Question about principle of choosing number of batches and particles

[Yue JIN]

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.

[Jiankai YU]

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 

[Paul Romano]

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

--
You received this message because you are subscribed to the Google Groups "OpenMC Users Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openmc-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/openmc-users/6b9ba141-5c8c-486b-9c72-f9722b67259ao%40googlegroups.com.

[Yue JIN]

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：

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

On Tue, Jul 14, 2020 at 8:54 AM Jiankai YU <rockf...@gmail.com> wrote:

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 

On Tuesday, July 14, 2020 at 8:00:20 AM UTC-4, Yue JIN wrote:

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.

--
You received this message because you are subscribed to the Google Groups "OpenMC Users Group" group.

To unsubscribe from this group and stop receiving emails from it, send an email to openmc...@googlegroups.com.