Open GoogleCodeExporter opened 9 years ago
This looks to be an issue with scaling when the initial tree is a random
(extremely low probability) tree. Perhaps dynamic scaling should do scaling
always for the first 1000 steps to get over this hump?
Original comment by ramb...@gmail.com
on 17 Dec 2014 at 10:14
Actually, I suspect that this is issue is related to multithreading. See the
issue I made today on GitHub: https://github.com/armanbilge/BEAST3/issues/21
Original comment by armanbilge
on 17 Feb 2015 at 7:53
I agree this is probably due to synchronisation issues with threads. This issue
is related: https://code.google.com/p/beast-mcmc/issues/detail?id=756
Need to do a deeper examination of the threading issues - I suspect some
synchronization on most objects referenced in TreeLikelihood in the initial
traversal of the tree.
Original comment by ramb...@gmail.com
on 17 Feb 2015 at 8:00
Thanks for the pointer—I think that I may have found the problem.
BaseSubstitutionModel.getEigenDecomposition() has decompose() in a synchronized
block. However, the HKY and TN93 models override getEigenDecomposition()
without using synchronization. Adding synchronized to their method signatures
for getEigenDecomposition() appears to solve the problem for me. It looks like
all of the issues reported involved analyses using HKY.
Original comment by armanbilge
on 17 Feb 2015 at 8:48
Thanks. I will look into this in more detail. My feeling is it is better to
build the thread synchronization at the point where these classes are accessed
(i.e., TreeLikelihood) rather than relying on all these classes being thread
safe.
Original comment by ramb...@gmail.com
on 17 Feb 2015 at 10:30
I am struggling to replicate this by creating a new XML. Could you post an
example XML where it happens?
Original comment by ramb...@gmail.com
on 5 Mar 2015 at 9:23
I too am getting the "Underflow calculating likelihood. Attempting a
rescaling... " in v1.8.2 on a dual Xeon (8 core total) MacPro Snow Leop. I
chose 8 cores from the pulldown menu rather than leave at the default of
"automatic". The analyses continues to run without errors.
I have attached the xml for replication
Original comment by wbsi...@gmail.com
on 28 Mar 2015 at 12:16
Attachments:
"Underflow calculating likelihood. Attempting a rescaling... ": These messages
don't mean anything is wrong. This is not related the issue above.
Original comment by ramb...@gmail.com
on 28 Mar 2015 at 12:26
Original issue reported on code.google.com by
Quinn775...@gmail.com
on 15 Oct 2014 at 10:05