Accelerating Monte Carlo Simulation Of Neutron Transport On The Intel MIC

At present,large scale and high performance computers are moving towards the development of hybrid heterogeneous direction of the commercial processor and the co-processor.In 2010,Intel released the MIC(Many Integrated Core)co-processor.MIC uses the x86 instruction set,and supports the traditional programming model,such as OpenMP,Intel Threading Building Blocks(TBB).Because of the advantages in the performance,more and more people pay attention to the MIC.Many researchers have applied MIC in the field of research,and have achieved good results.Particle transport simulation is an important application in the field of science and engineering.There are two numerical computation methods for this problem,namely,deterministic method and Monte Carlo method.Compared with the deterministic method,the Monte Carlo method is more suitable for solving the particle transport problem with complex geometric modes,but it needs a large amount of computation,and the hybrid heterogeneous computing system such as CPU/MIC supply the power to solve the problem.It is very necessary to research the parallel algorithm of particle transport based on the heterogeneous architecture.In this paper,we present a particle transport parallel algorithm for CPU/MIC hybrid heterogeneous system,which is based on the MCNP program.The main work includes the following aspects:A parallel pseudo random number generator based on MIC is proposed.The method of generating random number is linear congruence method(LCG).Firstly,the random number seed of each thread needs to be generated by stride at the program initialization.Every threads running on MIC can generate random number parallel by using random number seed.In this process,each thread can use 512 bits of the VPU to vector the algorithm,this method can improve the generation speed greatly.A parallel algorithm of Monte Carlo simulation based on MIC for particle transport is proposed.Based on the existing MCNP program,we proposed a high efficiency data structure and the method of dividing the task.Compared to the MCNP algorithm on the Xeon E5-2670 CPU,MCNP-MIC algorithm has achieved 5.6-fold speedup.A cooperative algorithm of Monte Carlo simulation of particle transport based on CPU/MIC hybrid heterogeneous architecture is proposed.The algorithm has designed multilevel task partitioning method considering the different computing ability of CPU and MIC,multilevel random number seed generation algorithm,and multilevel reduction algorithm.Test was carried out in the nodes of two E5-2670 CPU and 1 MIC cards.The results show that the algorithm has high performance.