resumed *Beast (Beast2) run exceeds designated number of MCMC steps

[Cathy]

Hi all,

I am running several *Beast analyses with 8 a priori "species" and large numbers of loci/partitions (20-100 loci for each run). Using Beast v.2.4.0 on a Linux HPC server, with -beagle_CPU, -beagle_SSE, and -threads flags, each run a total of 1 billion MCMC generations. The HPC system limits jobs to a wallclock time of 72 hours.

After the jobs are killed by HPC at 72 hours, I've successfully restarted them using -resume and -statefile. But I've noticed that the resumed jobs continue running past the 1 billion MCMC steps designated in the XML file. (Runs with my smallest data sets that finish in less than 72 hours behave normally, stopping at 1 billion.) This isn't a problem really, since I just manually killed the runs once I realized this was happening and deleted everything after 1 billion. Mainly I'm just curious if others are seeing this too, or if I'm doing something wrong.

[Graham]

I see this too, but it is what I expected. I interpret resume as "resume for another N generations" not "resume until N generations". I mainly use it when ESSs are too low after N, so I want another N.

Graham Jones

[Justin Bagley]

Hey Cathy!

Yeah, don't worry this also happened to me. It turns out that resuming runs results in the parameters being appended to any existing log files in the run directory. So, if you start a run and it goes x generations, then to get to N generations, you'll need to change the mcmc chain length in the XML file to N - x and then resume the run. This will work nicely and help avoid removing any extra samples. However, if you look at the x gens trace that you get the first go round in Tracer and you don't like it, one thing you can do is to re-run while overwriting the existing log file (using the -overwrite option) while specifying a different starting seed (-seed). 

Cheers,

Justin

Justin C. Bagley, Ph.D.

CNPq Young Talent Fellow Postdoc

& Senior Research Associate

Departamento de Zoologia

Universidade de Brasília

70910-900 Brasília, DF, Brasil

jcbagley (at) unb.br

Website: http://www.justinbagley.org

Lattes CV: http://lattes.cnpq.br/0028570120872581

[Cathy]

Thanks! That all makes sense. This actually turned out to be convenient, as my largest data set ended up needing more than 1 billion generations to reach stationarity, so I just let it keep running. I thought I remembered reading something on this forum a while ago about -resume -statefile allowing resumption of a run until completion of the original number of MCMC steps but not resumption for the purpose of adding additional steps for a longer run, so I was confused.