Research And Application Of Parallel Geometric Region Decomposition Parallel Algorithm For Monte Carlo Particle Transport

Numerical simulation has been acknowledged widely as one way having the same importance with theoretical analysis and experiments with the development of computer technology. For nuclear reactor design, criticality safety analysis and shield problem, the Monte Carlo (MC) method is a very powerful tool for particle transport simulation which has essential importance for these subjects. This method is suitable for analyzing particle configuration by it’s capability to deal with arbitrary geometrical complexity and physics complexity, which are remark characters of particle transport simulation for nuclear reactor design. So, the MC method is often used as a verification tool for deterministic methods which inevitably do some approximations and considered to be a realistic candidate for on-line nuclear reactor analysis in the future.Because of the independence among different sample paths, the MC method is thought as essentially parallel. Many particles can be simulated synchronously in different CPUs and each processor needs to store the whole geometry and tally information of target model. However, as the degree of fine geometrical and physical modeling increase gradually, the geometry information and tally data are too large to be stored in memory of a single CPU. For example, the full-core calculation of reactor often includes millions of cells and billions of tally data, A simple estimation can makes it clear that traditional MC code can not be qualified for problems like this.This thesis focuses on the domain decomposition algorithm of Monte Carlo method for full-core pin-by-pin simulation. Our object is to utilize the architecture features of multi-core parallel computer to calculate efficiently large-scale models with ten millions of cells and billions of tallies. Obviously, some key technology difficulties must be overcame for this purpose.Firstly, efficient domain decomposition algorithm and shadow region construction algorithm are designed based on tree data structure representing the whole model. The geometry information and the relevant tallies are partitioned into different parts, which are allocated to different cores separately. By this way, the memory problem is solved. Then, efficient irregular data communication algorithm is designed and implemented to support asynchronous particle transport simulation, which solves the problem of particle communication between different domains. For improving scalability, two-layer parallel computing is realized by combining the domain decomposition with traditional domain replication and identical results are obtained whatever the number of processors is by implementing a hierarchy of random number seeds. All above research results are integrated into JCOGIN infrastructure.Secondly, the Monte Carlo neutron-photon transport code JMCT is developed based on JCOGIN infrastructure. Transport module for encapsulating particle’s position movement and physics module for encapsulating physical processes in transport are designed by considering the characters of neutron and photon transport to realize the separation of physics and math. Furthermore, in physics module, parameter class has been designed to store static cross-section data hierarchically and collision class has been designed to calculate the cross section at current particle energy by interpolation and sample strictly according to physical processes. These two classes make the separation between data and operation be true. A number of international benchmark models are calculated to verify JMCT code which integrates these researches.Finally, two kinds of full-core reactor models are built up to test the domain decomposition algorithm. These models are based on the PWR reactor in the DAYAWAN nuclear power station in China and have hundred thousands and/or ten millions geometry cells respectively. A variety of domain decomposition patterns, including combination with domain replication, are tested by large-scale critical simulation.All the results show domain decomposition algorithm is efficient and the corresponding code is correct. This conclusion is gotten safely by it’s reducing memory requirements obviously and getting same result whatever the number of processors is. For a model containing over10millions cells and49tally energy groups whose memory requested only for tallies exceed12G, JMCT with domain decomposition algorithm can simulate it using eight domains, which is in stark contrast with traditional Monte Carlo code.In these calculations, more than10,240cores are used and80billions particles are simulated. A50%parallel efficiency is observed when using8192cores, which shows the efficiency and scalability of two-layer parallel combining domain decomposition and domain replication are well-behaved.