Speaker
Description
Implementing weight window (WW) is a usual method for variance reduction (VR) of Monte Carlo simulation, however as for a complex and large model simulation it frequently encounter the long histories (LH in abbreviation) problem in parallel computing. LH behavior shows as the running time of a single particle history is significantly longer than that of normal histories. It would take a disproportionate amount of time for Monte Carlo simulation to accomplish and place a detrimental effect on the efficiency of parallel computing. In this paper, the investigation of reason that causing LH was carried out firstly. A simple dog-log model was constructed to observe and analyze the LH phenomenon. Then comparative tests were carried out on a 3D model of the Chinese Fusion Engineering Testing Reactor (CFETR) with three approaches these are: a) analog running without any VR techniques; b) normal weight window VR technique; c) a novel approach proposed in this paper of limitation of weight window splitting. The results show that a suitable set of parameters in the improved WW module significantly improves the efficiency of variance reduction performance in parallel calculation, making the long history problem tractable without biasing results.
| Eligible for student paper award? | No |
|---|
