Research On Data Parallel Methodsfor Large-scale Calculations With Reactor Monte Carlo Code RMC

With the higher requirements for the safety and economy of nuclear reactors as well as the developments of new types of nuclear systems, traditional methods and tools for reactor analysis are challenged. The Monte Carlo(MC) method is becoming an important research direction as next generation methods for reactor physics calculations. However, prohibitive data are required to be stored in large-scale calculations in MC codes and excessive memory demand turns into a key obstacle for MC method to accomplishpin-wise three-dimensionalfull core calculations. To handle the memory problem, this work investigates the data parallel methods, which include data decomposition and spatial domain decomposition, by using the Reactor Monte Carlo code RMC.Various kinds of data in MC codes are quantificationallyevaluated firstly in this thesis. The memory requirements are modeled and sources of memory problem in MC large-scale calculations are specified. Analyses show that tally data, material data and isotope densities in depletion are three major parts of memory storage in MC codes.Tally data decomposition(TDD) method is proposed against the primary memory occupation of tally. In TDD method, tally data is decomposed and distributed to parallel processors, and tallying are performed throughparallel commutations. Synchronous and asynchronous data communication algorithms are designed using message passing interface. TDDalgorithms are implemented in RMC and parallel efficiencies are analyzed. Numerical tests show that TDD manages to reduce memory footprint of tally data directly and asynchronous algorithm deserves high parallel efficiency.Different from TDD, spatial domain decomposition(SDD) method divides spatial geometry into domains, which are simulated separately by parallel processors, and particles crossing domains are communicated. Domain partition strategies are studiedbased on constructive solid geometry. Two modes of domain description,independent and coupled ones, are proposed. Asynchronous particle communication algorithm is designed to decrease communication costs. SDD function is developed in RMC and numerical tests of simple and full core models are executed. An approach by sorting source neutrons and random number is proposed to achieve reproducibility of calculation with SDD.Data Parallel methods are applied into MC burnup calculation finally to enhance large-scale depletion capability of RMC. Integrating TDD and depletion parallel method, a combined data decomposition strategy for tally and isotope densities in depletion is implemented. Numerical test of three-dimensionalfull core burnup calculation demonstrates the effect of data decomposition on increasing scale of MC burnup. SDD and MC burnup are coupled as well. By utilizing consistent domain partition in transport process and depletion process, data of tally, material and isotope densities are all decomposed. Numerical tests indicate that domain decomposed MC codes are capable of pin-wise full core burnup calculations with millions of depletion regions.