Optimize ExaBayes run

[Tobias Senoner]

Hi everyone,

I am currently running ExaBays for about 1.5K sequences. It is running now for 10 days, and we are at 7.38% avg. standard deviation of split frequencies. I run it with 10 runs and 4 chains on 40 cores.

I have two questions:

1. Do you have any insights, if writing and reading the files from SSD would increase speed significantly (Currently, it stores it on an HDD)? And if so, by how much would you guesstimate?

2. Are my settings with 10 runs and 4 chains reasonable? Or would you suggest changing them? Also, are there some other settings that would speed up convergence, while maintaining high quality?

Best,

Tobias

[Alexandros Stamatakis]

Hi Tobias,

> I am currently running ExaBays for about 1.5K sequences. It is running
> now for 10 days, and we are at 7.38% avg. standard deviation of split
> frequencies. I run it with 10 runs and 4 chains on 40 cores.
>
> I have two questions:
> 1. Do you have any insights, if writing and reading the files from SSD
> would increase speed significantly (Currently, it stores it on an HDD)?
> And if so, by how much would you guesstimate?

It depends a bit on your thinning rate, but I don't have any insights on
this, for that we would need to measure how much time is spent on I/O by
using some profiling tool like gprof.

> 2. Are my settings with 10 runs and 4 chains reasonable? Or would you
> suggest changing them? Also, are there some other settings that would
> speed up convergence, while maintaining high quality?

The settings seems reasonable, albeit the question is if it makes sense
in general to run MCMC on 1.5K taxa. I would rather tend to use ML here,
but my opinion is biased of course.

One thing you could do is to assess the difficulty of your dataset using
Pythia: https://academic.oup.com/mbe/article/39/12/msac254/6832260

to quickly obtain a prior intuition about the difficulty of analyzing
this dataset.

What you can also do is to compute a couple of ML trees and provide them
as starting trees for the MCMC searches, we have pretty good experiences
with biasing searches like this, although the results are not published yet.

Alexis


>
> Best,
> Tobias
>
> --
> You received this message because you are subscribed to the Google
> Groups "ExaBayes" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to exabayes+u...@googlegroups.com
> <mailto:exabayes+u...@googlegroups.com>.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/exabayes/10caee4c-bdbc-48a5-891a-eed8b47289d2n%40googlegroups.com <https://groups.google.com/d/msgid/exabayes/10caee4c-bdbc-48a5-891a-eed8b47289d2n%40googlegroups.com?utm_medium=email&utm_source=footer>.

--
Alexandros (Alexis) Stamatakis

ERA Chair, Institute of Computer Science, Foundation for Research and
Technology - Hellas
Research Group Leader, Heidelberg Institute for Theoretical Studies
Full Professor, Dept. of Informatics, Karlsruhe Institute of Technology

www.biocomp.gr (Crete lab)
www.exelixis-lab.org (Heidelberg lab)