Beast fails to run when using beagle-lib and CUDA GPU on certain xml sets?

[throwern]

Hello and thank you in advance to anybody that can shed some light for
me.

So I believe I have properly built the latest versions of Beast and
beagle-lib to work with my CUDA installation. I can process the
benchmark.xml file using the following command from a folder inside
the Beastv1.5.2 directory

java -Djava.library.path=/usr/local/lib -cp ../lib/beast.jar:../lib/
beast-beagle.jar dr.app.beast.BeastMain -beagle ../examples/
benchmark.xml

Using BEAGLE resource 0 : CPU
Flags: DOUBLE SINGLE ASYNCH SYNCH COMPLEX CPU SSE
....
Time taken: 1.2012666666666665 minutes

by adding the option -Dbeagle.resource.order='1' or -
Dbeagle.resource.order='2' beagle will use the correct GPU instead of
the CPU

java -Dbeagle.resource.order='1' -Djava.library.path=/usr/local/lib -
cp ../lib/beast.jar:../lib/beast-beagle.jar dr.app.beast.BeastMain -
beagle data/Multigene.xml
Using BEAGLE resource 1 : Tesla C1060
Global memory (MB): 4096
Clock speed (Ghz): 1.30
Number of cores: 240
Flags: SINGLE ASYNCH SYNCH COMPLEX LSCALE GPU
....
Time taken: 2.3076 minutes

My problem occurs when trying to process a different set of data. In
this Multigene.nex file "The bayes block is written into the nexus
file and the evolutionary models are partitioned for each gene" I am a
cse guy so that is what the biologist told me. I used BEAUti to open
and export a beast Multigene.xml file from the Multigene.nex
original.

Regular beast runs the file in about 45 minutes on 8 cores.

Using beagle with the CPU takes about 25 minutes

Trying to use beagle with either of the GPU's leads to the following
error:

Error running file: Multigene.xml
The initial model is invalid because state has a zero probability.

If the log likelihood of the tree is -Inf, his may be because the
initial, random tree is so large that it has an extremely bad
likelihood which is being rounded to zero.

Alternatively, it may be that the product of starting mutation rate
and tree height is extremely small or extremely large.

Finally, it may be that the initial state is incompatible with
one or more 'hard' constraints (on monophyly or bounds on parameter
values. This will result in Priors with zero probability.

The individual components of the posterior are as follows:
The initial posterior is zero:
CompoundLikelihood(compoundModel)=
(OneOnX(ac)=0.0,
OneOnX(ag)=0.0,
OneOnX(at)=0.0,
OneOnX(cg)=0.0,
OneOnX(gt)=0.0,
OneOnX(constant.popSize)=1.6094,
CoalescentLikelihood(coalescentLikelihood)=63.6527),
CompoundLikelihood(compoundModel)=(BeagleTreeLikelihood
(treeLikelihood)=-63200.4734,
BeagleTreeLikelihood(treeLikelihood)=-Inf, BeagleTreeLikelihood
(treeLikelihood)=-41101.3766,
BeagleTreeLikelihood(treeLikelihood)=-14636.4653)

For more information go to <http://beast.bio.ed.ac.uk/>.
Exception in thread "main" java.lang.RuntimeException: ErrorLog:
Maximum number of errors reached.
Terminating BEAST
at dr.util.ErrorLogHandler.publish(Unknown Source)
at java.util.logging.Logger.log(Logger.java:476)
at java.util.logging.Logger.doLog(Logger.java:498)
at java.util.logging.Logger.log(Logger.java:521)
at java.util.logging.Logger.severe(Logger.java:1008)
at dr.app.beast.BeastMain.<init>(Unknown Source)
at dr.app.beast.BeastMain.main(Unknown Source)

Can anybody help me figure out how to get around this? Wouldn't all of
the initial settings and constraints and so on be the same for CPU or
GPU configurations? Or does the actual beast.xml file need to be
tailored to run on the GPU for larger datasets?

Thanks again,

-Nick

[ramonf]

I have the same problem (I apologise if you expected a reply to be a
decent answer!).

I would say that in my shell (bin/bash) environment, the single
quotation marks around -Dbeagle.resource.order='1' are not useful ...
all good without them.

Initially, and as an uninformed starter, I'd also surmise the same: a
nip and tuck of the xml file is prob required.

Myself I'll probably run a script to see which xml files beagle is
accepting and which not. Will report back. Cheers.

