Nested Sampling - multithreads

[Shook Ling Low]

Hi all,

I'm facing trouble in figure out how many hours should we spent to run jobs on using nested sampling. In my first trial for four combinations on speciation and clocks, I run for 1 million generation, with subsamplings 10000. In general, I spent ca 30 hours to get 1 particle done. However, calculation shown that I need 450 particles to achieve a standard deviation of less than 2.

I understand that we could reduce the runs timing by using multithreads. My question is would it be enough if we use 4 gpu and 4 cpu, for 7 particles runs (memory 400G). Is there anyway to reduce the usage of memory, as this keep my submitted job pending in a long queue.

Any suggestions /experiences are welcome. Many thanks.

[Aussie Dispensary]

Send me your Whatsapp 

--
You received this message because you are subscribed to the Google Groups "beast-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to beast-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/beast-users/7c8bc329-0b2a-4aa1-988e-0abdf44ffd78n%40googlegroups.com.

[higg...@gmail.com]

Do you really need a standard deviation less than 2 to compare the models you are interested in?

If not, it may be possible to run with fewer particles, and still get meaningful Bayes factors.

Another possibility is to combine results from independent runs, and reduce standard deviation that way.

Remco

[Shook Ling Low]

Dear Remco,

I have tested with several independent runs, and got a lower standard deviation (ca. 20) compare to only one run.

Here raise another concern, when I choose to go for higher generations like 100million generations, with subsamplings 100000 and 1 particle; the runs timing is again seem to be infinity. Therefore in this case, are the combination of independent runs (e.g. 1 millions generation each but run for 20-100 times) show the similar outputs? In addition, does the generations apply here must be the same as the generation we used for divergence dating?

Many thanks!