Question about principle of choosing number of batches and particles

User Support
1

Hi all,

To run a simulation, which setting below is recommended?
1000 batches * 1000 particles
or 100 batches * 10000 particles?
Does the choice depend on the tally that I want to obtain?

Thanks in advance,
Yue.

2

Hi Yue,

I am not the expert on sampling variation estimation or fission source convergence.
But I prefer the latter one, as more particles per batch is beneficial to reduce the shannon entropy and inter-batch correlation, so that finally accelerates the fission source convergence.

Best,
Jiankai

3

Yes, I generally tend to agree and much of what I’ve seen in the literature seems to point to larger batches being better. There are other practical benefits of using a larger batch, including less time on synchronization tasks between batches (for example, if you’re using MPI using larger batches as opposed to more batches will help reduce communication). However, it is possible to go “too far”. You’ll want to ensure that you have enough batches to ensure that you get a reliable estimate of the variance.

Best regards,
Paul

4

Thanks Paul for your help!

在 2020年7月15日星期三 UTC+8下午8:47:40，Paul Romano写道：