnetosphere from VPP Fortran to HPF/JA. The MHD code was fully vectorized and fully parallelized in VPP Fortran. We could successfully rewrite the 3-dimensional MHD code from VPP Fortran to HPF/JA in 3 weeks and we needed 2 more weeks to check final verification of the calculation results. The performance of the HPF MHD code was almost comparable with that of the VPP Fortran MHD code in the practical simulation and in large number of 56PEs (processing elements). Thus we can obtain a conclusion as: the fluid and MHD codes which are fully vectorized and fully parallelized in VPP Fortran could be relatively easily rewritten to HPF/JA, and the codes in HPF/JA can be expected to achieve comparable performance as those in VPP Fortran.

    The 3-dimensional global MHD simulation code of the earth's magnetosphere is mentioned in more detail. We solve the MHD and Maxwell's equations in the 3-dimensional Cartesian coordinates (x,y,z) as an initial and boundary value problem due to the modified leap-frog method. The MHD quantities are distributed in z-direction. The quantities in the neighbor grids can be calculated by HPF/JA instruction sentences of "shadow" and "reflect" when they exist in another PE. Unnecessary communication among PEs can be completely stopped by instruction of "independent, new" and "on home, local". Moreover, the lump transmission of data is used in calculation of the boundary condition by instruction of "asynchronous". It was not necessary to change the fundamental structure of the MHD code in the rewriting procedure. This was a big advantage to rewrite the MHD code from VPP Fortran to HPF/JA.

    I believe that we will not find essential difficulty to rewrite the program from VPP Fortran to HPF/JA and that we can expect almost comparable performance in Fujitsu VPP5000. The maximum performance of the 3-dimensional MHD code can be obtained as over 230 Gflops for 32PEs and over 400 Gflops for 56PEs. We hope that the MHD code rewritten in HPF/JA can be executed in other supercomputers such as Hitachi SR8000 and NEC SX-5 in near future. We have also disclosed a part of the boundary condition in the HPF/JA MHD code and a test program of the 3-dimensional wave equation in WWW as follows.
    http://gedas.stelab.nagoya-u.ac.jp/simulation/hpfja/hpf01.html


    Simulation Home Page.


    GEDAS Home Page
    STELab Home Page.