非定常输运问题适应于消息传递并行编程环境的香农熵计算方法

Efficient method of calculating Shannon entropy of non-static transport problem in message passing parallel programming environment

在多计算步的非定常输运问题的蒙特卡罗模拟中,为自动调整每一步的样本数以获得较高的计算效率,可以有多种准则.一种可选的方法是在每一步每隔若干样本监测一次系统中未死亡粒子属性分布对应的香农熵的收敛情况以决定何时停止追加样本,此种方法需要在每一步频繁计算香农熵值.由于在MPI消息传递并行编程环境下香农熵的经典计算方法必须广播大量的数据,导致每一步的计算时间随香农熵计算频率的提高而快速增大,这显然是不能满足实际需求的.本文提出了一种适应于消息传递并行编程环境的香农熵计算新方法,该方法计算得到的香农熵值并不等价于经典方法,但二者之间的差别会随着样本数的增加而趋于零.新方法的最大优势是高频计算香农熵值的时间代价大为降低,为最终实现基于香农熵收敛判断的每步样本数的自动调整奠定了必要的基础.

For the Monte Carlo simulation of the non-static transport problem, there must be many calculation steps. Because some particles cannot finish their transport in the last step, they are naturally used as the source particles of the present step. These particles are called undied particles. It is difficult to adjust the history number of each step to obtain higher efficiency because the adjusting rule is hard to find. The most direct method is to set a large enough history number for all steps. But evidently, it is unnecessary for some steps. Among all possible rules, one candidate of adjusting the history number is to check the convergence situation of Shannon entropy（corresponding to the distribution of some undied particle attributes） every some samples in each step to determine whether or not to simulate more particles. So,this method needs to calculate the Shannon entropy frequently. Because the classical method of calculating Shannon entropy in message passing parallel programming environment must reduce massive data, it is unpractical to be used in this situation for the great increasing of computation time with the high frequency of entropy calculation. In this paper,we propose an efficient method of calculating the entropy in the message passing parallel programming environment by letting each process calculate its entropy value based on the local data in each processer and calculating the final entropy by averaging all the entropy values gotten by all processes. The entropy value calculated by this method is not the same as that by the classical method when using finite history number, but the difference goes to zero when the history number goes to infinity. The most remarkable advantage of this method is the small increasing of computation time when calculating the entropy frequently. It is a suitable method of calculating Shannon entropy when adjusting the history number automatically based on the judgment of the convergence situation of Shannon entropy.