StarBeast3 multithreading

[Jimmy A Mcguire]

Hi All,

I'm attempting to run StarBeast3 on a 28-tip, 14-species, 500-locus data set using multiple threads. The species tree is almost entirely constrained to a particular topology using a series of monophyly/MRCA constraints. I have executed analyses a number of ways using either the GUI interface Mac version or running the Mac X86 version from the command line. The analyses suggest that all 20 (or 40) threads that I have called are being utilized, but the analysis never seems to actually use more than 1 processor and it moves VERY slowly (much more slowly than prior analyses I ran using StarBeast2 with comparable tips and loci). I have tried this with 1 or 2 instances as well but it doesn't make a difference. I'm hoping I'm overlooking something obvious that will fire up multiple processors. The commands I have used generally are like this: /Applications/BEAST\ 2.7.6_Mac/bin/beast -beagle -threads 20 Crot_28_500loci_constraint.xml. Any advice would be greatly appreciated. I would also be thrilled with an explanation of how one should utilize threads versus instances as this is confusing to me. Thanks!   

[higg...@gmail.com]

Hi Jimmy, 

The number of instances is the number of threads used per treelikelihood, while the number of threads is the maximum number of threads used at the same time. For StarBeast3 analyses, it is usually more efficient to have just 1 instance per treelikelihood. When using more instances the individual instance threads just gets in the way of running multiple threads for sampling individual gene trees. 

However, from what you describe, just having 1 instance is not helping getting more CPU usage, so I am not quite sure what is holding things up. If you can make the XML available, I can investigate further and see why there is not much multi-threading happening.

Cheers, Remco

[Jim McGuire]

Hi Remco,

Thanks so much for getting back to me! As it turns out, I figured out my mistake thanks to the StarBeast3 tutorial you have posted on github. I had mistakenly linked the tree and clock models and this was really slowing the analysis and apparently preventing parallelization. The impact was extreme - the analysis went from requiring 900+ hours per million generations to less than 4 hours per million generations. My initial analysis would have taken more than 10 years to complete whereas now it look like it will take about 2 weeks to finish. The analysis is now using 40 threads.

Many thanks!

Jim   

---
Jimmy A. McGuire
Professor, Department of Integrative Biology
Curator of Herpetology, Museum of Vertebrate Zoology
3101 Valley Life Sciences Building
University of California
Berkeley, CA 94720-3160
Phone: (510) 316-6201
Fax: (510) 643-8238
http://ib.berkeley.edu/labs/mcguire/
email: mcguirej at Berkeley.edu

--
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/5893e23b-9529-4f5d-8fad-72f341471cd7n%40googlegroups.com.