For parallel computation,two techniques are proposed to couple Fourier transform and fluid dynamics.One is muiti-physics coupled communication technique and the other is parallel Fourier transform.By comparing efficiency of the two techniques,it is shown that the muiti-physics coupled communication technique is better than calling parallel FFTW technique as the number of cores is relatively small,while calling parallel FFTW technique is better as the number of cores is over a thousand.Scalability is demonstrated with a parallel efficiency above 50% on 16 384 processors.Numerical simulation on laser filamentation is shown.

With support of Federation Computing in JASMIN,two serial codes named radiation hydrodynamics code RH2D and particle transport code Sn2D as Federal members are concatenated to an integrated program RHSn2D,which uses efficiently thousands of processors to simulate a multiphysics composition system.Federal members of RHSn2D have mesh patches and parallel algorithms respectively,encapsulating parallel communication between them based on JASMIN.For a typical model discretized by 90 720 meshes,100 patches in RH2D,2 835 patches in Sn2D,48 directions and 16 energy groups,it shows that the integrated program RHSn2D achieves parallel efficiency of 36% with 1 024 processors.

An FFT parallel solver is designed and implemented in JASMIN infrastructure to solve problems,such as computation scale restriction,great change of data structure,different communications among data structures,and arguments' diversity of FFTW functions as using FFTW directly.The solver encapsulates several distributed storages of data,different communications and 1D FFT computations with FFTW.It provides standard functions to users.With this solver FFT parallel computing is realized conveniently.It was used in numerical simulations of laser filament.Scalability was demonstrated with a parallel efficiency above 80% on 2048 processors.

A one-dimensional load balancing method is proposed with memory constraint.It is an iterative method which converges to optimal partitioning linearly.Experiments on(1 024) CPUs with typical load imbalance models and a three-dimensional molecular dynamics program show effectiveness of the method.