[Norm]

Any progress on this? Once I got beagle running (it gives me the
proper GPU info when I run & quit) my datasets all throw this sort of
error. Using BEAST 1.5.3.

On Nov 25, 1:09 pm, ramonf <ramonfal...@gmail.com> wrote:
> I have the same problem (I apologise if you expected a reply to be a
> decent answer!).
>
> I would say that in my shell (bin/bash) environment, the single
> quotation marks around -Dbeagle.resource.order='1' are not useful ...
> all good without them.
>
> Initially, and as an uninformed starter, I'd also surmise the same: a
> nip and tuck of the xml file is prob required.
>
> Myself I'll probably run a script to see which xml filesbeagleis
> accepting and which not. Will report back. Cheers.
>
> On Nov 7, 3:48 am, throwern <throw...@gmail.com> wrote:
>
> > Hello and thank you in advance to anybody that can shed some light for
> > me.
>
> > So I believe I have properly built the latest versions of Beast and
> >beagle-lib to work with my CUDA installation.  I can process the
> > benchmark.xml file using the following command from a folder inside
> > the Beastv1.5.2 directory
>
> > java -Djava.library.path=/usr/local/lib -cp ../lib/beast.jar:../lib/
> > beast-beagle.jar dr.app.beast.BeastMain -beagle../examples/
> > benchmark.xml
>

> >   UsingBEAGLEresource 0 : CPU

> >     Flags: DOUBLE SINGLE ASYNCH SYNCH COMPLEX CPU SSE
> >     ....
> >     Time taken: 1.2012666666666665 minutes
>
> > by adding the option -Dbeagle.resource.order='1'  or -

> > Dbeagle.resource.order='2'  beaglewill use the correct GPU instead of

> > the CPU
>
> > java -Dbeagle.resource.order='1' -Djava.library.path=/usr/local/lib -
> > cp ../lib/beast.jar:../lib/beast-beagle.jar dr.app.beast.BeastMain -
> >beagledata/Multigene.xml

> >   UsingBEAGLEresource 1 : Tesla C1060

> >     Global memory (MB): 4096
> >     Clock speed (Ghz): 1.30
> >     Number of cores: 240
> >     Flags: SINGLE ASYNCH SYNCH COMPLEX LSCALE GPU
> >     ....
> >    Time taken: 2.3076 minutes
>
> > My problem occurs when trying to process a different set of data. In
> > this Multigene.nex file "The bayes block is written into the nexus
> > file and the evolutionary models are partitioned for each gene" I am a
> > cse guy so that is what the biologist told me. I used BEAUti to open
> > and export a beast Multigene.xml file from the Multigene.nex
> > original.
>
> > Regular beast runs the file in about 45 minutes on 8 cores.
>

> > Usingbeaglewith the CPU takes about 25 minutes
>
> > Trying to usebeaglewith either of the GPU's  leads to the following

[Ramon Fallon]

hmmm... have had no ime to look at this, from briefest of cursory tests (in so-called "white box" mode):

Results meagre:

of the BEAST example files:

works w/beagle:
benchmark.xml testBSP.xml new_benchmark.xml testAscertainment.xml testGeneralSubstitutionModel.xml testIA.xml testLocalClock.xml testLognormalPrior.xml testMC3.xml testMCMC.xml testMixedModel.xml testML.xml testMultiThreading.xml testMVOU.xml testNormalModelGibbs.xml testPathSampling.xml testPMD.xml testRandomLocalClock.xml testRelaxedClock.xml testStrictClock.xml testTransmissionStatistic.xml testUncorrelatedRelaxedClock.xml

disna work:
arg-ecoli.xml arg-lepto.xml arg-rokas.xml testAncestralReconstruction.xml testBinaryDollo.xml testCodonLikelihood.xml testComparativeMethod.xml testCovarionHKY.xml testIBDweights.xml testLikelihood.xml testPosteriorIBD.xml testSequenceSimulator.xml


testPathSampling.xml and testAncestralReconstruction.xml deal with the same sequences, so I had a look to see why .. but they are too many differences to really get to the bottom of it (as I said, white box testing). Initially I thought it was the multi-alignment nature of testAncestralReconstruction.xml, but no that wasn't it. If I get some spare time later in the week, will look again.

[ramonf]

soryy, my references to "white-box testing" are incorrect, I should
have said "black box testing" (testing a system from its inputs and
outputs, rather than from (in this case) the source code (the inside)