BEAST consistently yielding different results in single- versus multi-threaded mode

[Mark Farman]

Is there a reason why BEAST should consistently arrive at different TMRCA when running the same dataset in single- versus multi-threaded mode? In single-threaded, we are routinely getting dates of 49 - 50 years but this is shortened to 46 years when running with  2 or 16 threads. Same pattern observed regardless of running BEAST v1.10.4, v2.6.3 or v2.6.7. The later date is definitely incorrect, while the earlier one is entirely reasonable. 

One possibly relevant piece of information is that the population being studied is so recently evolved that we only have 422 SNPs among 70+ individuals (but there is a strong temporal signal, r2 = 0.5).