非定常Monte Carlo输运问题的并行算法

文中给出了非定常MonteCarlo(下文简写为MC)输运问题的并行算法 ,对并行程序的加载运行模式进行了讨论和优化设计 .针对MC并行计算设计了一种理想情况下无通信的并行随机数发生器算法 .动态MC输运问题有大量的I/O操作 ,特别是读取剩余粒子数据文件需要大量的I/O时间 ,文中针对I/O问题 ,提出了三种并行I/O算法 .最后给出了并行算法的性能测试结果 ,对比串行计算时间 ,使用 6 4台处理机时的并行计算时间缩短了 30倍

Parallel Algorithm of Time-dependent Monte Carlo Transport

This paper is concerned with parallel computing code for time dependent MC transport. Two parallel algorithms is given and several loading problem of parallel program is explord. In order to reduce the time of entering and exiting parallel computing environment, the old code is modified to use MC computing code as a subroutine block. A parallel random number generator capable of MC parallel computing is proposed.Then the parallel I/O problems are discussed. The code discussed here is significantly different from the general particle transport MC simulations in the side of I/O requirements because it is designed to deal with the variational source problems. At each calculation step, it will randomly sample from a source file, and randomly write to another source file. These I/O operations consist of accesses to a large number of small, noncontiguous pieces of data. Moreover, the sizes of the source files and the amount of computations vary dynamically in each calculation step. So the I/O performance degrades drastically. To avoid this disadvantage, three parallel I/O algorithms are given. The first parallel I/O algorithm is direct parallelized from sequence code. In the second algorithm, each processor first writes the data to local memory, parallelly writes the local data to a file, then all processors parallelly read the file to local memory, and last all the processors can be able to access the small, noncontiguous pieces of data from local memory. In third algorithm, all the data are accessed in memory and the good performance of I/O is given. The experiments are performed on a 64 processor parallel machine that provides MPI2.0 and supports parallel file systems. The results indicate that the parallel algorithms proposed in the paper are practical and effective to simulate time dependent MC transport. The best speedup is up to 30.