beast slow down and stops at 1% run

[joan pons]

Dear beast users,

I am running beast 1.6.1 with beagle on a GPU Tesla C2050 board.
Generally beast jobs slow down when they reach about 1% of the
generations set but about 50% of the job stop after that point. I get
the following error

Feb 24, 2011 5:13:07 PM dr.inference.markovchain.MarkovChain runChain
SEVERE: State was not correctly calculated after an operator move.
Likelihood evaluation: -124017.69294376519
Full Likelihood evaluation: -124017.69428653862
Operator: dr.inference.operators.SwapOperator@6d7f11fb
swapOperator(branchRates.categories)
4000 -122493.9707 -195.2448 -122298.7259 119.694 1.15E-2
-
Exception in thread "Thread-1" java.lang.RuntimeException: One or more
evaluation errors occured during the test phase of this
run. These errors imply critical errors which may produce incorrect
results.
at dr.inference.markovchain.MarkovChain.runChain(Unknown Source)
at dr.inference.mcmc.MCMC.chain(Unknown Source)
at dr.inference.mcmc.MCMC.run(Unknown Source)
at java.lang.Thread.run(Thread.java:636)

Most of the times I try to run the same xml file and it goes beyond
the critical point of 1% generations and finishes the full run.

Have you find the same error? Any tips to solve it?

Thanks,

Joan

[Marc Suchard]

Dear Joan,

Looking at how small your sequence data likelihood is, I suspect
you're running an analysis either with a very large dataset or under a
substitution model with a larger number of states. In both cases, any
form rescaling to avoid underflows can be problematic (in any ML or
Bayesian analysis). BEAGLE implements several rescaling
procedures. The default is something we are calling "dynamic". The
"always" procedure is often more stable, but slightly slower. Using
the following BEAST command-line option to switch:

-beagle_scaling always

best, Marc

PS Also, if you are brave, you could ignored these very small (and
hopefully insignificant) differences by setting the <mcmc> XML block
attribute fullEvaluation="0".

PPS If using BEAGLE or any other modeling extension or inference add-
on (e.g., *BEAST), please cite them in addition to the BEAST release
publications.